{

"cells": [
{
"cell_type": "markdown",
"id": "c2dc16c4-c7f4-4945-ba91-6430a51e6f5a",
"metadata": {
"id": "c2dc16c4-c7f4-4945-ba91-6430a51e6f5a"
},
"source": [
"<a href=\"https://colab.research.google.com/github/mrdbourke/pytorch-deep-
learning/blob/main/03_pytorch_computer_vision.ipynb\" target=\"_parent\"><img
src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In
Colab\"/></a>\n",
"\n",
"[View Source
Code](https://github.com/mrdbourke/pytorch-deep-learning/blob/main/
03_pytorch_computer_vision.ipynb) | [View
Slides](https://github.com/mrdbourke/pytorch-deep-learning/blob/main/slides/
03_pytorch_computer_vision.pdf) | [Watch Video
Walkthrough](https://youtu.be/Z_ikDlimN6A?t=50417) "
]
},
{
"cell_type": "markdown",
"id": "08f47c6a-3318-4e3f-8bb3-c520e00e63dd",
"metadata": {
"id": "08f47c6a-3318-4e3f-8bb3-c520e00e63dd"
},
"source": [
"# 03. PyTorch Computer Vision\n",
"\n",
"[Computer vision](https://en.wikipedia.org/wiki/Computer_vision) is the art of
teaching a computer to see.\n",
"\n",
"For example, it could involve building a model to classify whether a photo is
of a cat or a dog ([binary classification](https://developers.google.com/machine-
learning/glossary#binary-classification)).\n",
"\n",
"Or whether a photo is of a cat, dog or chicken ([multi-class classification]
(https://developers.google.com/machine-learning/glossary#multi-class-
classification)).\n",
"\n",
"Or identifying where a car appears in a video frame ([object detection]
(https://en.wikipedia.org/wiki/Object_detection)).\n",
"\n",
"Or figuring out where different objects in an image can be separated
([panoptic segmentation](https://arxiv.org/abs/1801.00868)).\n",
"\n",
"![example computer vision
problems](https://raw.githubusercontent.com/mrdbourke/pytorch-deep-learning/main/
images/03-computer-vision-problems.png)\n",
"*Example computer vision problems for binary classification, multiclass
classification, object detection and segmentation.*"
]
},
{
"cell_type": "markdown",
"id": "19179a39-0c6c-40f7-9891-09e17d107ecf",
"metadata": {
 "id": "19179a39-0c6c-40f7-9891-09e17d107ecf"
},
"source": [
"## Where does computer vision get used?\n",
"\n",
"If you use a smartphone, you've already used computer vision.\n",
"\n",
"Camera and photo apps use [computer vision to
enhance](https://machinelearning.apple.com/research/panoptic-segmentation) and sort
images.\n",
"\n",
"Modern cars use [computer vision](https://youtu.be/j0z4FweCy4M?t=2989) to
avoid other cars and stay within lane lines.\n",
"\n",
"Manufacturers use computer vision to identify defects in various products.\n",
"\n",
"Security cameras use computer vision to detect potential intruders.\n",
"\n",
"In essence, anything that can be described in a visual sense can be a
potential computer vision problem."
]
},
{
"cell_type": "markdown",
"id": "412e8bd1-0e6b-4ad6-8506-b28a8f669dc1",
"metadata": {
"id": "412e8bd1-0e6b-4ad6-8506-b28a8f669dc1"
},
"source": [
"## What we're going to cover\n",
"\n",
"We're going to apply the PyTorch Workflow we've been learning in the past
couple of sections to computer vision.\n",
"\n",
"![a PyTorch workflow with a computer vision
focus](https://raw.githubusercontent.com/mrdbourke/pytorch-deep-learning/main/
images/03-pytorch-computer-vision-workflow.png)\n",
"\n",
"Specifically, we're going to cover:\n",
"\n",
"| **Topic** | **Contents** |\n",
"| ----- | ----- |\n",
"| **0. Computer vision libraries in PyTorch** | PyTorch has a bunch of built-
in helpful computer vision libraries, let's check them out. |\n",
"| **1. Load data** | To practice computer vision, we'll start with some images
of different pieces of clothing from
[FashionMNIST](https://github.com/zalandoresearch/fashion-mnist). |\n",
"| **2. Prepare data** | We've got some images, let's load them in with a
[PyTorch `DataLoader`](https://pytorch.org/docs/stable/data.html) so we can use
them with our training loop. |\n",
"| **3. Model 0: Building a baseline model** | Here we'll create a multi-class
classification model to learn patterns in the data, we'll also choose a **loss
function**, **optimizer** and build a **training loop**. | \n",
"| **4. Making predictions and evaluating model 0** | Let's make some
predictions with our baseline model and evaluate them. |\n",
"| **5. Setup device agnostic code for future models** | It's best practice to
write device-agnostic code, so let's set it up. |\n",
"| **6. Model 1: Adding non-linearity** | Experimenting is a large part of
machine learning, let's try and improve upon our baseline model by adding non-
linear layers. |\n",
"| **7. Model 2: Convolutional Neural Network (CNN)** | Time to get computer
vision specific and introduce the powerful convolutional neural network
architecture. |\n",
"| **8. Comparing our models** | We've built three different models, let's
compare them. |\n",
"| **9. Evaluating our best model** | Let's make some predictions on random
images and evaluate our best model. |\n",
"| **10. Making a confusion matrix** | A confusion matrix is a great way to
evaluate a classification model, let's see how we can make one. |\n",
"| **11. Saving and loading the best performing model** | Since we might want
to use our model for later, let's save it and make sure it loads back in correctly.
|"
]
},
{
"cell_type": "markdown",
"id": "cddf62c3-f5e5-4f7e-852a-2ad6d38b7399",
"metadata": {
"id": "cddf62c3-f5e5-4f7e-852a-2ad6d38b7399"
},
"source": [
"## Where can you get help?\n",
"\n",
"All of the materials for this course [live on
GitHub](https://github.com/mrdbourke/pytorch-deep-learning).\n",
"\n",
"If you run into trouble, you can ask a question on the course [GitHub
Discussions page](https://github.com/mrdbourke/pytorch-deep-learning/discussions)
there too.\n",
"\n",
"And of course, there's the [PyTorch
documentation](https://pytorch.org/docs/stable/index.html) and [PyTorch developer
forums](https://discuss.pytorch.org/), a very helpful place for all things PyTorch.
"
]
},
{
"cell_type": "markdown",
"id": "a0bedcfc-e12a-4a81-9913-84c6a888742a",
"metadata": {
"id": "a0bedcfc-e12a-4a81-9913-84c6a888742a"
},
"source": [
"## 0. Computer vision libraries in PyTorch\n",
"\n",
"Before we get started writing code, let's talk about some PyTorch computer
vision libraries you should be aware of.\n",
"\n",
"| PyTorch module | What does it do? |\n",
"| ----- | ----- |\n",
"| [`torchvision`](https://pytorch.org/vision/stable/index.html) | Contains
datasets, model architectures and image transformations often used for computer
vision problems. |\n",
"| [`torchvision.datasets`](https://pytorch.org/vision/stable/datasets.html) |
Here you'll find many example computer vision datasets for a range of problems from
image classification, object detection, image captioning, video classification and
more. It also contains [a series of base classes for making custom datasets]
(https://pytorch.org/vision/stable/datasets.html#base-classes-for-custom-datasets).
|\n",
"| [`torchvision.models`](https://pytorch.org/vision/stable/models.html) | This
module contains well-performing and commonly used computer vision model
architectures implemented in PyTorch, you can use these with your own problems. | \
n",
"|
[`torchvision.transforms`](https://pytorch.org/vision/stable/transforms.html) |
Often images need to be transformed (turned into numbers/processed/augmented)
before being used with a model, common image transformations are found here. | \n",
"|
[`torch.utils.data.Dataset`](https://pytorch.org/docs/stable/data.html#torch.utils.
data.Dataset) | Base dataset class for PyTorch. | \n",
"|
[`torch.utils.data.DataLoader`](https://pytorch.org/docs/stable/data.html#module-
torch.utils.data) | Creates a Python iterable over a dataset (created with
`torch.utils.data.Dataset`). |\n",
"\n",
"> **Note:** The `torch.utils.data.Dataset` and `torch.utils.data.DataLoader`
classes aren't only for computer vision in PyTorch, they are capable of dealing
with many different types of data.\n",
"\n",
"Now we've covered some of the most important PyTorch computer vision
libraries, let's import the relevant dependencies.\n"
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "c263a60d-d788-482f-b9e7-9cab4f6b1f72",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "c263a60d-d788-482f-b9e7-9cab4f6b1f72",
"outputId": "20ba933b-6026-475f-a8d9-12cf416aff74"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"PyTorch version: 2.0.1+cu118\n",
"torchvision version: 0.15.2+cu118\n"
]
}
],
"source": [
"# Import PyTorch\n",
"import torch\n",
"from torch import nn\n",
"\n",
"# Import torchvision \n",
"import torchvision\n",
"from torchvision import datasets\n",
"from torchvision.transforms import ToTensor\n",
"\n",
"# Import matplotlib for visualization\n",
"import matplotlib.pyplot as plt\n",
"\n",
 "# Check versions\n",
"# Note: your PyTorch version shouldn't be lower than 1.10.0 and torchvision
version shouldn't be lower than 0.11\n",
"print(f\"PyTorch version: {torch.__version__}\\ntorchvision version:
{torchvision.__version__}\")"
]
},
{
"cell_type": "markdown",
"id": "48d6bfe7-91da-44eb-9ab6-7c41c1e9fa8e",
"metadata": {
"id": "48d6bfe7-91da-44eb-9ab6-7c41c1e9fa8e"
},
"source": [
"## 1. Getting a dataset\n",
"\n",
"To begin working on a computer vision problem, let's get a computer vision
dataset.\n",
"\n",
"We're going to start with FashionMNIST.\n",
"\n",
"MNIST stands for Modified National Institute of Standards and Technology.\n",
"\n",
"The [original MNIST dataset](https://en.wikipedia.org/wiki/MNIST_database)
contains thousands of examples of handwritten digits (from 0 to 9) and was used to
build computer vision models to identify numbers for postal services.\n",
"\n",
"[FashionMNIST](https://github.com/zalandoresearch/fashion-mnist), made by
Zalando Research, is a similar setup. \n",
"\n",
"Except it contains grayscale images of 10 different kinds of clothing.\n",
"\n",
"![example image of
FashionMNIST](https://raw.githubusercontent.com/mrdbourke/pytorch-deep-learning/
main/images/03-fashion-mnist-slide.png)\n",
"*`torchvision.datasets` contains a lot of example datasets you can use to
practice writing computer vision code on. FashionMNIST is one of those datasets.
And since it has 10 different image classes (different types of clothing), it's a
multi-class classification problem.*\n",
"\n",
"Later, we'll be building a computer vision neural network to identify the
different styles of clothing in these images.\n",
"\n",
"PyTorch has a bunch of common computer vision datasets stored in
`torchvision.datasets`.\n",
"\n",
"Including FashionMNIST in
[`torchvision.datasets.FashionMNIST()`](https://pytorch.org/vision/main/generated/
torchvision.datasets.FashionMNIST.html).\n",
"\n",
"To download it, we provide the following parameters:\n",
"* `root: str` - which folder do you want to download the data to?\n",
"* `train: Bool` - do you want the training or test split?\n",
"* `download: Bool` - should the data be downloaded?\n",
"* `transform: torchvision.transforms` - what transformations would you like to
do on the data?\n",
"* `target_transform` - you can transform the targets (labels) if you like
too.\n",
"\n",
 "Many other datasets in `torchvision` have these parameter options."
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "486f8377-6810-4367-859d-69dccc7aef95",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "486f8377-6810-4367-859d-69dccc7aef95",
"outputId": "877f93b2-12c5-477e-92bf-3ec3f1449282"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Downloading
http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-images-idx3-
ubyte.gz\n",
"Downloading
http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-images-idx3-
ubyte.gz to data/FashionMNIST/raw/train-images-idx3-ubyte.gz\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"100%|██████████| 26421880/26421880 [00:01<00:00, 16189161.14it/s]\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Extracting data/FashionMNIST/raw/train-images-idx3-ubyte.gz to
data/FashionMNIST/raw\n",
"\n",
"Downloading
http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-labels-idx1-
ubyte.gz\n",
"Downloading
http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-labels-idx1-
ubyte.gz to data/FashionMNIST/raw/train-labels-idx1-ubyte.gz\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"100%|██████████| 29515/29515 [00:00<00:00, 269809.67it/s]\n"
]
},
{
"name": "stdout",
"output_type": "stream",
 "text": [
"Extracting data/FashionMNIST/raw/train-labels-idx1-ubyte.gz to
data/FashionMNIST/raw\n",
"\n",
"Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-
images-idx3-ubyte.gz\n",
"Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-
images-idx3-ubyte.gz to data/FashionMNIST/raw/t10k-images-idx3-ubyte.gz\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"100%|██████████| 4422102/4422102 [00:00<00:00, 4950701.58it/s]\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Extracting data/FashionMNIST/raw/t10k-images-idx3-ubyte.gz to
data/FashionMNIST/raw\n",
"\n",
"Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-
labels-idx1-ubyte.gz\n",
"Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-
labels-idx1-ubyte.gz to data/FashionMNIST/raw/t10k-labels-idx1-ubyte.gz\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"100%|██████████| 5148/5148 [00:00<00:00, 4744512.63it/s]"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Extracting data/FashionMNIST/raw/t10k-labels-idx1-ubyte.gz to
data/FashionMNIST/raw\n",
"\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"\n"
]
}
],
"source": [
"# Setup training data\n",
"train_data = datasets.FashionMNIST(\n",
" root=\"data\", # where to download data to?\n",
" train=True, # get training data\n",
 " download=True, # download data if it doesn't exist on disk\n",
" transform=ToTensor(), # images come as PIL format, we want to turn into
Torch tensors\n",
" target_transform=None # you can transform labels as well\n",
")\n",
"\n",
"# Setup testing data\n",
"test_data = datasets.FashionMNIST(\n",
" root=\"data\",\n",
" train=False, # get test data\n",
" download=True,\n",
" transform=ToTensor()\n",
")"
]
},
{
"cell_type": "markdown",
"id": "a63246f6-3645-49de-88fe-ec18e78bfbaf",
"metadata": {
"id": "a63246f6-3645-49de-88fe-ec18e78bfbaf"
},
"source": [
"Let's check out the first sample of the training data."
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "43bfd3d9-a132-41e8-8ccd-5ae25a7da59a",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "43bfd3d9-a132-41e8-8ccd-5ae25a7da59a",
"outputId": "1595e80b-6a3f-4171-a128-ec506b4d8326"
},
"outputs": [
{
"data": {
"text/plain": [
"(tensor([[[0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0039, 0.0000, 0.0000, 0.0510,\
n",
" 0.2863, 0.0000, 0.0000, 0.0039, 0.0157, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0039, 0.0039, 0.0000],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0118, 0.0000, 0.1412, 0.5333,\
n",
" 0.4980, 0.2431, 0.2118, 0.0000, 0.0000, 0.0000, 0.0039, 0.0118,\
n",
" 0.0157, 0.0000, 0.0000, 0.0118],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0235, 0.0000, 0.4000, 0.8000,\
n",
" 0.6902, 0.5255, 0.5647, 0.4824, 0.0902, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0471, 0.0392, 0.0000],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.6078, 0.9255,\
n",
" 0.8118, 0.6980, 0.4196, 0.6118, 0.6314, 0.4275, 0.2510, 0.0902,\
n",
" 0.3020, 0.5098, 0.2824, 0.0588],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0039, 0.0000, 0.2706, 0.8118, 0.8745,\
n",
" 0.8549, 0.8471, 0.8471, 0.6392, 0.4980, 0.4745, 0.4784, 0.5725,\
n",
" 0.5529, 0.3451, 0.6745, 0.2588],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0039, 0.0039, 0.0039, 0.0000, 0.7843, 0.9098, 0.9098,\
n",
" 0.9137, 0.8980, 0.8745, 0.8745, 0.8431, 0.8353, 0.6431, 0.4980,\
n",
" 0.4824, 0.7686, 0.8980, 0.0000],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.7176, 0.8824, 0.8471,\
n",
" 0.8745, 0.8941, 0.9216, 0.8902, 0.8784, 0.8706, 0.8784, 0.8667,\
n",
" 0.8745, 0.9608, 0.6784, 0.0000],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.7569, 0.8941, 0.8549,\
n",
" 0.8353, 0.7765, 0.7059, 0.8314, 0.8235, 0.8275, 0.8353, 0.8745,\
n",
" 0.8627, 0.9529, 0.7922, 0.0000],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0039, 0.0118, 0.0000, 0.0471, 0.8588, 0.8627, 0.8314,\
n",
" 0.8549, 0.7529, 0.6627, 0.8902, 0.8157, 0.8549, 0.8784, 0.8314,\
n",
" 0.8863, 0.7725, 0.8196, 0.2039],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0235, 0.0000, 0.3882, 0.9569, 0.8706, 0.8627,\
n",
" 0.8549, 0.7961, 0.7765, 0.8667, 0.8431, 0.8353, 0.8706, 0.8627,\
n",
" 0.9608, 0.4667, 0.6549, 0.2196],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0157, 0.0000, 0.0000, 0.2157, 0.9255, 0.8941, 0.9020,\
n",
" 0.8941, 0.9412, 0.9098, 0.8353, 0.8549, 0.8745, 0.9176, 0.8510,\
n",
" 0.8510, 0.8196, 0.3608, 0.0000],\n",
" [0.0000, 0.0000, 0.0039, 0.0157, 0.0235, 0.0275, 0.0078, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.9294, 0.8863, 0.8510, 0.8745,\
n",
" 0.8706, 0.8588, 0.8706, 0.8667, 0.8471, 0.8745, 0.8980, 0.8431,\
n",
" 0.8549, 1.0000, 0.3020, 0.0000],\n",
" [0.0000, 0.0118, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.2431, 0.5686, 0.8000, 0.8941, 0.8118, 0.8353, 0.8667,\
n",
" 0.8549, 0.8157, 0.8275, 0.8549, 0.8784, 0.8745, 0.8588, 0.8431,\
n",
" 0.8784, 0.9569, 0.6235, 0.0000],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0706, 0.1725, 0.3216, 0.4196,\
n",
" 0.7412, 0.8941, 0.8627, 0.8706, 0.8510, 0.8863, 0.7843, 0.8039,\
n",
" 0.8275, 0.9020, 0.8784, 0.9176, 0.6902, 0.7373, 0.9804, 0.9725,\
n",
" 0.9137, 0.9333, 0.8431, 0.0000],\n",
" [0.0000, 0.2235, 0.7333, 0.8157, 0.8784, 0.8667, 0.8784, 0.8157,\
n",
" 0.8000, 0.8392, 0.8157, 0.8196, 0.7843, 0.6235, 0.9608, 0.7569,\
n",
" 0.8078, 0.8745, 1.0000, 1.0000, 0.8667, 0.9176, 0.8667, 0.8275,\
n",
" 0.8627, 0.9098, 0.9647, 0.0000],\n",
" [0.0118, 0.7922, 0.8941, 0.8784, 0.8667, 0.8275, 0.8275, 0.8392,\
n",
" 0.8039, 0.8039, 0.8039, 0.8627, 0.9412, 0.3137, 0.5882, 1.0000,\
n",
" 0.8980, 0.8667, 0.7373, 0.6039, 0.7490, 0.8235, 0.8000, 0.8196,\
n",
" 0.8706, 0.8941, 0.8824, 0.0000],\n",
" [0.3843, 0.9137, 0.7765, 0.8235, 0.8706, 0.8980, 0.8980, 0.9176,\
n",
" 0.9765, 0.8627, 0.7608, 0.8431, 0.8510, 0.9451, 0.2549, 0.2863,\
n",
 " 0.4157, 0.4588, 0.6588, 0.8588, 0.8667, 0.8431, 0.8510, 0.8745,\
n",
" 0.8745, 0.8784, 0.8980, 0.1137],\n",
" [0.2941, 0.8000, 0.8314, 0.8000, 0.7569, 0.8039, 0.8275, 0.8824,\
n",
" 0.8471, 0.7255, 0.7725, 0.8078, 0.7765, 0.8353, 0.9412, 0.7647,\
n",
" 0.8902, 0.9608, 0.9373, 0.8745, 0.8549, 0.8314, 0.8196, 0.8706,\
n",
" 0.8627, 0.8667, 0.9020, 0.2627],\n",
" [0.1882, 0.7961, 0.7176, 0.7608, 0.8353, 0.7725, 0.7255, 0.7451,\
n",
" 0.7608, 0.7529, 0.7922, 0.8392, 0.8588, 0.8667, 0.8627, 0.9255,\
n",
" 0.8824, 0.8471, 0.7804, 0.8078, 0.7294, 0.7098, 0.6941, 0.6745,\
n",
" 0.7098, 0.8039, 0.8078, 0.4510],\n",
" [0.0000, 0.4784, 0.8588, 0.7569, 0.7020, 0.6706, 0.7176, 0.7686,\
n",
" 0.8000, 0.8235, 0.8353, 0.8118, 0.8275, 0.8235, 0.7843, 0.7686,\
n",
" 0.7608, 0.7490, 0.7647, 0.7490, 0.7765, 0.7529, 0.6902, 0.6118,\
n",
" 0.6549, 0.6941, 0.8235, 0.3608],\n",
" [0.0000, 0.0000, 0.2902, 0.7412, 0.8314, 0.7490, 0.6863, 0.6745,\
n",
" 0.6863, 0.7098, 0.7255, 0.7373, 0.7412, 0.7373, 0.7569, 0.7765,\
n",
" 0.8000, 0.8196, 0.8235, 0.8235, 0.8275, 0.7373, 0.7373, 0.7608,\
n",
" 0.7529, 0.8471, 0.6667, 0.0000],\n",
" [0.0078, 0.0000, 0.0000, 0.0000, 0.2588, 0.7843, 0.8706, 0.9294,\
n",
" 0.9373, 0.9490, 0.9647, 0.9529, 0.9569, 0.8667, 0.8627, 0.7569,\
n",
" 0.7490, 0.7020, 0.7137, 0.7137, 0.7098, 0.6902, 0.6510, 0.6588,\
n",
" 0.3882, 0.2275, 0.0000, 0.0000],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.1569,\
n",
" 0.2392, 0.1725, 0.2824, 0.1608, 0.1373, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000],\n",
" [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,\
n",
" 0.0000, 0.0000, 0.0000, 0.0000]]]),\n",
 " 9)"
]
},
"execution_count": 3,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# See first training sample\n",
"image, label = train_data[0]\n",
"image, label"
]
},
{
"cell_type": "markdown",
"id": "9ad9d782-06cb-4591-ae3c-3a8b2389a1b2",
"metadata": {
"id": "9ad9d782-06cb-4591-ae3c-3a8b2389a1b2"
},
"source": [
"### 1.1 Input and output shapes of a computer vision model\n",
"\n",
"We've got a big tensor of values (the image) leading to a single value for the
target (the label).\n",
"\n",
"Let's see the image shape."
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "c2997d9f-b574-4d23-aa34-1a4df1751226",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "c2997d9f-b574-4d23-aa34-1a4df1751226",
"outputId": "d9c4283b-aab8-410f-dd7f-03f08d943366"
},
"outputs": [
{
"data": {
"text/plain": [
"torch.Size([1, 28, 28])"
]
},
"execution_count": 4,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# What's the shape of the image?\n",
"image.shape"
]
},
{
"cell_type": "markdown",
 "id": "b5326a05-f807-448d-99a3-6d03fc8739f8",
"metadata": {
"id": "b5326a05-f807-448d-99a3-6d03fc8739f8"
},
"source": [
"The shape of the image tensor is `[1, 28, 28]` or more specifically:\n",
"\n",
"```\n",
"[color_channels=1, height=28, width=28]\n",
"```\n",
"\n",
"Having `color_channels=1` means the image is grayscale.\n",
"\n",
"![example input and output shapes of the fashionMNIST
problem](https://raw.githubusercontent.com/mrdbourke/pytorch-deep-learning/main/
images/03-computer-vision-input-and-output-shapes.png)\n",
"*Various problems will have various input and output shapes. But the premise
remains: encode data into numbers, build a model to find patterns in those numbers,
convert those patterns into something meaningful.*\n",
"\n",
"If `color_channels=3`, the image comes in pixel values for red, green and blue
(this is also known as the [RGB color
model](https://en.wikipedia.org/wiki/RGB_color_model)).\n",
"\n",
"The order of our current tensor is often referred to as `CHW` (Color Channels,
Height, Width).\n",
"\n",
"There's debate on whether images should be represented as `CHW` (color
channels first) or `HWC` (color channels last).\n",
"\n",
"> **Note:** You'll also see `NCHW` and `NHWC` formats where `N` stands for
*number of images*. For example if you have a `batch_size=32`, your tensor shape
may be `[32, 1, 28, 28]`. We'll cover batch sizes later.\n",
"\n",
"PyTorch generally accepts `NCHW` (channels first) as the default for many
operators.\n",
"\n",
"However, PyTorch also explains that `NHWC` (channels last) performs better and
is [considered best practice](https://pytorch.org/blog/tensor-memory-format-
matters/#pytorch-best-practice). \n",
"\n",
"For now, since our dataset and models are relatively small, this won't make
too much of a difference.\n",
"\n",
"But keep it in mind for when you're working on larger image datasets and using
convolutional neural networks (we'll see these later).\n",
"\n",
"Let's check out more shapes of our data."
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "fc4f768c-c3f6-454d-a633-673ad1d6eca0",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "fc4f768c-c3f6-454d-a633-673ad1d6eca0",
 "outputId": "fcac1ff4-5b9a-4459-a05e-77482f0e6776"
},
"outputs": [
{
"data": {
"text/plain": [
"(60000, 60000, 10000, 10000)"
]
},
"execution_count": 5,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# How many samples are there? \n",
"len(train_data.data), len(train_data.targets), len(test_data.data),
len(test_data.targets)"
]
},
{
"cell_type": "markdown",
"id": "6e0267d5-946b-4c53-af69-61acd3527972",
"metadata": {
"id": "6e0267d5-946b-4c53-af69-61acd3527972"
},
"source": [
"So we've got 60,000 training samples and 10,000 testing samples.\n",
"\n",
"What classes are there?\n",
"\n",
"We can find these via the `.classes` attribute."
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "e22849c6-d93f-4b38-8403-5ebf0deaf008",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "e22849c6-d93f-4b38-8403-5ebf0deaf008",
"outputId": "6e18aa0f-b8a0-45ee-9f4e-8931bcdfbec0"
},
"outputs": [
{
"data": {
"text/plain": [
"['T-shirt/top',\n",
" 'Trouser',\n",
" 'Pullover',\n",
" 'Dress',\n",
" 'Coat',\n",
" 'Sandal',\n",
" 'Shirt',\n",
" 'Sneaker',\n",
" 'Bag',\n",
" 'Ankle boot']"
 ]
},
"execution_count": 6,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# See classes\n",
"class_names = train_data.classes\n",
"class_names"
]
},
{
"cell_type": "markdown",
"id": "abdd225c-5742-4d9c-8e8d-fb30a9c3cb6e",
"metadata": {
"id": "abdd225c-5742-4d9c-8e8d-fb30a9c3cb6e"
},
"source": [
"Sweet! It looks like we're dealing with 10 different kinds of clothes.\n",
"\n",
"Because we're working with 10 different classes, it means our problem is
**multi-class classification**.\n",
"\n",
"Let's get visual."
]
},
{
"cell_type": "markdown",
"id": "fb625d80-6a98-471e-a758-4de0ce0f3a64",
"metadata": {
"id": "fb625d80-6a98-471e-a758-4de0ce0f3a64"
},
"source": [
"### 1.2 Visualizing our data"
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "b1df1f2c-28c9-43bf-aaef-cf996c9ae1c5",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 469
},
"id": "b1df1f2c-28c9-43bf-aaef-cf996c9ae1c5",
"outputId": "9bbdbb0d-eed3-408a-bd7b-03aa22cb35bb"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Image shape: torch.Size([1, 28, 28])\n"
]
},
{
 "data": {
"image/png":
"iVBORw0KGgoAAAANSUhEUgAAAaAAAAGzCAYAAABpdMNsAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIH
ZlcnNpb24zLjcuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/
bCgiHAAAACXBIWXMAAA9hAAAPYQGoP6dpAAAkhklEQVR4nO3de3SU9b3v8c/
kNgSYTAghNwkYUEAFYkshplhESYG0xwPK7tHWswo9Li0YXEXarQu3ilq70+La1lOLes7aLdS1xNuqyJZtOV
VogrQJyu1QaptCGgUlCRfNTMh1kvmdPzhGI9ffwyS/
JLxfa81aZOb58Px4eJJPnszMNz5jjBEAAL0szvUCAAAXJwoIAOAEBQQAcIICAgA4QQEBAJyggAAATlBAAAA
nKCAAgBMUEADACQoIAOAEBQT0kp07d2ru3LlKSUlRIBDQ7NmztWfPHtfLApzxMQsO6Hm7du3S9OnTlZubq+
9///uKRqN6+umn9fHHH+udd97R+PHjXS8R6HUUENALvvnNb6qiokL79+/X8OHDJUm1tbUaN26cZs+erd/
+9reOVwj0Pn4EB/
SCt99+W0VFRV3lI0nZ2dm67rrrtHHjRp04ccLh6gA3KCCgF7S1tSk5OfmU+wcPHqz29nbt27fPwaoAtyggo
BeMHz9elZWV6uzs7Lqvvb1d27dvlyR99NFHrpYGOEMBAb3grrvu0t///
nfdfvvteu+997Rv3z5997vfVW1trSSppaXF8QqB3kcBAb1g8eLFuv/+
+7Vu3TpdddVVmjRpkqqrq3XvvfdKkoYOHep4hUDvo4CAXvKTn/xE9fX1evvtt7V37169+
+67ikajkqRx48Y5Xh3Q+3gZNuDQtGnTVFtbqw8++EBxcXw/
iIsLZzzgyEsvvaR3331Xy5Yto3xwUeIKCOgFW7du1aOPPqrZs2dr+PDhqqys1Jo1a/
T1r39dr7/+uhISElwvEeh1nPVAL7jkkksUHx+vxx9/
XI2NjcrLy9Njjz2m5cuXUz64aHEFBABwgh88AwCcoIAAAE5QQAAAJyggAIATFBAAwAkKCADgRJ97A0I0GtX
hw4cVCATk8/
lcLwcAYMkYo8bGRuXk5Jx1ykefK6DDhw8rNzfX9TIAABfo0KFDGjly5Bkf73MFFAgEJEnX6htKUKLj1QAAb
HUoom16o+vr+Zn0WAGtXr1ajz/+uOrq6pSfn6+nnnpK06ZNO2fu0x+7JShRCT4KCAD6nf8/
X+dcT6P0yIsQXnrpJS1fvlwrV67Url27lJ+frzlz5ujIkSM9sTsAQD/
UIwX0xBNP6I477tD3vvc9XXnllXr22Wc1ePBg/
frXv+6J3QEA+qGYF1B7e7t27typoqKiz3YSF6eioiJVVFScsn1bW5vC4XC3GwBg4It5AR07dkydnZ3KzMzs
dn9mZqbq6upO2b60tFTBYLDrxivgAODi4PyNqCtWrFAoFOq6HTp0yPWSAAC9IOavgktPT1d8fLzq6+u73V9
fX6+srKxTtvf7/fL7/
bFeBgCgj4v5FVBSUpKmTJmizZs3d90XjUa1efNmFRYWxnp3AIB+qkfeB7R8+XItXLhQX/
nKVzRt2jQ9+eSTampq0ve+972e2B0AoB/
qkQK65ZZbdPToUT300EOqq6vT1VdfrU2bNp3ywgQAwMXLZ4wxrhfxeeFwWMFgUDM1j0kIANAPdZiIyrRBoV
BIKSkpZ9zO+avgAAAXJwoIAOAEBQQAcIICAgA4QQEBAJyggAAATlBAAAAnKCAAgBMUEADACQoIAOAEBQQAc
IICAgA4QQEBAJyggAAATlBAAAAnKCAAgBMUEADACQoIAOAEBQQAcIICAgA4QQEBAJyggAAATlBAAAAnKCAA
gBMUEADACQoIAOBEgusFAH2Kz2efMSb26ziN+OFp1plP5ozztK+UdZWectY8HG9fQqJ1xkTarTN9npdz1as
eOse5AgIAOEEBAQCcoIAAAE5QQAAAJyggAIATFBAAwAkKCADgBAUEAHCCAgIAOEEBAQCcoIAAAE5QQAAAJx
hGCnyOLz7eOmM6OqwzcVdfaZ356/eH2u+nxToiSUpsmmadSWiJ2u/
n9zusM706WNTLsFQP55B89tcCvXkcfAl2VeEzRjqPTwuugAAATlBAAAAnKCAAgBMUEADACQoIAOAEBQQAcI
ICAgA4QQEBAJyggAAATlBAAAAnKCAAgBMUEADACYaRAp9jO3RR8jaM9NCcVOvMbYVvW2f+eHSMdUaSPvBnW
WdMsv1+EooKrTPjnv7IOtPx/kHrjCTJGPuIh/PBi/
hhw7wFOzvtI+Gw1fbGnN8x4AoIAOAEBQQAcCLmBfTwww/L5/
N1u02YMCHWuwEA9HM98hzQVVddpbfeeuuznXj4uToAYGDrkWZISEhQVpb9k5gAgItHjzwHtH//
fuXk5GjMmDG67bbbdPDgmV+B0tbWpnA43O0GABj4Yl5ABQUFWrt2rTZt2qRnnnlGNTU1+trXvqbGxsbTbl9
aWqpgMNh1y83NjfWSAAB9UMwLqLi4WN/
61rc0efJkzZkzR2+88YYaGhr08ssvn3b7FStWKBQKdd0OHToU6yUBAPqgHn91QGpqqsaNG6cDBw6c9nG/
3y+/39/TywAA9DE9/j6gEydOqLq6WtnZ2T29KwBAPxLzAvrRj36k8vJyvf/++/rTn/
6km266SfHx8fr2t78d610BAPqxmP8I7sMPP9S3v/
1tHT9+XCNGjNC1116ryspKjRgxIta7AgD0YzEvoBdffDHWfyXQa6Ktrb2yn/YvnbDO/
FNwh3VmUFzEOiNJ5XFR68xHW+xfwdo52f44fPBEwDoT3f1V64wkDd9nP7gzZXetdebYjEusM0en2A9KlaTM
SvvMsLeqrbY30Xbp2Lm3YxYcAMAJCggA4AQFBABwggICADhBAQEAnKCAAABOUEAAACcoIACAExQQAMAJCgg
A4AQFBABwggICADjR47+QDnDC5/OWM/YDHk/8t2usM9+9ssw6Ux2xnyg/
Mulj64wkfStnp33ov9tnfll1nXWm6R9B60zcEG+DO+uusf8e/
aN59v9PJtJhnRm2y9uX77iF9daZcPsYq+07Iq3ShvNYi/
VKAACIAQoIAOAEBQQAcIICAgA4QQEBAJyggAAATlBAAAAnKCAAgBMUEADACQoIAOAEBQQAcIICAgA4QQEBA
JxgGjZ6l9cp1X3YNfe9Y525fuh7PbCSU10ib1Ogm0ySdaahc4h1ZuWV/
2mdOTouYJ2JGG9f6v59/1etMyc8TOuO77D/
vLjmf+y2zkjSgrR3rTOrfjvJavsOEzmv7bgCAgA4QQEBAJyggAAATlBAAAAnKCAAgBMUEADACQoIAOAEBQQ
AcIICAgA4QQEBAJyggAAATlBAAAAnGEaK3mW8Dcfsy/
afyLDOHE8Zap2p60i1zgyPP2GdkaRAXIt15tLEY9aZo532g0XjE6PWmXYTb52RpEeuet0603pFonUm0ddpn
fnqoMPWGUn61nvftc4M0T887etcuAICADhBAQEAnKCAAABOUEAAACcoIACAExQQAMAJCggA4AQFBABwggIC
ADhBAQEAnKCAAABOUEAAACcYRgpcoBF++4Gfg3wR60ySr8M6czgyzDojSftbxltn/h62H8o6N/
Mv1pmIh8Gi8fI2BNfLkNCcxE+sM63GfoCp/
Rl00vRM+8Giezzu61y4AgIAOEEBAQCcsC6grVu36sYbb1ROTo58Pp9ee+21bo8bY/
TQQw8pOztbycnJKioq0v79+2O1XgDAAGFdQE1NTcrPz9fq1atP+/iqVav0i1/8Qs8+
+6y2b9+uIUOGaM6cOWptbb3gxQIABg7rFyEUFxeruLj4tI8ZY/Tkk0/qgQce0Lx58yRJzz33nDIzM/
Xaa6/
p1ltvvbDVAgAGjJg+B1RTU6O6ujoVFRV13RcMBlVQUKCKiorTZtra2hQOh7vdAAADX0wLqK6uTpKUmZnZ7f
7MzMyux76otLRUwWCw65abmxvLJQEA+ijnr4JbsWKFQqFQ1+3QoUOulwQA6AUxLaCsrCxJUn19fbf76+vru
x77Ir/
fr5SUlG43AMDAF9MCysvLU1ZWljZv3tx1Xzgc1vbt21VYWBjLXQEA+jnrV8GdOHFCBw4c6Pq4pqZGe/
bsUVpamkaNGqVly5bpscce0+WXX668vDw9+OCDysnJ0fz582O5bgBAP2ddQDt27ND111/
f9fHy5cslSQsXLtTatWt17733qqmpSXfeeacaGhp07bXXatOmTRo0aFDsVg0A6Pd8xhhvU/
p6SDgcVjAY1EzNU4LPfkAf+jifzz4Sbz980nTYD+6UpPhh9sM7b634s/
1+fPafdkc7AtaZ1Phm64wklTfYDyP9y/HTP897No+O/w/rzK7mS60zOUn2A0Ilb8fv/fZ068zl/tO/
SvhsfvdJvnVGknIHfWyd+f2yGVbbd3S0alvZIwqFQmd9Xt/
5q+AAABcnCggA4AQFBABwggICADhBAQEAnKCAAABOUEAAACcoIACAExQQAMAJCggA4AQFBABwggICADhBAQ
EAnLD+dQzABfEwfN2XYH+aep2Gfej2K6wzNwx+3Trzp9ZLrDMjEhqtMxFjP0lckrL9IetMILPVOtPQOdg6k
5ZwwjrT2JlsnZGkwXFt1hkv/
09fTjpmnbnnrS9bZyQpMPG4dSYl0e5aJXqe1zZcAQEAnKCAAABOUEAAACcoIACAExQQAMAJCggA4AQFBABw
ggICADhBAQEAnKCAAABOUEAAACcoIACAEwwjRa/yJSZZZ6Kt9kMuvUr/
c7t15lhnonUmNa7ZOpPk67TOtHscRvrVtBrrzFEPAz93teRZZwLxLdaZEXH2A0IlKTfRfnDnn1tzrTNvNF1
mnbn9v7xlnZGkF/
73160zSZv+ZLV9nImc33bWKwEAIAYoIACAExQQAMAJCggA4AQFBABwggICADhBAQEAnKCAAABOUEAAACcoI
ACAExQQAMAJCggA4MTFPYzU5/MWS7AfPumL99D1cfaZaGub/X6i9kMuvTIR+2Gfvel//
q9fWmcOdaRaZ+oi9pnUePsBpp3ydo5XtgStM4Pizm8A5eeNSAhbZ8JR+6GnXjVGB1lnIh4GwHo5dvcN32+d
kaRXQ0Wecj2BKyAAgBMUEADACQoIAOAEBQQAcIICAgA4QQEBAJyggAAATlBAAAAnKCAAgBMUEADACQoIAOA
EBQQAcGLADCP1Jdj/
U0xHh6d9eRmoaexnDQ5ILfOmWWcOzbcflnrbl96xzkhSXUfAOrO7+VLrTDC+xTozJM5+0GyrsR+cK0mH24d
ZZ7wM1ExLOGGdyfAwwLTTePte+6OI/
XHwwsug2Q877I+dJDX+10brTOpznnZ1TlwBAQCcoIAAAE5YF9DWrVt14403KicnRz6fT6+99lq3xxctWiSf
z9ftNnfu3FitFwAwQFgXUFNTk/Lz87V69eozbjN37lzV1tZ23V544YULWiQAYOCxfua+uLhYxcXFZ93G7/
crKyvL86IAAANfjzwHVFZWpoyMDI0fP15LlizR8ePHz7htW1ubwuFwtxsAYOCLeQHNnTtXzz33nDZv3qyf/
exnKi8vV3FxsTo7T/
9S2tLSUgWDwa5bbm5urJcEAOiDYv4+oFtvvbXrz5MmTdLkyZM1duxYlZWVadasWadsv2LFCi1fvrzr43A4T
AkBwEWgx1+GPWbMGKWnp+vAgQOnfdzv9yslJaXbDQAw8PV4AX344Yc6fvy4srOze3pXAIB+xPpHcCdOnOh2
NVNTU6M9e/YoLS1NaWlpeuSRR7RgwQJlZWWpurpa9957ry677DLNmTMnpgsHAPRv1gW0Y8cOXX/99V0ff/
r8zcKFC/XMM89o7969+s1vfqOGhgbl5ORo9uzZ+vGPfyy/
3x+7VQMA+j2fMca4XsTnhcNhBYNBzdQ8Jfi8DVLsixKy7d8XFcnLtM58fMVg60xzls86I0lXf+Ov1plFmdu
sM0c77Z8XTPR5GzTb2JlsnclKbLDObAldaZ0ZmmA/jNTL0FNJ+nLy+9aZhqj9uZeT8Il15r4D/
2SdyRxsP4BTkv599BvWmYiJWmeqIvbfoAfi7IciS9LbzZdZZ9ZfOcJq+w4TUZk2KBQKnfV5fWbBAQCcoIAA
AE5QQAAAJyggAIATFBAAwAkKCADgBAUEAHCCAgIAOEEBAQCcoIAAAE5QQAAAJyggAIATFBAAwImY/
0puV9qKp1pnMv7lH572dXXKh9aZK5Ptp0C3Ru2ngQ+Ki1hn3mu5xDojSc3RJOvM/nb7qeChDvspy/E+
+4nEknSkPWCd+beaIuvM5mnPWmceODzXOhOX7G3Y/fHOodaZBUPDHvZkf45/f9RW68yYpCPWGUna2GT/
izQPR4ZZZzITQ9aZSxOPWmck6ebA360z62U3Dft8cQUEAHCCAgIAOEEBAQCcoIAAAE5QQAAAJyggAIATFBA
AwAkKCADgBAUEAHCCAgIAOEEBAQCcoIAAAE702WGkvoQE+Xznv7yCf33Xeh+zAn+xzkhSs/
FbZ7wMFvUy1NCLYEKzp1xbxP70ORJJ8bQvW+P8dZ5yN6Xssc5s/
WWBdeba1rutM9U3rLHObG6Jt85I0tEO+/
+nW2tusM7sOphrnbnm0hrrzKTAR9YZydsg3EB8q3Um0ddhnWmK2n8dkqTKVvtBsz2FKyAAgBMUEADACQoIA
OAEBQQAcIICAgA4QQEBAJyggAAATlBAAAAnKCAAgBMUEADACQoIAOAEBQQAcKLPDiOtXTJF8f5B5739w8Gn
rPex7uNrrDOSlDvoY+vM6KRj1pn85A+sM14E4uyHJ0rS+BT7AYobm0ZaZ8oaJlhnshMbrDOS9HbzWOvMiw8
/bp1ZdM8PrTOFbyy2zoQv9fY9ZscQY51JyT9unXngS/9pnUnydVpnGjrth4pKUpq/
yTqTGu9tuK8tL0ORJSkQ12KdiR9/mdX2prNN2n/
u7bgCAgA4QQEBAJyggAAATlBAAAAnKCAAgBMUEADACQoIAOAEBQQAcIICAgA4QQEBAJyggAAATlBAAAAn+u
ww0sFHoopPip739hvDV1vvY0zyUeuMJB2LBKwz/
+fEJOvMyORPrDPBePtBg5f566wzkrSnNdU6s+noVdaZnOSwdaY+ErTOSNLxyBDrTHPUfijkr37+hHXm3+qL
rDM3pe2yzkhSfpL9YNGGqP33s++1Z1lnGqPnP6T4U60m0TojSSEPQ0wDHj4HI8b+S3G8Of+vj5+XGmc/
LDU8abjV9h2RVoaRAgD6LgoIAOCEVQGVlpZq6tSpCgQCysjI0Pz581VVVdVtm9bWVpWUlGj48OEaOnSoFix
YoPr6+pguGgDQ/
1kVUHl5uUpKSlRZWak333xTkUhEs2fPVlPTZ7+06Z577tHrr7+uV155ReXl5Tp8+LBuvvnmmC8cANC/
WT3ztWnTpm4fr127VhkZGdq5c6dmzJihUCikX/
3qV1q3bp1uuOEGSdKaNWt0xRVXqLKyUtdc4+03kAIABp4Leg4oFApJktLS0iRJO3fuVCQSUVHRZ6/
WmTBhgkaNGqWKiorT/h1tbW0Kh8PdbgCAgc9zAUWjUS1btkzTp0/
XxIkTJUl1dXVKSkpSampqt20zMzNVV3f6l/
qWlpYqGAx23XJzc70uCQDQj3guoJKSEu3bt08vvvjiBS1gxYoVCoVCXbdDhw5d0N8HAOgfPL0RdenSpdq4c
aO2bt2qkSNHdt2flZWl9vZ2NTQ0dLsKqq+vV1bW6d9w5vf75ffbv5EPANC/WV0BGWO0dOlSrV+/
Xlu2bFFeXl63x6dMmaLExERt3ry5676qqiodPHhQhYWFsVkxAGBAsLoCKikp0bp167RhwwYFAoGu53WCwaC
Sk5MVDAZ1++23a/ny5UpLS1NKSoruvvtuFRYW8go4AEA3VgX0zDPPSJJmzpzZ7f41a9Zo0aJFkqSf//
zniouL04IFC9TW1qY5c+bo6aefjsliAQADh88YY1wv4vPC4bCCwaBmXPugEhLOf+jg1Cd3Wu9rXzjHOiNJm
YMarTOTh35onalqth/UeLglxTozOCFinZGk5Hj7XIexf91Lht/+eI/y2w/
TlKRAnP0gySRfp3Wm08Prf65KOmydOdgxzDojSXUdqdaZ95rtP5+GJdgPxvyzh8/
b5o4k64wktXXaP03e2mGfCfpbrTNT0z6wzkhSnOy/5K/7j+usto+2tuofj/2LQqGQUlLO/
DWJWXAAACcoIACAExQQAMAJCggA4AQFBABwggICADhBAQEAnKCAAABOUEAAACcoIACAExQQAMAJCggA4AQF
BABwwtNvRO0Ncdv2Ks6XeN7bv/L76db7eHDeK9YZSSpvmGCd2Vg3yToTbrf/
TbEjBjdZZ1IS7adNS1Jaov2+gh6mHw/ydVhnPukYYp2RpLa48z/
nPtUpn3Wmri1onflj9HLrTCQab52RpDYPOS/T0T9uT7fO5CSHrDONHec/Wf/
z3m9Ms84cCw21zrQOtv9SvK1zrHVGkuZm/cU6k3zE7hzvbDu/
7bkCAgA4QQEBAJyggAAATlBAAAAnKCAAgBMUEADACQoIAOAEBQQAcIICAgA4QQEBAJyggAAATlBAAAAnfMY
Y43oRnxcOhxUMBjVT85RgMYzUi9Bt13jKjbmryjozLbXGOrMrPMo6c9DD8MRI1Nv3IYlxUevM4MR268wgD0
Muk+I7rTOSFCf7T4eoh2GkQ+Ltj8OQhDbrTEpCq3VGkgLx9rk4n/354EW8h/+jd0KXxn4hZxDw8P/UYew/
BwuD1dYZSfp1zVetM8FvHLDavsNEVKYNCoVCSklJOeN2XAEBAJyggAAATlBAAAAnKCAAgBMUEADACQoIAOA
EBQQAcIICAgA4QQEBAJyggAAATlBAAAAnKCAAgBN9dxhp3M12w0ij3oZP9pamBQXWmYL737XPBOwHFE5Iqr
fOSFKi7IdPDvIwsHJInP2wz1aPp7WX78i2teRaZzo97GnLJ1dYZyIehlxKUn3zmQdInkmixwGwtqLG/
nxo6fA22DjUMsg6Ex9nf+61lqVbZ4a/Zz+kV5L8b9h/
XbHFMFIAQJ9GAQEAnKCAAABOUEAAACcoIACAExQQAMAJCggA4AQFBABwggICADhBAQEAnKCAAABOUEAAACf
67jBSzbMbRgrPfFMnecq1ZCVbZ/zH26wzjaPt95NS3WSdkaS4tg7rTPT//
tXTvoCBimGkAIA+jQICADhhVUClpaWaOnWqAoGAMjIyNH/
+fFVVVXXbZubMmfL5fN1uixcvjumiAQD9n1UBlZeXq6SkRJWVlXrzzTcViUQ0e/
ZsNTV1/3n7HXfcodra2q7bqlWrYrpoAED/
l2Cz8aZNm7p9vHbtWmVkZGjnzp2aMWNG1/2DBw9WVlZWbFYIABiQLug5oFAoJElKS0vrdv/
zzz+v9PR0TZw4UStWrFBzc/MZ/462tjaFw+FuNwDAwGd1BfR50WhUy5Yt0/Tp0zVx4sSu+7/
zne9o9OjRysnJ0d69e3XfffepqqpKr7766mn/
ntLSUj3yyCNelwEA6Kc8vw9oyZIl+t3vfqdt27Zp5MiRZ9xuy5YtmjVrlg4cOKCxY8ee8nhbW5va2j57b0g
4HFZubi7vA+pFvA/oM7wPCLhw5/s+IE9XQEuXLtXGjRu1devWs5aPJBUUFEjSGQvI7/fL7/
d7WQYAoB+zKiBjjO6+
+26tX79eZWVlysvLO2dmz549kqTs7GxPCwQADExWBVRSUqJ169Zpw4YNCgQCqqurkyQFg0ElJyerurpa69a
t0ze+8Q0NHz5ce/
fu1T333KMZM2Zo8uTJPfIPAAD0T1YF9Mwzz0g6+WbTz1uzZo0WLVqkpKQkvfXWW3ryySfV1NSk3NxcLViwQ
A888EDMFgwAGBisfwR3Nrm5uSovL7+gBQEALg6eX4aNgcO8+2dPuUExXseZpPypl3YkKdp7uwIuegwjBQA4
QQEBAJyggAAATlB
AAAAnKCAAgBMUEADACQoIAOAEBQQAcIICAgA4QQEBAJyggAAATlBAAAAnKCAAgBMUEADACQoIAOAEBQQAcI
ICAgA4QQEBAJyggAAATlBAAAAnKCAAgBMUEADACQoIAOAEBQQAcCLB9QK+yBgjSepQRDKOFwMAsNahiKTPv
p6fSZ8roMbGRknSNr3heCUAgAvR2NioYDB4xsd95lwV1cui0agOHz6sQCAgn8/
X7bFwOKzc3FwdOnRIKSkpjlboHsfhJI7DSRyHkzgOJ/
WF42CMUWNjo3JychQXd+ZnevrcFVBcXJxGjhx51m1SUlIu6hPsUxyHkzgOJ3EcTuI4nOT6OJztyudTvAgBA
OAEBQQAcKJfFZDf79fKlSvl9/tdL8UpjsNJHIeTOA4ncRxO6k/
Hoc+9CAEAcHHoV1dAAICBgwICADhBAQEAnKCAAABOUEAAACf6TQGtXr1al156qQYNGqSCggK98847rpfU6x
5++GH5fL5utwkTJrheVo/bunWrbrzxRuXk5Mjn8+m1117r9rgxRg899JCys7OVnJysoqIi7d+/
381ie9C5jsOiRYtOOT/
mzp3rZrE9pLS0VFOnTlUgEFBGRobmz5+vqqqqbtu0traqpKREw4cP19ChQ7VgwQLV19c7WnHPOJ/
jMHPmzFPOh8WLFzta8en1iwJ66aWXtHz5cq1cuVK7du1Sfn6+5syZoyNHjrheWq+76qqrVFtb23Xbtm2b6y
X1uKamJuXn52v16tWnfXzVqlX6xS9+oWeffVbbt2/XkCFDNGfOHLW2tvbySnvWuY6DJM2dO7fb+fHCCy/
04gp7Xnl5uUpKSlRZWak333xTkUhEs2fPVlNTU9c299xzj15//
XW98sorKi8v1+HDh3XzzTc7XHXsnc9xkKQ77rij2/
mwatUqRys+A9MPTJs2zZSUlHR93NnZaXJyckxpaanDVfW+lStXmvz8fNfLcEqSWb9+fdfH0WjUZGVlmccff
7zrvoaGBuP3+80LL7zgYIW944vHwRhjFi5caObNm+dkPa4cOXLESDLl5eXGmJP/94mJieaVV17p2uavf/
2rkWQqKipcLbPHffE4GGPMddddZ37wgx+4W9R56PNXQO3t7dq5c6eKioq67ouLi1NRUZEqKiocrsyN/
fv3KycnR2PGjNFtt92mgwcPul6SUzU1Naqrq+t2fgSDQRUUFFyU50dZWZkyMjI0fvx4LVmyRMePH3e9pB4V
CoUkSWlpaZKknTt3KhKJdDsfJkyYoFGjRg3o8+GLx+FTzz//
vNLT0zVx4kStWLFCzc3NLpZ3Rn1uGvYXHTt2TJ2dncrMzOx2f2Zmpv72t785WpUbBQUFWrt2rcaPH6/
a2lo98sgj+trXvqZ9+/YpEAi4Xp4TdXV1knTa8+PTxy4Wc+fO1c0336y8vDxVV1fr/
vvvV3FxsSoqKhQfH+96eTEXjUa1bNkyTZ8+XRMnTpR08nxISkpSampqt20H8vlwuuMgSd/
5znc0evRo5eTkaO/evbrvvvtUVVWlV1991eFqu+vzBYTPFBcXd/158uTJKigo0OjRo/Xyyy/
r9ttvd7gy9AW33npr158nTZqkyZMna+zYsSorK9OsWbMcrqxnlJSUaN+
+fRfF86Bnc6bjcOedd3b9edKkScrOztasWbNUXV2tsWPH9vYyT6vP/
wguPT1d8fHxp7yKpb6+XllZWY5W1TekpqZq3LhxOnDggOulOPPpOcD5caoxY8YoPT19QJ4fS5cu1caNG/
WHP/
yh2+8Py8rKUnt7uxoaGrptP1DPhzMdh9MpKCiQpD51PvT5AkpKStKUKVO0efPmrvui0ag2b96swsJChytz7
8SJE6qurlZ2drbrpTiTl5enrKysbudHOBzW9u3bL/
rz48MPP9Tx48cH1PlhjNHSpUu1fv16bdmyRXl5ed0enzJlihITE7udD1VVVTp48OCAOh/
OdRxOZ8+ePZLUt84H16+COB8vvvii8fv9Zu3atea9994zd955p0lNTTV1dXWul9arfvjDH5qysjJTU1Nj/
vjHP5qioiKTnp5ujhw54nppPaqxsdHs3r3b7N6920gyTzzxhNm9e7f54IMPjDHG/PSnPzWpqalmw4YNZu/
evWbevHkmLy/PtLS0OF55bJ3tODQ2Npof/
ehHpqKiwtTU1Ji33nrLfPnLXzaXX365aW1tdb30mFmyZIkJBoOmrKzM1NbWdt2am5u7tlm8eLEZNWqU2bJl
i9mxY4cpLCw0hYWFDlcde+c6DgcOHDCPPvqo2bFjh6mpqTEbNmwwY8aMMTNmzHC88u76RQEZY8xTTz1lRo0
aZZKSksy0adNMZWWl6yX1ultuucVkZ2ebpKQkc8kll5hbbrnFHDhwwPWyetwf/vAHI+mU28KFC40xJ1+K/
eCDD5rMzEzj9/vNrFmzTFVVldtF94CzHYfm5mYze/ZsM2LECJOYmGhGjx5t7rjjjgH3Tdrp/
v2SzJo1a7q2aWlpMXfddZcZNmyYGTx4sLnppptMbW2tu0X3gHMdh4MHD5oZM2aYtLQ04/f7zWWXXWb+
+Z//2YRCIbcL/wJ+HxAAwIk+/
xwQAGBgooAAAE5QQAAAJyggAIATFBAAwAkKCADgBAUEAHCCAgIAOEEBAQCcoIAAAE5QQAAAJ/
4fSFZm765APLcAAAAASUVORK5CYII=",
"text/plain": [
"<Figure size 640x480 with 1 Axes>"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"import matplotlib.pyplot as plt\n",
"image, label = train_data[0]\n",
"print(f\"Image shape: {image.shape}\")\n",
"plt.imshow(image.squeeze()) # image shape is [1, 28, 28] (colour channels,
height, width)\n",
"plt.title(label);"
]
},
{
"cell_type": "markdown",
"id": "adb19c5c-2f2b-4aaf-8300-256f3594e2db",
"metadata": {
"id": "adb19c5c-2f2b-4aaf-8300-256f3594e2db"
},
"source": [
"We can turn the image into grayscale using the `cmap` parameter of
`plt.imshow()`."
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "92f09917-88f7-4446-b65f-baae586914c9",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 452
},
"id": "92f09917-88f7-4446-b65f-baae586914c9",
 "outputId": "c702456b-607c-4214-8e03-4bd0b22b097f"
},
"outputs": [
{
"data": {
"image/png":
"iVBORw0KGgoAAAANSUhEUgAAAaAAAAGzCAYAAABpdMNsAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIH
ZlcnNpb24zLjcuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/
bCgiHAAAACXBIWXMAAA9hAAAPYQGoP6dpAAAomElEQVR4nO3de3RV5Z3G8eckJIdAksMl5FYCCTdh5KKDEC
Nyj0C0DBSseFmzoINamdAW0LGLmVbqtGtSsWNZVCq20wXWiSLO4lJdSoeLhCogBWHQGWUIBgFDwqXmJCTkQ
vLOHyzPeLiFd5vkTcL3s9ZecvZ5f9kvLzt53Dn7/
I7PGGMEAEALi3A9AQDAjYkAAgA4QQABAJwggAAAThBAAAAnCCAAgBMEEADACQIIAOAEAQQAcIIAAhoxZ84c
xcbGNjpu3LhxGjduXJMdd9y4cRo8eHCTfT2gtSGA0C79+te/ls/nU2ZmpuuptEn/8i//
og0bNrieBto5AgjtUn5+vtLT07Vnzx4VFha6nk6bQwChJRBAaHeKioq0c+dOPffcc+rRo4fy8/
NdTwnAFRBAaHfy8/PVtWtX3XPPPbr33nuvGEBHjx6Vz+fTL37xC/
3mN79R37595ff7NWLECP35z39u9BgHDhxQjx49NG7cOJ07d+6q42pqarRkyRL169dPfr9faWlpevLJJ1VTU
3Pdf599+/bpjjvuUExMjDIyMrRy5crLxpw6dUpz585VUlKSOnbsqGHDhumll166bFxlZaUef/xxpaWlye/
366abbtIvfvELfbUpvs/
nU2VlpV566SX5fD75fD7NmTPnuucLXDcDtDMDBw40c+fONcYYs2PHDiPJ7NmzJ2xMUVGRkWRuvfVW069fP/
PMM8+YpUuXmoSEBNOzZ09TW1sbGjt79mzTuXPn0OM9e/
aYrl27mrvuustUVVWF9o8dO9aMHTs29Li+vt5MmjTJdOrUySxYsMC8+OKLZv78+aZDhw5m2rRpjf49xo4da
1JTU01iYqKZP3+
+Wb58ubnzzjuNJPO73/0uNK6qqsoMGjTIREVFmYULF5rly5eb0aNHG0lm2bJloXENDQ1mwoQJxufzmYcfft
g8//zzZurUqUaSWbBgQWjcyy+/bPx+vxk9erR5+eWXzcsvv2x27tzZ+MIDlgggtCt79+41kszmzZuNMRd/
6Pbs2dP84Ac/CBv3ZQB1797d/OUvfwnt37hxo5Fk3njjjdC+rwbQu++
+a+Lj480999xjqqurw77mpQH08ssvm4iICPOnP/
0pbNzKlSuNJPPee+9d8+8yduxYI8n867/+a2hfTU2NueWWW0xiYmIoJJctW2YkmX//
938PjautrTVZWVkmNjbWlJeXG2OM2bBhg5Fkfvazn4Ud59577zU+n88UFhaG9nXu3NnMnj37mvMDvi5+BYd
2JT8/X0lJSRo/
fryki79OmjVrltasWaP6+vrLxs+aNUtdu3YNPR49erQk6dNPP71s7DvvvKPJkydr4sSJWrdunfx+/
zXn8vrrr2vQoEEaOHCgzpw5E9omTJgQ+nqN6dChg7773e+GHkdHR+u73/2uTp06pX379kmS3nrrLSUnJ+uB
Bx4IjYuKitL3v/
99nTt3TgUFBaFxkZGR+v73vx92jMcff1zGGL399tuNzgdoSgQQ2o36+nqtWbNG48ePV1FRkQoLC1VYWKjMz
EyVlpZq69atl9X06tUr7PGXYfTFF1+E7a+urtY999yjW2+9VWvXrlV0dHSj8zl8+LD++7//
Wz169AjbBgwYIOni6zaNSU1NVefOncP2fVl/9OhRSdJnn32m/v37KyIi/Nt50KBBoee//
G9qaqri4uKuOQ5oKR1cTwBoKtu2bdPJkye1Zs0arVmz5rLn8/PzNWnSpLB9kZGRV/xa5pJPqvf7/
br77ru1ceNGbdq0Sd/
85jcbnU9DQ4OGDBmi55577orPp6WlNfo1gPaMAEK7kZ+fr8TERK1YseKy59atW6f169dr5cqViomJsf7aPp
9P+fn5mjZtmr797W/r7bffbrTrQd++ffVf//Vfmjhxonw+n/UxJam4uFiVlZVhV0H/+7//
K0lKT0+XJPXu3VsHDx5UQ0ND2FXQJ598Enr+y/9u2bJFFRUVYVdBl4778u8LNDd+BYd24fz581q3bp2+
+c1v6t57771smz9/vioqKvSHP/zB8zGio6O1bt06jRgxQlOnTtWePXuuOf6+++7T559/rt/
+9rdXnG9lZWWjx7xw4YJefPHF0OPa2lq9+OKL6tGjh4YPHy5Juvvuu1VSUqLXXnstrO5Xv/
qVYmNjNXbs2NC4+vp6Pf/882HH+OUvfymfz6ecnJzQvs6dO6usrKzR+QFfB1dAaBf+8Ic/qKKiQn/
zN39zxedvv/320JtSZ82a5fk4MTExevPNNzVhwgTl5OSooKDgqv3a/vZv/
1Zr167VY489pnfeeUejRo1SfX29PvnkE61du1Z//
OMfddttt13zeKmpqXrmmWd09OhRDRgwQK+99poOHDig3/
zmN4qKipIkPfroo3rxxRc1Z84c7du3T+np6fqP//gPvffee1q2bFnoamfq1KkaP368/umf/
klHjx7VsGHD9J//+Z/
auHGjFixYoL59+4aOO3z4cG3ZskXPPfecUlNTlZGRQVsjND3Xt+EBTWHq1KmmY8eOprKy8qpj5syZY6Kios
yZM2dCt2E/++yzl42TZJYsWRJ6fOn7gIwx5syZM+av/uqvTHJysjl8+LAx5vLbsI25eDv0M888Y26+
+Wbj9/tN165dzfDhw83TTz9tgsHgNf9OY8eONTfffLPZu3evycrKMh07djS9e/c2zz///
GVjS0tLzXe+8x2TkJBgoqOjzZAhQ8yqVasuG1dRUWEWLlxoUlNTTVRUlOnfv7959tlnTUNDQ9i4Tz75xIwZ
M8bExMQYSdySjWbhM+aSV1sBAGgBvAYEAHCCAAIAOEEAAQCcIIAAAE4QQAAAJwggAIATre6NqA0NDSouLlZ
cXBztQACgDTLGqKKiQqmpqZc1yf2qVhdAxcXFNGkEgHbg+PHj6tmz51Wfb3W/
gru0VTwAoG1q7Od5swXQihUrlJ6ero4dOyozM7PRxo1f4tduANA+NPbzvFkC6LXXXtOiRYu0ZMkSffDBBxo
2bJgmT558XR/
ABQC4QTRHg7mRI0ea3Nzc0OP6+nqTmppq8vLyGq0NBoNGEhsbGxtbG98aa7jb5FdAtbW12rdvn7Kzs0P7Ii
IilJ2drV27dl02vqamRuXl5WEbAKD9a/
IAOnPmjOrr65WUlBS2PykpSSUlJZeNz8vLUyAQCG3cAQcANwbnd8EtXrxYwWAwtB0/
ftz1lAAALaDJ3weUkJCgyMhIlZaWhu0vLS1VcnLyZeP9fr/
8fn9TTwMA0Mo1+RVQdHS0hg8frq1bt4b2NTQ0aOvWrcrKymrqwwEA2qhm6YSwaNEizZ49W7fddptGjhypZc
uWqbKyUt/
5znea43AAgDaoWQJo1qxZOn36tJ566imVlJTolltu0aZNmy67MQEAcOPyGWOM60l8VXl5uQKBgOtpAAC+pm
AwqPj4+Ks+7/
wuOADAjYkAAgA4QQABAJwggAAAThBAAAAnCCAAgBMEEADACQIIAOAEAQQAcIIAAgA4QQABAJwggAAAThBAA
AAnCCAAgBMEEADACQIIAOAEAQQAcIIAAgA4QQABAJwggAAAThBAAAAnCCAAgBMEEADACQIIAOAEAQQAcIIA
AgA40cH1BIDWxOfzWdcYY5phJpeLi4uzrrnzzjs9Hevtt9/2VGfLy3pHRkZa11y4cMG6prXzsnZeNdc5zhU
QAMAJAggA4AQBBABwggACADhBAAEAnCCAAABOEEAAACcIIACAEwQQAMAJAggA4AQBBABwggACADhBM1LgKy
Ii7P+frL6+3rqmX79+1jUPP/ywdc358+etaySpsrLSuqa6utq6Zs+ePdY1LdlY1EvDTy/
nkJfjtOQ62DaANcaooaGh0XFcAQEAnCCAAABOEEAAACcIIACAEwQQAMAJAggA4AQBBABwggACADhBAAEAnC
CAAABOEEAAACcIIACAEzQjBb7Ctumi5K0Z6YQJE6xrsrOzrWtOnDhhXSNJfr/fuqZTp07WNXfddZd1zb/
9279Z15SWllrXSBebatrycj54ERsb66nuepqEXqqqqsrTsRrDFRAAwAkCCADgRJMH0E9+8hP5fL6wbeDAgU
19GABAG9csrwHdfPPN2rJly/
8fpAMvNQEAwjVLMnTo0EHJycnN8aUBAO1Es7wGdPjwYaWmpqpPnz566KGHdOzYsauOrampUXl5edgGAGj/
mjyAMjMztXr1am3atEkvvPCCioqKNHr0aFVUVFxxfF5engKBQGhLS0tr6ikBAFqhJg+gnJwcffvb39bQoUM
1efJkvfXWWyorK9PatWuvOH7x4sUKBoOh7fjx4009JQBAK9Tsdwd06dJFAwYMUGFh4RWf9/
v9nt70BgBo25r9fUDnzp3TkSNHlJKS0tyHAgC0IU0eQE888YQKCgp09OhR7dy5U9/61rcUGRmpBx54oKkPB
QBow5r8V3AnTpzQAw88oLNnz6pHjx668847tXv3bvXo0aOpDwUAaMOaPIDWrFnT1F8SaDG1tbUtcpwRI0ZY
16Snp1vXeGmuKkkREfa/HPnjH/9oXXPrrbda1yxdutS6Zu/
evdY1kvThhx9a13z88cfWNSNHjrSu8XIOSdLOnTuta3bt2mU13hhzXW+poRccAMAJAggA4AQBBABwggACAD
hBAAEAnCCAAABOEEAAACcIIACAEwQQAMAJAggA4AQBBABwggACADjR7B9IB7jg8/
k81RljrGvuuusu65rbbrvNuuZqH2t/LZ07d7aukaQBAwa0SM2f//
xn65qrfbjltcTGxlrXSFJWVpZ1zYwZM6xr6urqrGu8rJ0kPfzww9Y1NTU1VuMvXLigP/
3pT42O4woIAOAEAQQAcIIAAgA4QQABAJwggAAAThBAAAAnCCAAgBMEEADACQIIAOAEAQQAcIIAAgA4QQABA
JwggAAATviMl/a/
zai8vFyBQMD1NNBMvHapbilevh12795tXZOenm5d44XX9b5w4YJ1TW1tradj2aqurrauaWho8HSsDz74wLr
GS7duL+s9ZcoU6xpJ6tOnj3XNN77xDU/
HCgaDio+Pv+rzXAEBAJwggAAAThBAAAAnCCAAgBMEEADACQIIAOAEAQQAcIIAAgA4QQABAJwggAAAThBAAA
AnCCAAgBMdXE8AN5ZW1vu2SXzxxRfWNSkpKdY158+ft67x+/
3WNZLUoYP9j4bY2FjrGi+NRWNiYqxrvDYjHT16tHXNHXfcYV0TEWF/
LZCYmGhdI0mbNm3yVNccuAICADhBAAEAnCCAAABOEEAAACcIIACAEwQQAMAJAggA4AQBBABwggACADhBAAE
AnCCAAABOEEAAACdoRgp8TZ06dbKu8dJ80ktNVVWVdY0kBYNB65qzZ89a16Snp1vXeGlo6/
P5rGskb2vu5Xyor6+3rvHaYDUtLc1TXXPgCggA4AQBBABwwjqAduzYoalTpyo1NVU+n08bNmwIe94Yo6eee
kopKSmKiYlRdna2Dh8+3FTzBQC0E9YBVFlZqWHDhmnFihVXfH7p0qVavny5Vq5cqffff1+dO3fW5MmTPX3w
FACg/bK+CSEnJ0c5OTlXfM4Yo2XLlulHP/qRpk2bJkn6/e9/
r6SkJG3YsEH333//15stAKDdaNLXgIqKilRSUqLs7OzQvkAgoMzMTO3ateuKNTU1NSovLw/
bAADtX5MGUElJiSQpKSkpbH9SUlLouUvl5eUpEAiEttZ0iyAAoPk4vwtu8eLFCgaDoe348eOupwQAaAFNGk
DJycmSpNLS0rD9paWloecu5ff7FR8fH7YBANq/Jg2gjIwMJScna+vWraF95eXlev/
995WVldWUhwIAtHHWd8GdO3dOhYWFocdFRUU6cOCAunXrpl69emnBggX62c9+pv79+ysjI0M//
vGPlZqaqunTpzflvAEAbZx1AO3du1fjx48PPV60aJEkafbs2Vq9erWefPJJVVZW6tFHH1VZWZnuvPNObdq0
SR07dmy6WQMA2jyf8dLZrxmVl5crEAi4ngaaiZemkF4aQnpp7ihJsbGx1jX79+
+3rvGyDufPn7eu8fv91jWSVFxcbF1z6Wu/1+OOO+6wrvHS9NRLg1BJio6Otq6pqKiwrvHyM8/
rDVtezvG5c+daja+vr9f+/
fsVDAav+bq+87vgAAA3JgIIAOAEAQQAcIIAAgA4QQABAJwggAAAThBAAAAnCCAAgBMEEADACQIIAOAEAQQA
cIIAAgA4QQABAJyw/
jgG4Ovw0nw9MjLSusZrN+xZs2ZZ11zt036v5fTp09Y1MTEx1jUNDQ3WNZLUuXNn65q0tDTrmtraWusaLx2+
6+rqrGskqUMH+x+RXv6dunfvbl2zYsUK6xpJuuWWW6xrvKzD9eAKCADgBAEEAHCCAAIAOEEAAQCcIIAAAE4
QQAAAJwggAIATBBAAwAkCCADgBAEEAHCCAAIAOEEAAQCcoBkpWpSXpoZeGlZ69dFHH1nX1NTUWNdERUVZ17
RkU9bExETrmurqauuas2fPWtd4WbuOHTta10jemrJ+8cUX1jUnTpywrnnwwQetayTp2Wefta7ZvXu3p2M1h
isgAIATBBAAwAkCCADgBAEEAHCCAAIAOEEAAQCcIIAAAE4QQAAAJwggAIATBBAAwAkCCADgBAEEAHDihm5G
6vP5PNV5aQoZEWGf9V7mV1dXZ13T0NBgXePVhQsXWuxYXrz11lvWNZWVldY158+ft66Jjo62rjHGWNdI0un
Tp61rvHxfeGkS6uUc96qlvp+8rN3QoUOtayQpGAx6qmsOXAEBAJwggAAAThBAAAAnCCAAgBMEEADACQIIAO
AEAQQAcIIAAgA4QQABAJwggAAAThBAAAAnCCAAgBPtphmpl2Z+9fX1no7V2htqtmZjxoyxrpk5c6Z1zahRo
6xrJKmqqsq65uzZs9Y1XhqLduhg/+3q9Rz3sg5evgf9fr91jZcGpl6bsnpZBy+8nA/
nzp3zdKwZM2ZY17zxxhuejtUYroAAAE4QQAAAJ6wDaMeOHZo6dapSU1Pl8/
m0YcOGsOfnzJkjn88Xtk2ZMqWp5gsAaCesA6iyslLDhg3TihUrrjpmypQpOnnyZGh79dVXv9YkAQDtj/
Wrmjk5OcrJybnmGL/
fr+TkZM+TAgC0f83yGtD27duVmJiom266SfPmzbvmXUI1NTUqLy8P2wAA7V+TB9CUKVP0+9//
Xlu3btUzzzyjgoIC5eTkXPV20Ly8PAUCgdCWlpbW1FMCALRCTf4+oPvvvz/05yFDhmjo0KHq27evtm/
frokTJ142fvHixVq0aFHocXl5OSEEADeAZr8Nu0+fPkpISFBhYeEVn/f7/YqPjw/
bAADtX7MH0IkTJ3T27FmlpKQ096EAAG2I9a/
gzp07F3Y1U1RUpAMHDqhbt27q1q2bnn76ac2cOVPJyck6cuSInnzySfXr10+TJ09u0okDANo26wDau3evxo
8fH3r85es3s2fP1gsvvKCDBw/qpZdeUllZmVJTUzVp0iT99Kc/
9dTzCQDQfvmM1y59zaS8vFyBQMD1NJpct27drGtSU1Ota/
r3798ix5G8NTUcMGCAdU1NTY11TUSEt98u19XVWdfExMRY1xQXF1vXREVFWdd4aXIpSd27d7euqa2tta7p1
KmTdc3OnTuta2JjY61rJG/
NcxsaGqxrgsGgdY2X80GSSktLrWsGDRrk6VjBYPCar+vTCw4A4AQBBABwggACADhBAAEAnCCAAABOEEAAAC
cIIACAEwQQAMAJAggA4AQBBABwggACADhBAAEAnCCAAABONPlHcrty++23W9f89Kc/
9XSsHj16WNd06dLFuqa+vt66JjIy0rqmrKzMukaSLly4YF1TUVFhXeOly7LP57OukaTz589b13jpznzfffd
Z1+zdu9e6Ji4uzrpG8taBPD093dOxbA0ZMsS6xus6HD9+3LqmqqrKusZLR3WvHb579+7tqa45cAUEAHCCAA
IAOEEAAQCcIIAAAE4QQAAAJwggAIATBBAAwAkCCADgBAEEAHCCAAIAOEEAAQCcIIAAAE602makERERVg0ll
y9fbn2MlJQU6xrJW5NQLzVemhp6ER0d7anOy9/JS7NPLwKBgKc6L40af/
7zn1vXeFmHefPmWdcUFxdb10hSdXW1dc3WrVutaz799FPrmv79+1vXdO/
e3bpG8tYINyoqyromIsL+WqCurs66RpJOnz7tqa45cAUEAHCCAAIAOEEAAQCcIIAAAE4QQAAAJwggAIATBB
AAwAkCCADgBAEEAHCCAAIAOEEAAQCcIIAAAE74jDHG9SS+qry8XIFAQA899JBVk0wvDSGPHDliXSNJsbGxL
VLj9/uta7zw0jxR8tbw8/
jx49Y1Xhpq9ujRw7pG8tYUMjk52bpm+vTp1jUdO3a0rklPT7eukbydr8OHD2+RGi//
Rl6aino9ltfmvrZsmjV/lZfv99tvv91qfENDgz7//
HMFg0HFx8dfdRxXQAAAJwggAIATBBAAwAkCCADgBAEEAHCCAAIAOEEAAQCcIIAAAE4QQAAAJwggAIATBBAA
wAkCCADgRAfXE7ia06dPWzXN89LkMi4uzrpGkmpqaqxrvMzPS0NIL40Qr9Us8Fr+8pe/
WNd89tln1jVe1uH8+fPWNZJUXV1tXXPhwgXrmvXr11vXfPjhh9Y1XpuRduvWzbrGS8PPsrIy65q6ujrrGi/
/
RtLFppq2vDT79HIcr81IvfyMGDBggNX4Cxcu6PPPP290HFdAAAAnCCAAgBNWAZSXl6cRI0YoLi5OiYmJmj5
9ug4dOhQ2prq6Wrm5uerevbtiY2M1c+ZMlZaWNumkAQBtn1UAFRQUKDc3V7t379bmzZtVV1enSZMmqbKyMj
Rm4cKFeuONN/T666+roKBAxcXFmjFjRpNPHADQtlndhLBp06awx6tXr1ZiYqL27dunMWPGKBgM6ne/
+51eeeUVTZgwQZK0atUqDRo0SLt377b+VD0AQPv1tV4DCgaDkv7/jpl9+/
aprq5O2dnZoTEDBw5Ur169tGvXrit+jZqaGpWXl4dtAID2z3MANTQ0aMGCBRo1apQGDx4sSSopKVF0dLS6d
OkSNjYpKUklJSVX/
Dp5eXkKBAKhLS0tzeuUAABtiOcAys3N1UcffaQ1a9Z8rQksXrxYwWAwtHl5vwwAoO3x9EbU+fPn680339SO
HTvUs2fP0P7k5GTV1taqrKws7CqotLRUycnJV/xafr9ffr/fyzQAAG2Y1RWQMUbz58/
X+vXrtW3bNmVkZIQ9P3z4cEVFRWnr1q2hfYcOHdKxY8eUlZXVNDMGALQLVldAubm5euWVV7Rx40bFxcWFXt
cJBAKKiYlRIBDQ3LlztWjRInXr1k3x8fH63ve+p6ysLO6AAwCEsQqgF154QZI0bty4sP2rVq3SnDlzJEm//
OUvFRERoZkzZ6qmpkaTJ0/Wr3/96yaZLACg/fAZY4zrSXxVeXm5AoGAhgwZosjIyOuu+
+1vf2t9rDNnzljXSFLnzp2ta7p3725d46VR47lz56xrvDRPlKQOHexfQvTSdLFTp07WNV4amEre1iIiwv5e
Hi/fdpfeXXo9vvomcRtemrl+8cUX1jVeXv/
18n3rpYGp5K2JqZdjxcTEWNdc7XX1xnhpYpqfn281vqamRs8//7yCweA1mx3TCw4A4AQBBABwggACADhBAA
EAnCCAAABOEEAAACcIIACAEwQQAMAJAggA4AQBBABwggACADhBAAEAnCCAAABOePpE1Jbw4YcfWo1ft26d9
TH+7u/+zrpGkoqLi61rPv30U+ua6upq6xovXaC9dsP20sE3OjrausamK/qXampqrGskqb6+3rr
GS2frqqoq65qTJ09a13htdu9lHbx0R2+pc7y2tta6RvLWkd5LjZcO2l46dUu67INEr0dpaanV+Otdb66AAA
BOEEAAACcIIACAEwQQAMAJAggA4AQBBABwggACADhBAAEAnCCAAABOEEAAACcIIACAEwQQAMAJn/
HarbCZlJeXKxAItMixcnJyPNU98cQT1jWJiYnWNWfOnLGu8dII0UvjSclbk1AvzUi9NLn0MjdJ8vl81jVev
oW8NID1UuNlvb0ey8vaeeHlOLbNNL8OL2ve0NBgXZOcnGxdI0kHDx60rrnvvvs8HSsYDCo+Pv6qz3MFBABw
ggACADhBAAEAnCCAAABOEEAAACcIIACAEwQQAMAJAggA4AQBBABwggACADhBAAEAnCCAAABOtNpmpD6fz6r
poJdmfi1p/Pjx1jV5eXnWNV6annpt/hoRYf//L16ahHppRuq1waoXp06dsq7x8m33+eefW9d4/
b44d+6cdY3XBrC2vKxdXV2dp2NVVVVZ13j5vti8ebN1zccff2xdI0k7d+70VOcFzUgBAK0SAQQAcIIAAgA4
QQABAJwggAAAThBAAAAnCCAAgBMEEADACQIIAOAEAQQAcIIAAgA4QQABAJxotc1I0XIGDhzoqS4hIcG6pqy
szLqmZ8+e1jVHjx61rpG8Na08cuSIp2MB7R3NSAEArRIBBABwwiqA8vLyNGLECMXFxSkxMVHTp0/
XoUOHwsaMGzcu9Fk+X26PPfZYk04aAND2WQVQQUGBcnNztXv3bm3evFl1dXWaNGmSKisrw8Y98sgjOnnyZG
hbunRpk04aAND2WX3U5KZNm8Ier169WomJidq3b5/
GjBkT2t+pUyclJyc3zQwBAO3S13oNKBgMSpK6desWtj8/P18JCQkaPHiwFi9efM2Pta2pqVF5eXnYBgBo/
6yugL6qoaFBCxYs0KhRozR48ODQ/gcffFC9e/dWamqqDh48qB/
+8Ic6dOiQ1q1bd8Wvk5eXp6efftrrNAAAbZTn9wHNmzdPb7/9tt59991rvk9j27ZtmjhxogoLC9W3b9/
Lnq+pqVFNTU3ocXl5udLS0rxMCR7xPqD/x/uAgKbT2PuAPF0BzZ8/X2++
+aZ27NjR6A+HzMxMSbpqAPn9fvn9fi/
TAAC0YVYBZIzR9773Pa1fv17bt29XRkZGozUHDhyQJKWkpHiaIACgfbIKoNzcXL3yyivauHGj4uLiVFJSIk
kKBAKKiYnRkSNH9Morr+juu+9W9+7ddfDgQS1cuFBjxozR0KFDm+UvAABom6wC6IUXXpB08c2mX7Vq1SrNm
TNH0dHR2rJli5YtW6bKykqlpaVp5syZ+tGPftRkEwYAtA/
Wv4K7lrS0NBUUFHytCQEAbgx0wwYANAu6YQMAWiUCCADgBAEEAHCCAAIAOEEAAQCcIIAAAE4QQAAAJwggAI
ATBBAAwAkCCADgBAEEAHCCAAIAOEEAAQCcIIAAAE4QQAAAJwggAIATBBAAwAkCCADgBAEEAHCCAAIAOEEAA
QCcIIAAAE4QQAAAJwggAIATrS6AjDGupwAAaAKN/
TxvdQFUUVHhegoAgCbQ2M9zn2lllxwNDQ0qLi5WXFycfD5f2HPl5eVKS0vT8ePHFR8f72iG7rEOF7EOF7EO
F7EOF7WGdTDGqKKiQqmpqYqIuPp1TocWnNN1iYiIUM+ePa85Jj4+/
oY+wb7EOlzEOlzEOlzEOlzkeh0CgUCjY1rdr+AAADcGAggA4ESbCiC/368lS5bI7/
e7nopTrMNFrMNFrMNFrMNFbWkdWt1NCACAG0ObugICALQfBBAAwAkCCADgBAEEAHCCAAIAONFmAmjFihVKT
09Xx44dlZmZqT179rieUov7yU9+Ip/PF7YNHDjQ9bSa3Y4dOzR16lSlpqbK5/Npw4YNYc8bY/
TUU08pJSVFMTExys7O1uHDh91Mthk1tg5z5sy57PyYMmWKm8k2k7y8PI0YMUJxcXFKTEzU9OnTdejQobAx1
dXVys3NVffu3RUbG6uZM2eqtLTU0Yybx/
Wsw7hx4y47Hx577DFHM76yNhFAr732mhYtWqQlS5bogw8+0LBhwzR58mSdOnXK9dRa3M0336yTJ0+Gtnfff
df1lJpdZWWlhg0bphUrVlzx+aVLl2r58uVauXKl3n//
fXXu3FmTJ09WdXV1C8+0eTW2DpI0ZcqUsPPj1VdfbcEZNr+CggLl5uZq9+7d2rx5s+rq6jRp0iRVVlaGxix
cuFBvvPGGXn/
9dRUUFKi4uFgzZsxwOOumdz3rIEmPPPJI2PmwdOlSRzO+CtMGjBw50uTm5oYe19fXm9TUVJOXl+dwVi1vyZ
IlZtiwYa6n4ZQks379+tDjhoYGk5ycbJ599tnQvrKyMuP3+82rr77qYIYt49J1MMaY2bNnm2nTpjmZjyunT
p0ykkxBQYEx5uK/fVRUlHn99ddDYz7++GMjyezatcvVNJvdpetgjDFjx441P/
jBD9xN6jq0+iug2tpa7du3T9nZ2aF9ERERys7O1q5duxzOzI3Dhw8rNTVVffr00UMPPaRjx465npJTRUVFK
ikpCTs/AoGAMjMzb8jzY/
v27UpMTNRNN92kefPm6ezZs66n1KyCwaAkqVu3bpKkffv2qa6uLux8GDhwoHr16tWuz4dL1+FL+fn5SkhI0
ODBg7V48WJVVVW5mN5Vtbpu2Jc6c+aM6uvrlZSUFLY/KSlJn3zyiaNZuZGZmanVq1frpptu0smTJ/
X0009r9OjR+uijjxQXF+d6ek6UlJRI0hXPjy+fu1FMmTJFM2bMUEZGho4cOaJ//Md/
VE5Ojnbt2qXIyEjX02tyDQ0NWrBggUaNGqXBgwdLung+REdHq0uXLmFj2/
P5cKV1kKQHH3xQvXv3Vmpqqg4ePKgf/vCHOnTokNatW+dwtuFafQDh/
+Xk5IT+PHToUGVmZqp3795au3at5s6d63BmaA3uv//+0J+HDBmioUOHqm/
fvtq+fbsmTpzocGbNIzc3Vx999NEN8TrotVxtHR599NHQn4cMGaKUlBRNnDhRR44cUd+
+fVt6mlfU6n8Fl5CQoMjIyMvuYiktLVVycrKjWbUOXbp00YABA1RYWOh6Ks58eQ5wflyuT58+SkhIaJfnx/
z58/Xmm2/
qnXfeCfv8sOTkZNXW1qqsrCxsfHs9H662DleSmZkpSa3qfGj1ARQdHa3hw4dr69atoX0NDQ3aunWrsrKyHM
7MvXPnzunIkSNKSUlxPRVnMjIylJycHHZ+lJeX6/3337/hz48TJ07o7Nmz7er8MMZo/
vz5Wr9+vbZt26aMjIyw54cPH66oqKiw8+HQoUM6duxYuzofGluHKzlw4IAkta7zwfVdENdjzZo1xu/
3m9WrV5v/+Z//MY8++qjp0qWLKSkpcT21FvX444+b7du3m6KiIvPee+
+Z7Oxsk5CQYE6dOuV6as2qoqLC7N+/3+zfv99IMs8995zZv3+/
+eyzz4wxxvz85z83Xbp0MRs3bjQHDx4006ZNMxkZGeb8+fOOZ960rrUOFRUV5oknnjC7du0yRUVFZsuWLea
v//qvTf/+/
U11dbXrqTeZefPmmUAgYLZv325OnjwZ2qqqqkJjHnvsMdOrVy+zbds2s3fvXpOVlWWysrIczrrpNbYOhYWF
5p//+Z/N3r17TVFRkdm4caPp06ePGTNmjOOZh2sTAWSMMb/
61a9Mr169THR0tBk5cqTZvXu36ym1uFmzZpmUlBQTHR1tvvGNb5hZs2aZwsJC19Nqdu+8846RdNk2e/
ZsY8zFW7F//OMfm6SkJOP3+83EiRPNoUOH3E66GVxrHaqqqsykSZNMjx49TFRUlOndu7d55JFH2t3/
pF3p7y/JrFq1KjTm/Pnz5u///u9N165dTadOncy3vvUtc/
LkSXeTbgaNrcOxY8fMmDFjTLdu3Yzf7zf9+vUz//AP/
2CCwaDbiV+CzwMCADjR6l8DAgC0TwQQAMAJAggA4AQBBABwggACADhBAAEAnCCAAABOEEAAACcIIACAEwQQ
AMAJAggA4MT/AcBjvi3QnOhnAAAAAElFTkSuQmCC",
"text/plain": [
"<Figure size 640x480 with 1 Axes>"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"plt.imshow(image.squeeze(), cmap=\"gray\")\n",
"plt.title(class_names[label]);"
]
},
{
"cell_type": "markdown",
"id": "9a09388a-d754-485f-aa26-4e7a0f782967",
 "metadata": {
"id": "9a09388a-d754-485f-aa26-4e7a0f782967"
},
"source": [
"Beautiful, well as beautiful as a pixelated grayscale ankle boot can get.\n",
"\n",
"Let's view a few more."
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "7188ed7a-5959-48c4-ac7f-19129a2adc83",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 752
},
"id": "7188ed7a-5959-48c4-ac7f-19129a2adc83",
"outputId": "98d50938-b984-4725-8949-d85bf3143555"
},
"outputs": [
{
"data": {
"image/png":
"iVBORw0KGgoAAAANSUhEUgAAAswAAALfCAYAAAB1k5QvAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIH
ZlcnNpb24zLjcuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/
bCgiHAAAACXBIWXMAAA9hAAAPYQGoP6dpAACmVUlEQVR4nOzdd3xVVb7//08MJISEhBYICZBA6EVQQLAgRR
AVRB1QYdQBbIyKZcYZv5Y7V51Rx4qoWOfnKCIOlgErqKioI+hgAwWl9xpK6E1h//
7wQa5hvddmHxJIez0fj3ncy4e1zt5nn7XXWR72Z33igiAIDAAAAIB0TEmfAAAAAFCasWAGAAAAQrBgBgAAA
EKwYAYAAABCsGAGAAAAQrBgBgAAAEKwYAYAAABCsGAGAAAAQrBgBgAAAEJU+AXz0KFDLSUl5ZDtunfvbt27
dy+243bv3t3atGlTbK8HFFVcXJyNGDHikO2ef/
55i4uLs6VLlx75kwKAco51SNlQJhfMTzzxhMXFxVnnzp1L+lTKpHvuucdef/31kj4NHEXff/
+9DRw40LKzs61KlSqWlZVlvXv3tscee+yIH5vxhqPhwH/I/fp/
derUsR49etjkyZNL+vRQzrAOKZqy+L1QJhfM48aNs5ycHJsxY4YtXLiwpE+nzCmLAxWHb/
r06daxY0ebNWuWXXHFFTZ69Gi7/PLL7ZhjjrFHHnkk5te75JJLbNeuXZadnR2pPeMNR9Nf//
pXGzt2rL3wwgt200032fr16+2ss86yt99+u6RPDeUI65CiKYvfC5VK+gRitWTJEps+fbpNmDDBhg8fbuPGj
bPbb7+9pE8LKLXuvvtuS0tLsy+//NKqV69e6O/
y8vJifr34+HiLj48PbRMEge3evduSkpJifn2gKM4880zr2LFjwZ8vu+wyq1u3rv3rX/
+yfv36leCZobxgHVIxlblfmMeNG2c1atSwvn372sCBA23cuHFOm6VLl1pcXJw9+OCD9swzz1hubq4lJiZap
06d7MsvvzzkMWbOnGnp6enWvXt32759u7fdnj177Pbbb7cmTZpYYmKiNWjQwG666Sbbs2dP5Pfz9ddf20kn
nWRJSUnWqFEje+qpp5w2eXl5BZN+lSpVrF27djZmzBin3Y4dO+zGG2+0Bg0aWGJiojVv3twefPBBC4KgoE1
cXJzt2LHDxowZU/DPlkOHDo18vih7Fi1aZK1bt3YWy2ZmderUcWKvv/
66tWnTxhITE61169b27rvvFvp79QxzTk6O9evXz9577z3r2LGjJSUl2dNPP814Q4mrXr26JSUlWaVK//
f70IMPPmgnnXSS1apVy5KSkqxDhw722muvOX137dpl1113ndWuXduqVatm/fv3t1WrVllcXJzdcccdR/
FdoDRhHVJB1yFBGdOiRYvgsssuC4IgCD799NPAzIIZM2YUarNkyZLAzILjjjsuaNKkSXDfffcF999/
f1C7du2gfv36wd69ewvaDhkyJEhOTi7484wZM4IaNWoEvXv3Dnbu3FkQ79atW9CtW7eCP+/bty84/
fTTg6pVqwY33HBD8PTTTwcjRowIKlWqFJxzzjmHfB/
dunULMjMzgzp16gQjRowIHn300eCUU04JzCx49tlnC9rt3LkzaNmyZVC5cuXgD3/4Q/
Doo48GXbt2DcwsGDVqVEG7/fv3Bz179gzi4uKCyy+/
PBg9enRw9tlnB2YW3HDDDQXtxo4dGyQmJgZdu3YNxo4dG4wdOzaYPn36oS88yqzTTz89qFatWvD999+HtjO
zoF27dkG9evWCv/
3tb8GoUaOCxo0bB1WrVg02bNhQ0O65554LzCxYsmRJQSw7Ozto0qRJUKNGjeDmm28OnnrqqWDq1KmMNxw1B
8blBx98EKxfvz7Iy8sLZs+eHQwfPjw45phjgvfff7+gbf369YOrr746GD16dDBy5MjghBNOCMwsePvttwu9
5gUXXBCYWXDJJZcEjz/+eHDBBRcE7dq1C8wsuP3224/
yO0RpwTqkYq5DytSC+auvvgrMLJgyZUoQBL98OPXr1w+uv/76Qu0ODNRatWoFmzZtKoi/
8cYbgZkFb731VkHs1wP1s88+C1JTU4O+ffsGu3fvLvSaBw/UsWPHBsccc0zwn//
8p1C7p556KjCzYNq0aaHvpVu3boGZBQ899FBBbM+ePUH79u2DOnXqFNxMo0aNCswsePHFFwva7d27NzjxxB
ODlJSUYOvWrUEQBMHrr78emFlw1113FTrOwIEDg7i4uGDhwoUFseTk5GDIkCGh54fy4/333w/
i4+OD+Pj44MQTTwxuuumm4L333is0YQfBLwvmhISEQmNl1qxZgZkFjz32WEHMt2A2s+Ddd991js94w9FwYF
we/L/ExMTg+eefL9T214uQIPhlTm3Tpk3Qs2fPgtjXX3/
tfNEHQRAMHTqUBXMFxjrkFxVxHVKmHskYN26c1a1b13r06GFmv/
ysf+GFF9r48eNt3759TvsLL7zQatSoUfDnrl27mpnZ4sWLnbZTp061Pn362GmnnWYTJkywxMTE0HN59dVXr
WXLltaiRQvbsGFDwf969uxZ8HqHUqlSJRs+fHjBnxMSEmz48OGWl5dnX3/9tZmZTZo0yTIyMmzw4MEF7SpX
rmzXXXedbd++3T755JOCdvHx8XbdddcVOsaNN95oQRCQJV6B9e7d2z7//HPr37+/
zZo1y+6//37r06ePZWVl2Ztvvlmoba9evSw3N7fgz8cee6ylpqbKe+ZgjRo1sj59+hT7+QOxePzxx23KlCk
2ZcoUe/HFF61Hjx52+eWX24QJEwra/
PrZ+vz8fNuyZYt17drVvvnmm4L4gUeRrr766kKvf+211x7hd4DSjHXILyriOqTMLJj37dtn48ePtx49etiS
JUts4cKFtnDhQuvcubOtW7fOPvzwQ6dPw4YNC/
35wKDNz88vFN+9e7f17dvXjjvuOHvllVcsISHhkOezYMECmzNnjqWnpxf6X7NmzcwsWjJVZmamJScnF4od6
H/g+dBly5ZZ06ZN7ZhjCn9ULVu2LPj7A/
83MzPTqlWrFtoOFVOnTp1swoQJlp+fbzNmzLBbbrnFtm3bZgMHDrQffvihoN3B94zZL/
fNwfeM0qhRo2I9Z+BwnHDCCdarVy/
r1auXXXTRRfbOO+9Yq1atbMSIEbZ3714zM3v77betS5cuVqVKFatZs6alp6fbk08+aVu2bCl4nWXLltkxxx
zjjOsmTZoc1feD0oN1SMVeh5SZXTI++ugjW7NmjY0fP97Gjx/v/
P24cePs9NNPLxTzZfIHv3r43MwsMTHRzjrrLHvjjTfs3XffjZRJvX//fmvbtq2NHDlS/
n2DBg0O+RrA0ZaQkGCdOnWyTp06WbNmzWzYsGH26quvFmR4R71nFHbEQGl0zDHHWI8ePeyRRx6xBQsW2KZN
m6x///526qmn2hNPPGH16tWzypUr23PPPWcvvfRSSZ8uSjHWIRVbmVkwjxs3zurUqWOPP/
6483cTJkywiRMn2lNPPXVYX9pxcXE2btw4O+ecc+z888+3yZMnH7KaTm5urs2aNctOO+00i4uLi/
mYZmarV6+2HTt2FPqvu/nz55vZL7sOmJllZ2fbd999Z/v37y/0X3dz584t+PsD//
eDDz6wbdu2Ffqvu4PbHXi/wIGtt9asWXNEj8N4Q0n7+eefzcxs+/bt9u9//9uqVKli7733XqF/
8n7uuecK9cnOzrb9+/
fbkiVLrGnTpgVx9tytuFiHVOx1SJl4JGPXrl02YcIE69evnw0cOND534gRI2zbtm3O85ixSEhIsAkTJlinT
p3s7LPPthkzZoS2v+CCC2zVqlX2j3/8Q57vjh07DnnMn3/+2Z5++umCP+/
du9eefvppS09Ptw4dOpiZ2VlnnWVr1661l19+uVC/
xx57zFJSUqxbt24F7fbt22ejR48udIyHH37Y4uLi7MwzzyyIJScn2+bNmw95figfpk6dKn8hnjRpkpmZNW/
e/Igen/GGkvTTTz/Z+++/bwkJCdayZUuLj4+3uLi4Qs+bLl261CmicOB5/CeeeKJQ/
GhUx0TpwzqEdUiZ+IX5zTfftG3btln//
v3l33fp0sXS09Nt3LhxduGFFx72cZKSkuztt9+2nj172plnnmmffPKJt876JZdcYq+88or9/ve/
t6lTp9rJJ59s+/
bts7lz59orr7xSsB9tmMzMTLvvvvts6dKl1qxZM3v55Zdt5syZ9swzz1jlypXNzOzKK6+0p59+2oYOHWpff
/
215eTk2GuvvWbTpk2zUaNGFfxX3Nlnn209evSw2267zZYuXWrt2rWz999/39544w274YYbCiVydejQwT744
AMbOXKkZWZmWqNGjSjvWY5de+21tnPnTjvvvPOsRYsWtnfvXps+fbq9/
PLLlpOTY8OGDTuix2e84WiaPHlywS9aeXl59tJLL9mCBQvs5ptvttTUVOvbt6+NHDnSzjjjDPvtb39reXl5
9vjjj1uTJk3su+++K3idDh062IABA2zUqFG2ceNG69Kli33yyScFv76VxV/
IcPhYh7AOKRPbyp199tlBlSpVgh07dnjbDB06NKhcuXKwYcOGgu1cHnjgAaedHbQd0MH7HwZBEGzYsCFo1a
pVkJGRESxYsCAIAnc7lyD4ZVuV+
+67L2jdunWQmJgY1KhRI+jQoUNw5513Blu2bAl9T926dQtat24dfPXVV8GJJ54YVKlSJcjOzg5Gjx7ttF23
bl0wbNiwoHbt2kFCQkLQtm3b4LnnnnPabdu2LfjDH/4QZGZmBpUrVw6aNm0aPPDAA8H+/
fsLtZs7d25w6qmnBklJSYGZlbmtXRCbyZMnB5deemnQokWLICUlJUhISAiaNGkSXHvttcG6desK2plZcM01
1zj9s7OzC40R37Zyffv2lcdnvOFoUNvKValSJWjfvn3w5JNPFpoHn3322aBp06ZBYmJi0KJFi+C5554Lbr/
99uDgr8QdO3YE11xzTVCzZs0gJSUlOPfcc4N58+YFZhbce++9R/
stogSxDmEdEhcEEbJ5AACAzZw504477jh78cUX7aKLLirp0wFwlJSJZ5gBADjadu3a5cRGjRplxxxzjJ166
qklcEYASkqZeIYZAICj7f7777evv/
7aevToYZUqVbLJkyfb5MmT7corr2TLLqCC4ZEMAACEKVOm2J133mk//PCDbd+
+3Ro2bGiXXHKJ3XbbbVapEr83ARUJC2YAAAAgBM8wAwAAACFYMAMAAAAhWDADAAAAISJnLVDVCEdKST5GXx
7GtXoP6pomJyfL/
oMGDXJi27dvd2L5+fmyf0ZGhhPbtm2bbDtx4kQZL48Y1yiPGNcoj6KMa35hBgAAAEKwYAYAAABCsGAGAAAA
QrBgBgAAAEJQqggohaIm8oXFD9a3b18Zr1GjhhOrXLmyE1PJfWZmbdu2dWItW7aUbY9m0l8s1xAAgDD8wgw
AAACEYMEMAAAAhGDBDAAAAIRgwQwAAACEYMEMAAAAhIgLIqaNU5JSa968uRNLT0+XbXft2uXE1G4EZmZ79+
6N1Hbfvn2y//79+yPFfMc65hj3v6VUzEyPjWrVqsm23377rRNTZZiPlvIwrlNTU53YgAEDnFjHjh1l/
+nTpzux//f//p8TU7thmJmtXr3aif3tb3+TbVV57mXLljmxKVOmyP5btmyR8dKIEsIojxjX5Y/
vupbGXYUyMzNlXO3i5FvzzJw504lRGhsAAAAoIhbMAAAAQAgWzAAAAEAIFswAAABACJL+BF9ym3qA/
O6773Zi9erVk/337NnjxHwlhFWCX9WqVZ2YStgz0+WOfXbv3u3EKlVyq6avXLlS9ldDyPew/
aOPPurEJk2adKhTPGJK67g+9thjndhxxx0n26rP+ueff3ZiDRs2lP3VWFPXJTc3V/Z///
33ndiCBQtk2yZNmkQ6vi9pdP369U7MlyC4cOFCGT9aSjJhRs1hsZxPLPdFaUwMwpFD0l/
RqPdQ1HtTxXzfwer7olGjRrLtvHnznNiOHTsOdYqhsrOzZbxu3bpOTCWJ+6SlpTkxlRBvZvbaa685sSjvi1
+YAQAAgBAsmAEAAIAQLJgBAACAECyYAQAAgBAsmAEAAIAQ7lYIiCljVe18oXadMNOlsefPny/
bVqlSxYnFx8c7sU2bNsn+tWrVcmK+3T8SEhIiHct3XX766ScnlpiYKNsuWrRIxiuqCy64QMZbtWrlxJYsWS
Lb5ufnOzE1LtTnZKazi1WG9Ycffij7qx05ateuLduqMujqvFS5bTOz6tWrO7EhQ4bItv/
+97+dmCqJWtGpe33fvn2R+6uxqsaEj2+nH3UOakcV37ym3pfa/
ac4dl5Qc6OK+c41luul7s2kpKTIr6nazp07V7ZV9ytKVlF3KWnevLkTS09Pl2137tzpxHxjRRkwYEDktmoX
MDWG1XeAmR6ram1j5l+jHQq/
MAMAAAAhWDADAAAAIVgwAwAAACFYMAMAAAAhSPorIpWw4kuYUQ+g+x5KV4koqm1WVpbsrxJDVMKLmU6OUck
pvlKbqq1KzjE7/
Ifty4OMjAwn5iuNPmfOHCfmS05Sn59K7PFde5U0qJJGt27dKvurhBFfwpG6N9R5+Upjq7aLFy+Wbdu3b+/
EKkrSXyyJQbEk+J1++ulOrG/
fvk5s6dKlsv+GDRucWNu2bWVblcSjEqd9c6hKyFb3kC8RL5ZkwKiv67vWURP5zHRpY3W91bUy0+Xpp02bJt
u+/
vrrMo7DdyRKi6vX9I1fNTf7kvRTUlKcWJ06dZzYZZddJvurublmzZqyrVqfqHtbJSKa6fvFd78d7mfAL8wA
AABACBbMAAAAQAgWzAAAAEAIFswAAABACJL+ikglIakkLDOdyOR7AF4lzakH1X1VsmJJOFHnq17Xdyx1rpm
ZmbKt74H9iqBFixZObPPmzbJtLNXztmzZ4sRiGSvq81Pn5auwpMaVL8FUJTepqpa+RKxt27Y5MV+CoDpflQ
hzJJJwyrp77rlHxlUinUrOU1UWzfTn75sTTjzxRCemqkLGMt+qseq7L9S5+hKflajzqplOblLX2sxs3rx5T
kxdb1/i7dVXX+3EunXrJtt+9NFHMo7SzzcHq6TRHTt2yLbqHlJJfytXrpT9VRVZ37HUvaE2NfD1V/
e2b244XPzCDAAAAIRgwQwAAACEYMEMAAAAhGDBDAAAAIRgwQwAAACEYJeMIkpOTnZivjKTKotTZf2b6UxWl
eHvK3+qsq59GaMqm1q19WVtq10WfJmsqjRuReErF67EkmFftWpVJ6bGha80thqvajcCXwlitXOAb6wo6rr4
xrW6Lr5dFlSGttppZP369Yc6xTJHXVP1mZqZffDBB07s8ccfl23VjiYDBgxwYldddZXsX69ePSe2evVq2Va
NoTPPPNOJqTLyZnpcq91XfDtfqPm6qKV2ValhM72jga+8eHZ2thPr2rVr5HNS78F3bzdt2lTGUXKi7vSjdp
gwM2vUqJET882B6jsjPT3diW3atEn2z8jIcGK+tYGamzdu3OjEfGNV3du+7zzfDiKHwi/
MAAAAQAgWzAAAAEAIFswAAABACBbMAAAAQAiS/opIlY70JZGoB/
NVWVkzncikHor3JRimpqY6MV+pVBVXD8v7ks5UMpHvYfuKTCX8rFmzRrZViRnLli2TbVUinEqA8CU6RE3Q8
yWG1KpVK9LxfXGVSOZLhlVtVeKtmR6vKmGqPCb9qXuye/
fusm3dunWd2MSJE2Xbl19+2YmpcZWbmyv7q88qJydHth05cqQTU/
PaySefLPurEtJqbvYl8qm5VSVcmelxrd5rfn6+7P/
jjz86MV+SZqtWrZyYSoSKZb5XicNh54DSz5dgqpKhfXOgugd8CdmKWh/
5khGjzsO+ZNZYEtXVpgpR8AszAAAAEIIFMwAAABCCBTMAAAAQggUzAAAAEIIFMwAAABCCXTKEqGVOzXSGta
8Etcrw92VoK+p1fVnMKq4yVs105vnmzZudmG83ArUjhu9YFZnaTWLRokWybbt27ZyYL4tY7SihSiP7MvwVN
dZUdrWvrW/nDZWNr3ZJ+M9//
iP7H3vssZGPpcawKs1cUQwbNkzGr7322siv0bBhQye2du3ayP3VWK1evbpse9lllzmxv/
3tb07MtyNPixYtnJiag33UHOq7h5YvX+7E1q1b58S2bt0q+6vdM3y7h6gxvGDBAifm20FJfb81a9ZMtvWVn
UfJiVoa27dDhLpffHO7mkPVmqVx48ayvyqN7btfa9as6cTU+POV1lbU96DZ4e/+wi/
MAAAAQAgWzAAAAEAIFswAAABACBbMAAAAQIgKn/QX9QF6H5WA4Uv627JlixPzlZtWiXifffaZE/
OVO1bH8iWRqHhSUpIT27Bhg+yfmZnpxL799lvZtiJTCW+
+JCBVGluVzzXzl2KPSvVXYziW0tq+tirhY+bMmU7MV9Y1Ly/PifnuN5X04ktcLW9U6df+/
fvLtkOGDIn8umpeUPOl7/NXn9XSpUtl2w4dOjixwYMHO7FZs2bJ/h999JETO+GEE5zY4sWLZX81t2/
cuFG2Pfvss53YtGnTnJhvDlaJVOq9mpl9+OGHTkyNa1+CorrfVXKXmX5fKBt8yXUrV650YtnZ2bKtSuhV8+
r27dtlf5X8r+YQM7O5c+c6MZW46tsoQcVjaRsFvzADAAAAIVgwAwAAACFYMAMAAAAhWDADAAAAIUj6K2LSn
0ra8lXYUQl+vmOppKdWrVpFPq9Vq1Y5sZ9//lm2VdW3VIKaSjo0M+vXr58T+
+abbw51iuWaqpx0zDHuf5/6kqNUYoQvUUG9bixV/VR/
dSxfEolvvCuq8lIsyRrqHkpPT5dtVSKTL+GkvPnd737nxCZPnhy5v2/8+KrHRaXmOzX+zHQVzNNOO82J+So
NqsqaqlLh119/Lfu/8MILTsyXIKmSUdVYXbZsmex/
+eWXOzGVDGsWvVqhb25R1dN891ubNm0iHQtHT9T1yYoVK2RcJfj5kj5V8rlKzvMlXk+fPt2J+ZLU1fpEzde
+xG1Vvc/
33XS43wP8wgwAAACEYMEMAAAAhGDBDAAAAIRgwQwAAACEYMEMAAAAhCiVu2QUdeeKI0Wd165du5yYKgdpps
vV+jI+VcaoKpftywKtW7du5GMpKmu2c+fOkfsvWLAgctvySJXVVWNFZfaa6c/aR2Udq/
Hj2yUl6i4bvjLederUiXROvmOpbH7fuarX9e3coMq1Vq9eXbYtb1R2+5gxYyL39823KkNe7bDgy5pXfLs+R
C3hfMUVV8j+U6ZMcWLz5893Yr75+tZbb3VivjlUjbU+ffo4MVWa28zs888/
d2Kq5LxZ9F1pfPeg2qlG7Zxhpnf6QNFEXd/
4dqqJ2t93D6oduHzHUt9Par5VaxszPdZ88vPznVjNmjWdWFZWluy/
cOFCJ7Zjxw7Z1rez0qHwCzMAAAAQggUzAAAAEIIFMwAAABCCBTMAAAAQolQm/
R3NBL9YSgir5BT1AL0qK20WvSywmU6OUsfylStWSRy+B+CVBg0aRDq+jy+JpKJQiRHqs/
Ylt6lkC5XIaaZLCKvEIF+pXJVcpO6LzMxM2V8lPPnuAZWIpMa6L2FJJe1lZGTItt99950TU5+LLzHFl5BZF
qhEzHnz5kXu70sYimW+VGJJblJjQCWhrVy5UvY/5ZRTnJgqietLZv3xxx+dmK9UtLreGzZscGIff/yx7K/
uTd9nEPV+iSVpzHes0pBsX95Evaa+dlHvwcaNG8u4mu/
T0tJkW5Worr6zfOcUS0KwSsRT78FXXn7jxo2RXtPM/
717KPzCDAAAAIRgwQwAAACEYMEMAAAAhGDBDAAAAIQolUl/
R1MsSQ1NmzZ1YuoBdl9ynXow3le1SSV2FLXqUixJBMuXL3divuQodQ1UpbuKRFUEUwkQvmu6Zs0aJ6YSlsy
ij2FfIp36/NT4i6XymO9Yijp/X4KiSsTzJb6
qaxtLwotK2ior1PVTCTw+vgRPNVZimZdiSfpTbdXnr8aqmVleXp4TizrWzfR870sQVIlI6r7w3e+xJFOqhC
X1HmKZ7333K5X+il9RKxmrz0qNq/r168v+27Ztc2K+5DhVaU9V8fUlSKtz9a0NVPL2Dz/
84MR8Sb4q8dZX2XXJkiUyfij8wgwAAACEYMEMAAAAhGDBDAAAAIRgwQwAAACEYMEMAAAAhKhQu2TEUoJX6d
27txNTGafVqlWT/
VWGdizZySoT1VeCWJWxVmWBfa+hsux9uxGoXUHat28v27711lsyXt5E3TnC95moTHhfGXU1rmIpf6vOS8V8
n7/
aEcRXelTdg1FjZnqXC1+GuTovdV1UyfuyTn1WsezE4NuRRVHXvzh2aFBzkBpXvnGtxLIbgdpVJJadgtRY9e
2KpOZ73+eljqXmhlh2H/FdF3bJiCaWnS+i3oex9D/hhBOcmG9XI7U28H23zJ49O9Lr+nYUat26tRPzfed9+
+23Mn6wJk2ayLia8zZv3izbHu645hdmAAAAIAQLZgAAACAEC2YAAAAgBAtmAAAAIMQRSfqLJbkkliSMqHxl
dX0Pth/
svPPOk3FVZlG9pi+xQyWM+M4patKeL2FJxX1JACqRRj1Av3PnTtlfJWj5SlJWFOoeiCWRT7X1JZOqtipB1J
ewFDUZ1pcIpvrHktykkph8iVzqevmuoboH1LXyJaGUZSoZ2Ve+VollrMRS8r2oyWlRy537XleNCV/
J8Kiluc30/
RJ1DjDT18WXDBlLkqyiziGWxEu4ilraWvF9ps2aNXNiar7duHGj7J+Tk+PEfMlxah5u1KiRE/
OV4Vb9Fy1aFLmtel/
5+fmyv5obfHN7LInCv8YvzAAAAEAIFswAAABACBbMAAAAQAgWzAAAAEAIFswAAABAiMgpsLGUflRxX8Zn1O
xe37FUxqkvE1np0aOHEzv22GNl25UrVzoxVQK6Ro0asr/
aZcK384HK5lYZ2r7sVpVJ6ruGqjS2yjj1nasaGzVr1pRtK4qon58vW1fdF75xrY4Vtdy1L66y433nqsa1Lz
tZjRV1XXzZ+aq/b5eFzMxMJ7Z9+3YndrgZ06WZmquuvPJK2fbvf/+7E/ONtagZ/
r6dS9S49n1+UXeOiGWHCPX5xzIH+6xbt86JxbJzgmrruy7qGqhr5buH1P0Wy3dmSYplHRK1v8+R2MHLzCw1
NdWJqblK7WZhps9rx44dTqx58+ayv/
puV7vqmOldjdRY882h6nXVPWimPxt1LLWDmJnZpk2bnNiaNWtk26jz2MH4hRkAAAAIwYIZAAAACMGCGQAAA
AjBghkAAAAIcUSS/
pTDfcj6cPhK+Pbr18+JNWnSxImph8fNzGrXru3E1MP2sVyXrVu3yrgqa6muYSyJl77EDvVgf4MGDZxYLMk5
KomhIomaiOdLzIllDKnyoaq0ue81VdKUStpbu3at7K+SSHylsVU8lnLJaqz67iE1BlWC4tGcm46W7777zok
99dRTsq1K+vNd0+zsbCc2e/
ZsJ+ZLAopaGj0sfjDfuFbJbXXq1HFivtLYeXl5Tkydv5lZRkaGE1MJ2b7kqliS0aLyJV6qUuK+a+grO16ax
FJavaiJfL4E0+TkZCfWuHFj2bZ69epOTI0LX8Ka6q++b9TaxCy2z1/
NjXXr1nVivu8GtWZS94pZ9CTbxYsXy7h6v6+88opsm5ubG+lYB+MXZgAAACAEC2YAAAAgBAtmAAAAIAQLZg
AAACBE5KQ/
9fB3rVq1ZNuWLVu6B4qh6pB6iF8lFpnpB9DVg+a+11APq6vEEDOdNDd37lwnphL2zMzq1asX6TXNdOKiivm
SEGKp5qReQyU+
+j4Dlciljl+RqGut7gFVZdHMbMWKFU5MJayZRa8qGEvCi+rvq4imkot8yUK+CoBRqbHqSxaJmgjj+wzKsnf
eeceJff/997LtSSed5MSmT58u26oxrOYF31hTc7vv8ytq9TnVX31nXXTRRbL/zJkzIx/rf/
7nf5xYnz59nJhKJDTT19CXtOebh6NSyZQqSdjMn/xZUopa1U8lzJnpyrTp6elOzDd/
RU2cNtPzeP369Z2YLxl1/
fr1TiwtLc2JbdmyRfZXiavqNc3MOnXq5MTU94BvzaW+B3xrAzUG1ToklqrRvnvlcOcWfmEGAAAAQrBgBgAA
AEKwYAYAAABCsGAGAAAAQrBgBgAAAEJE3iVDadq0qYyr8qm+TPSi7gaxYcMGJ+bLLt2+fbsTU5m0qp1Z9Ax
xVRLVTGeH+jJp1e4fsZyryoRVu3SYmaWmpjqx1atXOzHfZ6gyYYuayV3WqWxudZ1SUlJkf7WjgW9XGpUhrc
agb6caNa7Urji+zzRqCWMzfQ+oezuW0ti+MtzqPahdMirKWL3ppptk/
A9/+IMT8+2S8fbbbzuxY4891omp7HYzvSOGL2O9qOWio+5coHakMdP3i+9c1W4S6li+nS/
UdYnlflX3oO9c1dzk+379z3/+I+OlSY0aNWRcza2+uUp9LsuXL3divtLoiu+aqnNQ362+8a/
mO7Xm8e1wotYBXbp0kW3VWiwrK8uJqTWEmdmSJUucmG+3JzVfq2vlK62tvjM///
xz2da3dj0UfmEGAAAAQrBgBgAAAEKwYAYAAABCsGAGAAAAQkRO+lPlort27Srbrlu3zon5HmBXSTjbtm1zY
snJybK/
KnXpS6SLWi7Yl4ilEgPUw+6+Uq8qYcCXRKAejPcl+CnqHGJJjlCJBSoBwHdeKjHAzF9Cs7xR40olNfg+k08
++cSJdejQQbZVyYBREyh8bdW9otr5+BKWVCKMOi/
ffKHaquQcM7MmTZo4MfUeilquuzRS12/27NmyrUoQffPNN2VbNQeq6+e7z1Uimi85Sn1W6vixlJBetmyZEz
v99NNl/x9//NGJ+RLp1PfT4sWLnZhvrKn36ruHFHW/+pJZVdKUSp43M/
vss88in0NJadWqlYxnZmY6MV9ZZpU0p74Dfd/
Xah3ja6vGihrDvu9bNS7U973vvarvC9+5qpLd6rzmzJkj+6vy3Kq0tplOXlffo7731bBhQyf2+OOPy7a+pO
ZD4RdmAAAAIAQLZgAAACAEC2YAAAAgBAtmAAAAIAQLZgAAACBE5DTczp07O7F+/frJtj/
88IMT85VpVNmZapeNVatWyf4qi9O3m4TKrlQ7RKjMTDOdNauym9UuH2Y6O1adv5nOGI0lu1XxlRBWmd9q9w
bfLgnqdX2Z62q3lfJIXROVyezbDWLjxo1OzLfLhdplwJdhraj7Qp2r7zP1ZS0rapcA1V+VETfT49KXoX3ii
Sc6MTWGd+/eLfuXZSqT3vc5/fe//3ViJ5xwgmy7dOlSJ6Y+K98uGerz981halyrmG9eU/
eWmptHjBgh+6tsfrWbgpneJUHFfCWEfTuFKGoeUP19JYg//
fRTJ3bnnXdGPn5p45sX1ZrDNy+qeSGWXanq1q3rxHzfdfn5+U5MvQfffdGxY0cnpkprr169WvZX96DaUcRM
z/nz5893Yr4dXdT19r0vdW+q1/XdK+pc1c4ZvvOKgl+YAQAAgBAsmAEAAIAQLJgBAACAECyYAQAAgBCRk/
7eeustJ+ZLtjj33HOdWLNmzWRblXCmknDWrl0r+6sHxX0P66tEKlU+1FeGW8VVaW6VAGCmS5L6kiHVNRg7d
qwTu+CCC2R/
lczoK+vqK+V9MF8ypEpYUEkIZvp6lUdqrKl7wJcAofr7kh3UGFYJEL77Vb2uek1fyfg1a9Y4MV/
5U0WNH18ij0pY8ZV8VonCKpl2wYIFhzrFMsd3rypqrKmyzj5qvvV9/
moM+dqqcak+f3X+ZnoMqfti2rRpsr+aL30J3b7xejA1/sx0MqEvSVPdb6qM9/fffy/
7+75zFF9ScklR81JWVpZsq75r1q9fL9tmZ2c7MZVg7PucYykBrb4bo34H+6jS5jVr1pRt1Vjzra9UMmTLli
2dmG9MqfsllpLfah7zfWeqser7vHwJ7IfCL8wAAABACBbMAAAAQAgWzAAAAEAIFswAAABAiMhJf8q///
3vyPEWLVrItoMHD3ZijRs3dmJNmzaV/dUD9L4HvVUShXpQ3JfooOKbN292Yr7KY3/5y1+cmC/
hJKqHH35YxtV5+ZIhoybH+Cr9qevqSzBs0KCBjJc3USvK+ZKIFF/
Sn0rmiyWRTlFJKL6kQfX5+5LOot5vvveq4r5KfSqJRB1LzTdmumJpWeG7/5QPP/
zQiX300UeyrUqOUuPCV9FOJU77EoTVeFWfXyzvdcWKFU7MlzRaXsVSVdA355eU5s2bOzFf9cXly5c7Md93u
/
q+VGPFNwcqvnnJt6nAwVSCo5lOsFPvy1ctVX2mvvtV3VvqO9yXyBdLJWJ1v6trGMt3SyxrwSj4hRkAAAAIw
YIZAAAACMGCGQAAAAjBghkAAAAIwYIZAAAACFGkXTJ8WZwqA3Hu3Lmy7e233x7pWL7sVpU1W69ePdm2du3a
Tkxl0vvKRK5evdqJzZs3T7Y9Wn7/+9/
L+Lp165yYKp9pprNmVYatLxtZZcL6SmXm5eU5sfHjx8u2ZZnaFUaVcPWV9VVmzpwp4+3bt3dialz7MpbV56
+yi3391Y4avuxk9Roqa79WrVqyv8pm91HnkJmZWaTXrCh8mehLly49uieCYlfadr6Ihfr+GDBggGyr5kDfz
hFqN4hYdopSc5ivrXpdtfuGb/cYtSOFaqvKyPv4zlXN1zt37nRivl0nYrmGO3bscGLqPRRHuWvf/HYo/
MIMAAAAhGDBDAAAAIRgwQwAAACEYMEMAAAAhIgLIj797Eu6A4rqcB/
ALw5HalyrJI7q1as7MV8ChC9BU+ndu7cTO+2005zYqlWrIh8rIyPDifnOdeXKlU4sJSVFtlXJMdWqVXNiKr
HEzOyFF15wYrGUX1Wf95Eaf+VxXAOlbVyrpGMzneBbo0YN2VaVhlaJbL4S0CruS0JTmyWoY/
lKvqv5btu2bU7MN1+ra+jbwEElTvraFpW6Xir5O5brkp+fL9vOmDHDiUUZ1/
zCDAAAAIRgwQwAAACEYMEMAAAAhGDBDAAAAIRgwQwAAACEYJcMlLjSlnVdHqjs4jZt2si2NWvWdGIqa9y3G
4Xa0cKXSa2ywVV5+blz58r+ZQnjGuUR4xrlEbtkAAAAAEXEghkAAAAIwYIZAAAACMGCGQAAAAgROekPAAAA
qIj4hRkAAAAIwYIZAAAACMGCGQAAAAjBghkAAAAIwYIZAAAACMGCGQAAAAjBghkAAAAIwYIZQLF5/
vnnLS4uzpYuXRpz36FDh1pOTk6xnxMAwMV8HZtyvWCOi4uL9L+PP/64pE8VOGzff/
+9DRw40LKzs61KlSqWlZVlvXv3tscee6ykTw04KhYtWmTDhw+3xo0bW5UqVSw1NdVOPvlke+SRR2zXrl1H5
JgvvfSSjRo16oi8Nsov5uuyq1JJn8CRNHbs2EJ/
fuGFF2zKlClOvGXLlkfztIBiM336dOvRo4c1bNjQrrjiCsvIyLAVK1bYF198YY888ohde+21JX2KwBH1zjv
v2Pnnn2+JiYn2u9/9ztq0aWN79+61zz77zP785z/bnDlz7Jlnnin247700ks2e/
Zsu+GGG4r9tVE+MV+XbeV6wXzxxRcX+vMXX3xhU6ZMceIH27lzp1WtWvVIntoRsWPHDktOTi7p08BRdPfdd
1taWpp9+eWXVr169UJ/l5eXVzInBRwlS5YssUGDBll2drZ99NFHVq9evYK/
u+aaa2zhwoX2zjvvlOAZAv+H+bpsK9ePZETRvXt3a9OmjX399dd26qmnWtWqVe3WW281s18G8GWXXWZ169a
1KlWqWLt27WzMmDGF+n/88cfysY6lS5daXFycPf/88wWxtWvX2rBhw6x+/
fqWmJho9erVs3POOcd5fmjy5MnWtWtXS05OtmrVqlnfvn1tzpw5hdoMHTrUUlJSbNGiRXbWWWdZtWrV7KKL
Liq264KyYdGiRda6dWtn8jUzq1OnTsH//
9xzz1nPnj2tTp06lpiYaK1atbInn3zS6ZOTk2P9+vWzzz77zE444QSrUqWKNW7c2F544QWn7Zw5c6xnz56W
lJRk9evXt7vuusv279/vtHvjjTesb9++lpmZaYmJiZabm2t/+9vfbN++fUV786jw7r//ftu+fbs9+
+yzhRbLBzRp0sSuv/
56MzP7+eef7W9/+5vl5uZaYmKi5eTk2K233mp79uwp1CfKeO3evbu98847tmzZsoJH+yra85yIHfN12Vauf
2GOauPGjXbmmWfaoEGD7OKLL7a6devarl27rHv37rZw4UIbMWKENWrUyF599VUbOnSobd68uWASjsWAAQNs
zpw5du2111pOTo7l5eXZlClTbPny5QWT7dixY23IkCHWp08fu+++
+2znzp325JNP2imnnGLffvttoUn5559/tj59+tgpp5xiDz74YJn8VRxFk52dbZ9//
rnNnj3b2rRp42335JNPWuvWra1///5WqVIle+utt+zqq6+2/
fv32zXXXFOo7cKFC23gwIF22WWX2ZAhQ+yf//
ynDR061Dp06GCtW7c2s1/+469Hjx72888/280332zJycn2zDPPWFJSknPs559/3lJSUuyPf/
yjpaSk2EcffWT/+7//a1u3brUHHnigeC8IKpS33nrLGjdubCeddNIh215+
+eU2ZswYGzhwoN1444323//+1/7+97/bjz/
+aBMnTixoF2W83nbbbbZlyxZbuXKlPfzww2ZmlpKScmTeJMoN5usyLqhArrnmmuDgt9ytW7fAzIKnnnqqUH
zUqFGBmQUvvvhiQWzv3r3BiSeeGKSkpARbt24NgiAIpk6dGphZMHXq1EL9lyxZEphZ8NxzzwVBEAT5+fmBm
QUPPPCA9/y2bdsWVK9ePbjiiisKxdeuXRukpaUVig8ZMiQws+Dmm2+O/P5R/rz//
vtBfHx8EB8fH5x44onBTTfdFLz33nvB3r17C7XbuXOn07dPnz5B48aNC8Wys7MDMws+/
fTTglheXl6QmJgY3HjjjQWxG264ITCz4L///
W+hdmlpaYGZBUuWLAk99vDhw4OqVasGu3fvLogNGTIkyM7OjvzeUbFt2bIlMLPgnHPOOWTbmTNnBmYWXH75
5YXif/
rTnwIzCz766KOCWNTx2rdvX8YrYsJ8XbZV+EcyzMwSExNt2LBhhWKTJk2yjIwMGzx4cEGscuXKdt1119n27
dvtk08+iekYSUlJlpCQYB9//LHl5+fLNlOmTLHNmzfb4MGDbcOGDQX/
i4+Pt86dO9vUqVOdPldddVVM54HypXfv3vb5559b//79bdasWXb//fdbnz59LCsry958882Cdr/
+JWHLli22YcMG69atmy1evNi2bNlS6DVbtWplXbt2Lfhzenq6NW/
e3BYvXlwQmzRpknXp0sVOOOGEQu3UY0G/
Pva2bdtsw4YN1rVrV9u5c6fNnTu3aBcAFdbWrVvNzKxatWqHbDtp0iQzM/vjH/
9YKH7jjTeamRV6zpnxiiOF+bpsY8FsZllZWZaQkFAotmzZMmvatKkdc0zhS3RgR41ly5bFdIzExES77777b
PLkyVa3bl079dRT7f7777e1a9cWtFmwYIGZmfXs2dPS09ML/e/
99993kgIqVapk9evXj+k8UP506tTJJkyYYPn5+TZjxgy75ZZbbNu2bTZw4ED74YcfzMxs2rRp1qtXL0tOTr
bq1atbenp6wbP6B0/
ADRs2dI5Ro0aNQv+hd+D+OFjz5s2d2Jw5c+y8886ztLQ0S01NtfT09ILE24OPDUSVmppqZr98qR/
KsmXL7JhjjrEmTZoUimdkZFj16tULzeeMVxxJzNdlF88wm8nneKKKi4uTcfWA/A033GBnn322vf766/
bee+/ZX/7yF/
v73/9uH330kR133HEFD+CPHTvWMjIynP6VKhX+uBITE50FPSquhIQE69Spk3Xq1MmaNWtmw4YNs1dffdUuv
vhiO+2006xFixY2cuRIa9CggSUkJNikSZPs4YcfdhI/4uPj5esHQRDzOW3evNm6detmqamp9te//
tVyc3OtSpUq9s0339j/+3//TyadAFGkpqZaZmamzZ49O3If33x9AOMVRwvzddnDgtkjOzvbvvvuO9u/
f3+hRemBf5LIzs42s1/+S87sl4H2a75foHNzc+3GG2+0G2+80RYsWGDt27e3hx56yF588UXLzc01s1+yZXv
16lXcbwkVSMeOHc3MbM2aNfbWW2/
Znj177M033yz0a4R6xCeq7Ozsgn8R+bV58+YV+vPHH39sGzdutAkTJtipp55aEF+yZMlhHxs4oF+/
fvbMM8/Y559/bieeeKK3XXZ2tu3fv98WLFhQaN/
9devW2ebNmwvm81jG66EW30BUzNdlAz9Pepx11lm2du1ae/nllwtiP//
8sz322GOWkpJi3bp1M7NfBmJ8fLx9+umnhfo/
8cQThf68c+dO2717d6FYbm6uVatWrWBboz59+lhqaqrdc8899tNPPznntH79+mJ5byg/
pk6dKn9JOPDMZvPmzQt+gfh1uy1btthzzz132Mc966yz7IsvvrAZM2YUxNavX2/jxo0r1E4de+/
evc79ARyOm266yZKTk+3yyy+3devWOX+/
aNEie+SRR+yss84yM3Mq840cOdLMzPr27WtmsY3X5OTkCv9P1IgN83XZxi/MHldeeaU9/
fTTNnToUPv6668tJyfHXnvtNZs2bZqNGjWqINEkLS3Nzj//fHvssccsLi7OcnNz7e2333aeN54/
f76ddtppdsEFF1irVq2sUqVKNnHiRFu3bp0NGjTIzH75J8Ynn3zSLrnkEjv+
+ONt0KBBlp6ebsuXL7d33nnHTj75ZBs9evRRvxYova699lrbuXOnnXfeedaiRQvbu3evTZ8+3V5+
+WXLycmxYcOG2bp16ywhIcHOPvtsGz58uG3fvt3+8Y9/WJ06dWzNmjWHddybbrrJxo4da2eccYZdf/
31BdsUHfiXmQNOOukkq1Gjhg0ZMsSuu+46i4uLs7Fjxx7WPxcCB8vNzbWXXnrJLrzwQmvZsmWhSn/
Tp08v2Ar0+uuvtyFDhtgzzzxT8M/OM2bMsDFjxti5555rPXr0MLPYxmuHDh3s5Zdftj/+8Y/
WqVMnS0lJsbPPPvtoXwKUIczXZVzJbM5RMnzbyrVu3Vq2X7duXTBs2LCgdu3aQUJCQtC2bduCbeJ+bf369c
GAAQOCqlWrBjVq1AiGDx8ezJ49u9C2chs2bAiuueaaoEWLFkFycnKQlpYWdO7cOXjllVec15s6dWrQp0+fI
C0tLahSpUqQm5sbDB06NPjqq68K2gwZMiRITk4+/
IuBcmHy5MnBpZdeGrRo0SJISUkJEhISgiZNmgTXXnttsG7duoJ2b775ZnDssccGVapUCXJycoL77rsv+Oc/
/+lsKZSdnR307dvXOU63bt2Cbt26FYp99913Qbdu3YIqVaoEWVlZwd/
+9rfg2WefdV5z2rRpQZcuXYKkpKQgMzOzYCslO2g7xoq4TRGKx/
z584MrrrgiyMnJCRISEoJq1aoFJ598cvDYY48VbIX1008/
BXfeeWfQqFGjoHLlykGDBg2CW265pdBWWUEQfbxu3749+O1vfxtUr149MDPGLg6J+bpsiwsC/
tMBAAAA8OEZZgAAACAEC2YAAAAgBAtmAAAAIAQLZgAAACAEC2YAAAAgBAtmAAAAIAQLZgAAACBE5Ep/
cXFxR/I8SszGjRud2IYNG2Tb/fv3O7GUlBQnNn/+fNm/Ro0aTqxy5cqy7fbt251YzZo1ndjMmTNl/
wsvvFDGS6OS3Aq8vI5rlDzG9dFx6qmnynjPnj2dWNWqVZ1YlSpVZH9V9nr58uWy7bPPPuvE1PdFecC4RnkU
ZVzzCzMAAAAQggUzAAAAEIIFMwAAABAiLoj4QFJpfXZInZfvLTVv3tyJzZ0714mtXLlS9o+Pj3diiYmJTsz
37NqaNWsi9ffFt23b5sT27t0r+3
fo0EHGSyOeiUN5xLh2xTJfK6tWrXJial420/PwMce4vxElJyfL/iq/
xXes+vXrO7FTTjnFiU2bNk32L0sY1yiPeIYZAAAAKCIWzAAAAEAIFswAAABACBbMAAAAQAgWzAAAAECIyJX
+SqtYMnb/+c9/OrHVq1c7sRUrVsj+KkNXVfpLSEiQ/Xfu3OnEfFnXavcL9V59xwKA4qYqk/
7000+R+6v5as+ePbLt0KFDnZjaPUjtPmSmd79Qx1q2bJnsr+Z2X1XAJUuWOLGPP/
7YifkquypqRw+z8ltBECjt+IUZAAAACMGCGQAAAAjBghkAAAAIwYIZAAAACFHmk/
5icdJJJzmxhQsXOrGaNWtGfk1fYoaiEl58SSA///xzpJgqyQoAR4JK8Iul3LUvwU/
Jzs52Ylu2bHFi1atXl/2rVavmxNLS0pyY71x37drlxNQc7It///33sm1UJPcBpQu/
MAMAAAAhWDADAAAAIVgwAwAAACFYMAMAAAAhWDADAAAAIcrlLhkdOnSQ8Y0bNzoxld2ssr7NdBlrlaG9b98
+2d8Xj9q2UiX34/
JliKuysDt27Ih8fACIwrfLhKLmpccee0y2Pfvss53YihUrnFhmZqbsn5SU5MReeuklJ6Z23jAzO//
8852YbwelxYsXOzFVxvuTTz6R/
W+99VYnNm3aNNlWiWWnEgCHh1+YAQAAgBAsmAEAAIAQLJgBAACAECyYAQAAgBDlMunvhBNOkHFVhlqVaq1R
o4bsr0pbq+Q8X7ns1NRUGVfUufrKsioq4YSkPwBFoRKf1RzoS45TiWzp6emy7Zo1a5yYmsPy8vJkf/
W6c+fOdWLfffed7D948GAnlp+fL9vu3r3biak5PCsrS/Z/
8803ndiwYcMit1XH2rt3r+wP4PDwCzMAAAAQggUzAAAAEIIFMwAAABCCBTMAAAAQolwm/
Z1zzjkyHrWq39atW2V/VTmqatWqkc9LVerbv3+/
bKuqNPmSCRXfewCAwxW1Wulll10m41WqVHFi69ati3x8lcysEu7MdILfGWec4cS6d+8u+6s5eOnSpbKtSrp
TCZK+RLxNmzY5sSuuuEK2VUl/JPgBRx6/
MAMAAAAhWDADAAAAIVgwAwAAACFYMAMAAAAhWDADAAAAIcrlLhkNGjSQ8Z9+
+smJxbLzhMqEVmW0VSa3mdnGjRudmK/
ctdpRQ+3o4cswj6WMNo6OWMaaytBXsaPp+OOPl3G1U8xnn30W+XXVuPZR10DdK2bR74Fq1arJ+LZt2yKfFw
pTZaXNzHbt2uXEfDtvqM9PxRISEmR/
Nd8nJyc7saZNm8r+am71jVX1PaDelyrt7WubkZEh20blm298OzMBCMcvzAAAAEAIFswAAABACBbMAAAAQAg
WzAAAAECIcpn0l5OTI+NbtmxxYiphSSWLmOmyrqok6ahRo2T/m2++2YmtWLFCtlXJJepcv/rqK9kfpc/
RTLZR48eXNKgSoS699FIn5ktCWr58uRNr27atbPvss886sVjK+qoEP19yX1ZWlhN79NFHndjmzZtl/
wULFjix1157TbZduHChjFcEsYw1VS7al/TnS5A7mC/
Jevv27ZHaLlu2TPZX7yE9PV22VeeqkkZ9CYpKWlqajKtS3h9//HHk1wWKmy8ZtqiJ6h9+
+KETGzNmjGz7wgsvFOlYUfALMwAAABCCBTMAAAAQggUzAAAAEIIFMwAAABCCBTMAAAAQoszvkhF1Nwkzs7y
8vEiv6cvsrFOnjhO7+uqrndjTTz8t+6tdMmIp66syzOfMmSP7o2RF3TngSGUXx9J/
586dTkztCKNKw5uZbdq0yYnVqlVLtn3kkUec2F133eXEVq1aJfur+6JFixaRj1W3bl0nNn78eNm/
Zs2aTuzkk0+WbSvyLhnNmzd3YklJSbKtGpe+nSPUa6g50Lf7jNr9RR1LjXUzsz179jgx344uW7dudWLqvfr
KsKsdPXz32ymnnOLE2CUDR0ssOxUpp512moxPnDjRiW3YsMGJqR2czMwmTJjgxNR9ZabnkSj4hRkAAAAIwY
IZAAAACMGCGQAAAAjBghkAAAAIUeaT/
jp06BC5rSp5rRJGGjVqJPurxIonn3wy8vFjETVB7Pvvvz8ix0fRRE26K2pyX3Ho2bOnE+vfv78TU0l0ZmYX
XHCBE/v0009lW5UwcvfddzsxXxLTzJkzndh1110n26qS3epYzZo1k/
1VaW1fgmFFpsqg+8pVq0Q6X+KzouZw3z2k5stdu3Y5MV9ikOJLFjrmGPe3JxVT52+mz9WXzNitWzcnphJnf
f2BolAJfr7E3f/5n/9xYpdffrlsO23aNCe2ZcsWJ9a7d2/Z/
7777nNi11xzjWyr7s0o+IUZAAAACMGCGQAAAAjBghkAAAAIwYIZAAAACFHmk/
46duwYua16WH3fvn1OzPcAe58+fSIdx1dpUPE9fK6SQHbv3u3EPv/
888jHQjS+6nuxtFWJTKpKmKqSZmZWvXp1J+ZLRlVJTy+//
LJsq6jEDHX+Q4YMkf1VsoVKmDPTSVPr1693YieccILs37lzZyc2adIk2VZVcDv33HOdmO9+VdfA1zaWMVPe
dOrUyYn5Es7UfOebb1WlOxXzJdKpBD/1mfqOr+4r9X1hFr0ypm+
+jzpfmPmTVFG8oiZy+vjugZJOxoxahdZHJXmPGTNGtp09e7YTW7ZsmWyrKnuq78Fnn31W9r/
ppptkXImlMuGv8QszAAAAEIIFMwAAABCCBTMAAAAQggUzAAAAEIIFMwAAABCizO+SkZOT48R8WagqwzklJc
WJ/ec//
5H9fVnLB9u5c2ekdmb+7HoVr127thObO3du5GPBpa6z7zNRmcS+DHu1o4kaqyeeeKLsv23btkivaWbWunVr
J9amTRsn1rhxY9lfva97773XiV199dWy/y233OLEfDt6tGzZ0omprOdFixbJ/
hkZGU7s9NNPl21V1rXa5SI/
P1/2V7sv+HbJqFatmoxXBPXq1XNivix0NTfXqlVLtlVltNVY9R1L7ciidjnw7Xyh+HZJUOel7tc6derI/
ps3b3Zivu8xtatMRRbLDjW+3SDUWFHj4mjucBHLWFO7rPh2j4llR4w333zTibVt29aJ+dYhO3bscGK+70xV
8n3kyJFOLJbdMIobvzADAAAAIVgwAwAAACFYMAMAAAAhWDADAAAAIcp80p8qy7px40bZVj2wr0qaPvPMM0U
/MUElscSSsLB9+/biPB14xJIU4UvEU1Sywpw5c2TbL7/80ompxBQznQh3/
vnnOzFfEskDDzzgxNLT053Y4sWLZf+
+ffs6sbffflu2veGGG5yYSkZUJbB95+BLZlTvVyVTJiYmyv4qkU+VW/
Ydq6LIzMx0Yr4EaXWd1qxZI9uuW7fOialkUvWZmulEKJUgqEpYm+l5wDdfq+TxvLw8J7Zq1SrZX91vvvelx
mXdunWdmLp+5VEs87VP1MRPNdeZmQ0YMMCJqTFhZvbQQw85sf/+979OLJYEQ1+Cn/KHP/
zBiankOjNd2lqN67S0NNlfra981+U3v/
mNE5s4caJsW1SHO2Yq7iwPAAAARMCCGQAAAAjBghkAAAAIwYIZAAAACFHmk/
4aNmzoxFR1GTOd3KMSpo7Ug+ZbtmyJ3FYlrKxdu7Y4Twemk3h8yRYq2caXmHPeeec5saysLCfmGxN///
vfnViNGjVk248//tiJqcSS/v37y/7qPagkpD/+8Y+y/
1/+8hcn1r17d9lWJdesXr3aifk+A1XVUN0rvtdo0qSJE/
MlnY0ZM8aJvfHGG5GPVVGoOdiXeN20aVMn5ptvVQXG4447zomtWLFC9lf3tko6jCXx2pccppKbVEW+r776S
va//fbbndh3330n26pKaaraYnlM+osluVa19VWFbNCggRN74oknnJhK3DfTc5gvIfx///d/
ndi8efOcmBoTZmbVq1d3YgMHDnRi1113neyvvnMuueQS2VYlCGZnZzsx3/3eokULJ+arbjtjxgwZL034hRk
AAAAIwYIZAAAACMGCGQAAAAjBghkAAAAIwYIZAAAACFHmd8lQ5aJr1qwp26pdMlR2686dO4t+YoLKmlUZz2
Y6w/fHH38s9nOq6GLZ3cC3I4aidnNQWe++0thqDPp21Pjkk08itfXtBtC2bVsnpkq13nrrrbJ/
ly5dnJjvukbN3Fe7DpjpMsa+zHV1v99zzz1OzLfzheK7hhW5NLbavUXtEGGmy+Ju2rRJtlX3myoP77v2vt1
TolLlc33l6aP2//
TTT2VbNa58Ozqoc1DzzcyZMw9xhmWPuqa+MsexzO1qp58pU6Y4sUcffVT2P+WUU5yYKpdtZpaTk+PE1O49w
4cPl/
3VvbVkyRIn9swzz8j+K1eudGK+OXTq1KlOTO3IctJJJ8n+CxcudGKLFi2Sbdu0aePEqlat6sROO+002b9+/
fpOTO1+YmY2bNgwGT+UijvLAwAAABGwYAYAAABCsGAGAAAAQrBgBgAAAEKU+aQ/
9QC570HvXbt2OTFfYsWRoMpH+s5VJYEsX7682M+polOJBqp8r5nZhx9+6MS2bt0q286aNcuJqcSO+fPny/
7jx4+XcSUtLc2JdejQwYmp8qtmOjklOTnZifmSDt98800nphLuzHQZZVVqVd2rZjrJ15f08+WXXzqxWBL8V
DKZL5HIdw4VQVJSUuS26ppu2LBBts3IyHBiqjS1LxEzatl7XyJfLJ//Tz/
95MRUcpRq5xNLyff27ds7sXHjxkU+Vlmh7rP09HTZVs1hS5culW0/+
+wzJ3b55Zc7sd69e8v+HTt2dGJr166VbV977TUnphL5VIKzmU6crVatmhNT321mZmeccUbkY3377beRYr5E
PkUllJvp70z1ndW5c2fZX52DSoY0M2vevHnYKXrxCzMAAAAQggUzAAAAEIIFMwAAABCCBTMAAAAQggUzAAA
AEKLM75Lx+eefO7GuXbvKtirD1pdhfSSobF7fLh2qNLDKjvVR76siZ/L79OzZ04mpjGczs0GDBjkxX4a/
+vzULhN//vOfZf8777zTibVs2VK2Vdn0qtRpVlaW7L948WInFrUkqplZ//
79nZjKUDfT56p2GlG7YZiZ7dixQ8aVunXrOjGVJT937lzZf9WqVU6sRYsWsu1f//
rXyOdVlmVmZjoxNdZ98+ru3budmG+sqB1VNm/
e7MR8u1wcifnOV4ZblfdWu3z4svPVDkqxvK9GjRrJtuVNbm6uExsxYoRsO2PGDCdWs2ZN2VZ9Lvn5+U7Mt8
vJ5MmTnZhvhwa104aar9X8ZabvLbVLhm/nCvW+fLu/qF1p1Gfg2xVJ3RfqWpnptZDagec///mP7K/
OtVmzZrKtb7eUQ+EXZgAAACAEC2YAAAAgBAtmAAAAIAQLZgAAACBEmU/6U+WiYymV6kusOBJUud/U1NTI/
ffu3VucpwMze/TRRyO3Pf30052YKklrZnbeeec5MZXw5Ev6vPvuu52YL4mkdu3aTkwl+PlKWzdu3NiJXX/
99U5MJaaYmVWtWtWJJSQkyLbfffedE1OJXL4kJl8yoKJeV5UL/uqrr2T/
vLw8J+ZLGoqlNGxZVr9+fSemkm18yXGqBPDvfvc72VaNV5UgeqSS/tT7UgmOZjppSiWo+hLUVDKZ7/
zVnOFL6C1v0tLSnJgak2b6mvo2BFD3r5qrfPO9+h5XpbXNzH788Ucn9vTTT0c+lkpGVslxOTk5sr9KRlWJg
Gb6Gqr70vfdoPjWZ1HL1qvvUTOz4447zonNmTNHtl29enXYKXrxCzMAAAAQggUzAAAAEIIFMwAAABCCBTMA
AAAQoswn/X322WdOzPcAukqiUEkoR0osCX4qaUZVzUHRtGvXzompqklmZp988okTe//992Xb+++/
P9LxfclR1atXd2K+qkWqAp+qvKTeayzn5buvVBKK71iqotn3338fqZ2Z2ezZs53YsmXLZNui3ttUy3SpaqX
qmvgSc1QinS9pSyVtqsQgX5UxdV6+imaKugd8CUuVK1d2YipBVSWtmekKmL5jqWuoqiKWR998840Tu/
LKK2VblaTdoUMH2faUU05xYirhzZf0O3/+fCf2xhtvyLYqGa9Tp05OzPfdcM4550Rq65uvVdKeL0lbJZjWq
lXLiakxaabvN9/7Uuer7u2mTZvK/ioh+K677pJtDxe/
MAMAAAAhWDADAAAAIVgwAwAAACFYMAMAAAAhWDADAAAAIcr8Lhlr1qxxYr6sa5VNn5ycXOzn5KPKqvqyyVV
2qS8TFYdP7dDQo0cP2VaVtfXtXKLKmH/99ddObN68ebK/et0vvvhCto1q/
PjxRepf0ahdFnw7F1QUqgx7LLuJqGx8X8l2NQeq1/
XNiyob31dGW4ml5LfK8FfvVe06YKbfayw7F9StW1e2rQh8ZZ1ffvnlSDGfOnXqOLHMzEzZtlGjRk6sTZs2s
q3avUftiqR2TjEze/PNN51Y1F1azPQ9VLVqVdlWUeeqvu/M9I4avh2Q1LVV99vrr78u+z/
11FMyrhzuPM4vzAAAAEAIFswAAABACBbMAAAAQAgWzAAAAECIMp/
0pyxcuFDG1UP06mF3XwLFunXrinResZTVLWpZV0SjrumHH34o26q4L4lIlYZu2bKlE7vqqqtkf5UItW3bNt
lWJY6qseobfxs2bHBiWVlZkY5jZpaUlOTEfEkVKpFJtfWVEFaJs76kMXVe6j34El5U/
xUrVsi2sSQTlWXqWqkSwr77Qo3BWBIEFV8ini8eVSylsdX7Vf19SX8q7iv5rRKs1PFTU1Nlf1VCGK68vLxI
MTOzmTNnOrGJEycW9ymhmMSyFvs1fmEGAAAAQrBgBgAAAEKwYAYAAABCsGAGAAAAQrBgBgAAAEKUy10yfJn
IKutaxXwZ+kXdJUNlMvuyNVU2ttohACXLV5b3m2+
+iRQjkxplSUpKihOLZfeeWOYwtauRmtt9O1dELaPt240iFuocYinDXbNmTSfm280i6i4X7du3l/FPP/
008nkB+D/8wgwAAACEYMEMAAAAhGDBDAAAAIRgwQwAAACEKJdJf/
Xr15fxzZs3OzGVrFG5cuXiPiUz00ksvoSZWMqqAsDRoOYwNVdVq1ZN9lfznS8RUB1L8SXXFTURT/
ElaavXVWXYs7OzZf///ve/
Tiw3N1e2VYnqKiG9Tp06sj+Aw8MvzAAAAEAIFswAAABACBbMAAAAQAgWzAAAAECIcpn0p5L7zHRyytGsqLd
gwQInpio8menz2rt3b7GfEwBEVaNGDSe2atUqJ+arlvrOO+84MZUcZ2Y2YsQIJzZz5kwn5ksOjJq87Uvki6
WCoaogqBIBU1NTZf9evXo5senTp8u2GRkZTkx9t9WqVUv2B3B4+IUZAAAACMGCGQAAAAjBghkAAAAIwYIZA
AAACMGCGQAAAAhRLnfJyM/Pl3GV4a3KTderV6/Yz8lM73wRC5UJHcuxfNngABBF06ZNnZial5KSkmR/
tSPGtddeK9uqXTIaNGjgxHbt2iX7q12F1Hzvm1fVLhe+0tpVq1Z1YtWrV3dizz//
vOyvzuv777+XbXNycmQ8yjkBOHz8wgwAAACEYMEMAAAAhGDBDAAAAIRgwQwAAACEKJdJf77kOlWGOiEhwYm
1bdtW9n/77beLdF4qYcRX1lXFY0n6A4DippLuVFnon376Sfb/
5ptvIh9LJa2NHj3aiZ166qmyv0qOW7p0qROLZV5V79XMbO3atU7sxhtvdGLjx4+PfKzHHntMxs844wwnppI
sW7VqFflYAA6NFRgAAAAQggUzAAAAEIIFMwAAABCCBTMAAAAQggUzAAAAEKJc7pLx0ksvyfhxxx3nxDZs2O
DEpkyZUuznZGa2ZcsWJ+bL0N62bZsTmz17duRjUQYbQHHr2LGjE1O7EiUmJsr+qjS2jyp5fdlll0XuH1Xly
pVlvFq1ak5MzeFm/t0zimLmzJkyrsqTp6WlObE1a9YU9ykBFRq/
MAMAAAAhWDADAAAAIVgwAwAAACFYMAMAAAAh4gKywwAAAAAvfmEGAAAAQrBgBgAAAEKwYAYAAABCsGAGAAA
AQrBgBgAAAEKwYAYAAABCsGAGAAAAQrBgBgAAAEKwYI4oLi7O7rjjjoI/P//
88xYXF2dLly4tsXMCSpOlS5daXFycPfjggyV9KqjAmKtREcTFxdmIESMO2Y7xX3zK7YL5wCA58L8qVapYs2
bNbMSIEbZu3bqSPj3gsHz//fc2cOBAy87OtipVqlhWVpb17t3bHnvssZI+NeCwMFcDhZXkPH/PPffY66+/
fsSPUxZVKukTONL++te/
WqNGjWz37t322Wef2ZNPPmmTJk2y2bNnW9WqVUv69IDIpk+fbj169LCGDRvaFVdcYRkZGbZixQr74osv7JF
HHrFrr722pE8ROGzM1UDxz/OXXHKJDRo0yBITEyO1v+eee2zgwIF27rnnHsbZl2/
lfsF85plnWseOHc3M7PLLL7datWrZyJEj7Y033rDBgweX8NkdOTt27LDk5OSSPg0Uo7vvvtvS0tLsyy+/
tOrVqxf6u7y8vJI5qaNs586dLJ7KKeZqoPjn+fj4eIuPjw9tEwSB7d6925KSkmJ+/
Yqk3D6S4dOzZ08zM1uyZIl1797dunfv7rQZOnSo5eTkHNbrP/
HEE9a6dWtLTEy0zMxMu+aaa2zz5s0Ffz9ixAhLSUmxnTt3On0HDx5sGRkZtm/
fvoLY5MmTrWvXrpacnGzVqlWzvn372pw5c5zzTUlJsUWLFtlZZ51l1apVs4suuuiwzh+l16JFi6x169bOJG
pmVqdOnYL//8Czba+//rq1adPGEhMTrXXr1vbuu+86/VatWmWXXnqp1a1bt6DdP//5z0Jt9u7da//7v/
9rHTp0sLS0NEtOTrauXbva1KlTD3nOQRDYlVdeaQkJCTZhwoSC+IsvvmgdOnSwpKQkq1mzpg0aNMhWrFhRq
G/37t2tTZs29vXXX9upp55qVatWtVtvvfWQx0T5wFyNiijqPH/
AoeZ59QxzTk6O9evXz9577z3r2LGjJSUl2dNPP21xcXG2Y8cOGzNmTMEjUkOHDi3md1h2VbgF86JFi8zMrF
atWsX+2nfccYddc801lpmZaQ899JANGDDAnn76aTv99NPtp59+MjOzCy+80Hbs2GHvvPNOob47d+60t956y
wYOHFjwX4Njx461vn37WkpKit133332l7/8xX744Qc75ZRTnAf4f/75Z+vTp4/
VqVPHHnzwQRswYECxvz+UrOzsbPv6669t9uzZh2z72Wef2dVXX22DBg2y+++/
33bv3m0DBgywjRs3FrRZt26ddenSxT744AMbMWKEPfLII9akSRO77LLLbNSoUQXttm7dav/f//f/
Wffu3e2+++6zO+64w9avX299+vSxmTNnes9h3759NnToUHvhhRds4sSJ9pvf/MbMfvkF5Xe/
+501bdrURo4caTfccIN9+OGHduqppxZasJiZbdy40c4880xr3769jRo1ynr06BHTNUPZxVyNiqi453mfefP
m2eDBg6137972yCOPWPv27W3s2LGWmJhoXbt2tbFjx9rYsWNt+PDhxfG2yoegnHruuecCMws+
+OCDYP369cGKFSuC8ePHB7Vq1QqSkpKClStXBt26dQu6devm9B0yZEiQnZ1dKGZmwe233+68/
pIlS4IgCIK8vLwgISEhOP3004N9+/YVtBs9enRgZsE///nPIAiCYP/+/UFWVlYwYMCAQq//
yiuvBGYWfPrpp0EQBMG2bduC6tWrB1dccUWhdmvXrg3S0tIKxYcMGRKYWXDzzTfHeplQhrz//
vtBfHx8EB8fH5x44onBTTfdFLz33nvB3r17C7UzsyAhISFYuHBhQWzWrFmBmQWPPfZYQeyyyy4L6tWrF2zY
sKFQ/0GDBgVpaWnBzp07gyAIgp9//
jnYs2dPoTb5+flB3bp1g0svvbQgtmTJksDMggceeCD46aefggsvvDBISkoK3nvvvYI2S5cuDeLj44O77767
0Ot9/
/33QaVKlQrFu3XrFphZ8NRTT8V6qVCGMFcD/6e45/
mDx38QBEF2dnZgZsG7777rHD85OTkYMmRIsb+v8qDc/
8Lcq1cvS09PtwYNGtigQYMsJSXFJk6caFlZWcV6nA8+
+MD27t1rN9xwgx1zzP9d1iuuuMJSU1MLfqWIi4uz888/3yZNmmTbt28vaPfyyy9bVlaWnXLKKWZmNmXKFNu
8ebMNHjzYNmzYUPC/+Ph469y5s/zn8KuuuqpY3xNKl969e9vnn39u/
fv3t1mzZtn9999vffr0saysLHvzzTcLte3Vq5fl5uYW/PnYY4+11NRUW7x4sZn98qjEv//
9bzv77LMtCIJCY6xPnz62ZcsW++abb8zsl2fgEhISzMxs//79tmnTJvv555+tY8eOBW1+be/evXb+
+efb22+/bZMmTbLTTz+94O8mTJhg+/
fvtwsuuKDQMTMyMqxp06bOuE5MTLRhw4YVzwVEqcZcDRTvPB+mUaNG1qdPn2I///
Ks3Cf9Pf7449asWTOrVKmS1a1b15o3b15okiwuy5YtMzOz5s2bF4onJCRY48aNC/7e7Jd/
6hs1apS9+eab9tvf/ta2b99ukyZNsuHDh1tcXJyZmS1YsMDM/
u85voOlpqYW+nOlSpWsfv36xfZ+UDp16tTJJkyYYHv37rVZs2bZxIkT7eGHH7aBAwfazJkzrVWrVmZm1rBh
Q6dvjRo1LD8/38zM1q9fb5s3b7ZnnnnGnnnmGXmsXyeYjBkzxh566CGbO3duwT9Zm/0y6R7s73//
u23fvt0mT57sPHe6YMECC4LAmjZtKo9ZuXLlQn/
OysoqWKyjfGOuBn5RXPN8GDV3I1y5XzCfcMIJBZnXB4uLi7MgCJz4rxM5joQuXbpYTk6OvfLKK/
bb3/7W3nrrLdu1a5ddeOGFBW32799vZr88G5eRkeG8RqVKhT+6xMTEI/
LlgtIpISHBOnXqZJ06dbJmzZrZsGHD7NVXX7Xbb7/
dzMybFX1gvB8YXxdffLENGTJEtj322GPN7JcEvaFDh9q5555rf/7zn61OnToWHx9vf//73wueM/
21Pn362Lvvvmv333+/de/e3apUqVLwd/
v377e4uDibPHmyPMeUlJRCfyZru+JgrgYKK+o8H4a5NXblfsEcpkaNGvKfLn79C0NU2dnZZvbLg/
SNGzcuiO/du9eWLFlivXr1KtT+ggsusEceecS2bt1qL7/8suXk5FiXLl0K/
v7AP7PUqVPH6Qv82oFFxpo1ayL3SU9Pt2rVqtm+ffsOOb5ee+01a9y4sU2YMKHgVzUzK5i0D9alSxf7/
e9/b/369bPzzz/
fJk6cWLBoyM3NtSAIrFGjRtasWbPI54uKjbkaFd3hzPOH49dzPAqr0P+Zm5uba3PnzrX169cXxGbNmmXTpk
2L+bV69eplCQkJ9uijjxb6r7tnn33WtmzZYn379i3U/sILL7Q9e/bYmDFj7N1337ULLrig0N/
36dPHUlNT7Z577in0T+AH/PqcUTFMnTpV/nIwadIkM3P/iTlMfHy8DRgwwP7973/LbOxfj68Dv2L8+tj//
e9/7fPPP/
e+fq9evWz8+PH27rvv2iWXXFLwK9xvfvMbi4+PtzvvvNN5L0EQRMruRsXDXI2Kojjn+cORnJzs7FaEX1ToX
5gvvfRSGzlypPXp08cuu+wyy8vLs6eeespat25tW7dujem10tPT7ZZbbrE777zTzjjjDOvfv7/
NmzfPnnjiCevUqZNdfPHFhdoff/
zx1qRJE7vttttsz549hf6Jz+yX596efPJJu+SSS+z444+3QYMGWXp6ui1fvtzeeecdO/
nkk2306NFFvgYoO6699lrbuXOnnXfeedaiRQvbu3evTZ8+veBXr1iT4+69916bOnWqde7c2a644gpr1aqVb
dq0yb755hv74IMPbNOmTWZm1q9fP5swYYKdd9551rdvX1uyZIk99dRT1qpVq0LJUAc799xz7bnnnrPf/
e53lpqaak8//bTl5ubaXXfdZbfccostXbrUzj33XKtWrZotWbLEJk6caFdeeaX96U9/
KtJ1QvnDXI2Korjn+Vh16NDBPvjgAxs5cqRlZmZao0aNrHPnzkf0mGVGCezMcVQc2Erlyy+/
DG334osvBo0bNw4SEhKC9u3bB+
+9995hbVV0wOjRo4MWLVoElStXDurWrRtcddVVQX5+vjz2bbfdFphZ0KRJE+/
5TZ06NejTp0+QlpYWVKlSJcjNzQ2GDh0afPXVVwVthgwZEiQnJ4e+T5R9kydPDi699NKgRYsWQUpKSpCQkB
A0adIkuPbaa4N169YVtDOz4JprrnH6Z2dnO9sFrVu3LrjmmmuCBg0aBJUrVw4yMjKC0047LXjmmWcK2uzfv
z+45557guzs7CAxMTE47rjjgrffftu5T369rdyvPfHEE4GZBX/6058KYv/
+97+DU045JUhOTg6Sk5ODFi1aBNdcc00wb968gjbdunULWrdufbiXC2UEczXwf4p7nvdtK9e3b195/
Llz5wannnpqkJSUFJgZW8z9SlwQRHg6HAAAAKigKvQzzAAAAMChsGAGAAAAQrBgBgAAAEKwYAYAAABCsGAG
AAAAQrBgBgAAAEKwYAYAAABCRK70V5bqi6enp8v4Oeec48S2bNnixFasWBH5WCtXrnRilSrpy5qQkODEUlJ
SZNtu3bo5sU8++cSJffPNN4c6xVKvJLcCL0vjGmUL47r4NWzY0ImtWrVKtt23b1+xH3/AgAEy/
u9//7vYj1XUz/BIjT/Gdck6/fTTnViDBg2cmCrTbmbWtm1bJ/aPf/xDtp0/
f74TU59BeSjnEeU98AszAAAAEIIFMwAAABCCBTMAAAAQIi6I+PBJST871KZNGxnv27evE/
M9Q6yeF1ax+Ph42T8/P9+J7dmzx4nt3LlT9k9LS3NivnNVtm/f7sQqV64s286bN8+J/etf/
4p8rKOJZ+JQHpXHcV3U5xfbt2/
vxHbt2iXbZmZmOrGXX37ZiflyVh544AEntn79eieWm5sr+1900UVO7Jhj9G9MEyZMcGIvvfSSExs+fLjsf+
6558q4or6f1Gewf//+yK8Zi/I4rksjlcdkZjZ69Ggntnr1aifmWxv06NHDic2cOVO2Pe6440LO8NDU/
XKkxmVR8QwzAAAAUEQsmAEAAIAQLJgBAACAECyYAQAAgBAsmAEAAIAQZWaXjFtuuUXGVYb1smXLZFuVdZ2d
ne3E1G4YZjoTtVmzZpHameldLho1aiTbqt035syZ48SSk5Nl/
7p16zoxtXOGmdnkyZOd2NHMbiXrGuVReRzXRZ0XNmzY4MQWLFgQ+Viqv2+XCxWP5fzV98h3330n29asWdOJ
Va1aNdLxzfTcrHbp8Dma1dfK47gujR577DEZ7969uxNTu1yoe8XMbNCgQU7so48+km3ff/
99JzZmzBjZtqxjlwwAAACgiFgwAwAAACFYMAMAAAAhWDADAAAAIUpl0l/
9+vWd2A033CDbrly50on99NNPsq1KulPHqlGjhuw/d+7cSK/
pk5GR4cQaNmwo286aNcuJxVKGu3Hjxk4sNTVVtv3rX/
8q40cLSSQojyryuPYlLKlyv2oONzOrVKlSpGOpctdmOpmvSpUqTsw3hyq+MtyqDLFKukpISJD9W7du7cSee
+452fa+++5zYupa/fzzz7J/UVXkcR2LYcOGyfjxxx/vxFSSfkpKiuy/b98+J1a7du3I/
ZW1a9fKeGJiohPbvHmzE9u0aZPsf9NNNzmxvLw82baky2iT9AcAAAAUEQtmAAAAIAQLZgAAACAEC2YAAAAg
BAtmAAAAIESp3CWjXbt2TuzGG2+UbRcuXOjEfKWpt2zZ4sTi4+OdWL169WT/tLQ0J7ZkyRIn5stOVWW4f/
zxR9k26u4b6vzNdMluH3bJAIpfRR7XX331lYyr+Wrbtm2yrdq9Ipb3pXaJ2LFjhxOrVq2a7K/
O1Ze1r143KSnJiandNMzMkpOTnZjve6hRo0YyfjDftSrquKzI49rn5JNPdmJ33HGHbLt3714npnbAUqXVzf
TOFWqXDLUjjJnZ/PnznZhv9xe1DlGfgW/
NM2PGDCd2zTXXyLYljV0yAAAAgCJiwQwAAACEYMEMAAAAhGDBDAAAAISIVnv0KFNJGL4kOJUA4Su9qB6WV2
Umly5dKvur8pUtWrRwYr5z/
e677yId30yfq0osadKkieyvHuJftmyZbAsAh0vNQTVr1pRtVeK1bw5UyUUqMceXiKf6q3nxp59+kv1V0qAv
aU8lXa1bt86JNW3aVPZX5+BL+opaQvhIJf3BNXDgQCe2atUq2VYl46nPyleyfePGjU5MJc76+qt7U5VxN4s
+1nzl6XNycpzYKaecItt+9tlnTizqHHC08AszAAAAEIIFMwAAABCCBTMAAAAQggUzAAAAEKJUJv2pCkdr16
6VbVWFnZNOOkm2/de//uXE1AP46kF3M/2w/ebNm2VbRSV2+BJWKlVyPxr1EL+v6pM6VwAobqqCqS/
hTCV079q1S7ZVydMq5qsypqrn7d6924n55nuV4Ld161bZVlWBjSUhXCU+rly5UrZVc/
6iRYucGMl9xc+XZK82BFAJrj5qHeK7L2rUqOHE1qxZ48RURUEzs8zMTCfmSxBUGw2otYnvHlKJsxdddJFsq
5L+StsY5hdmAAAAIAQLZgAAACAEC2YAAAAgBAtmAAAAIAQLZgAAACBEqdwlQ2WB+rI4586d68S6d+8u2z7z
zDNOLD4+3on5SrWqrGnVX5W1NjNLSkpyYr5M2CVLljgxlWHuy9r98ccfnZjK5DaLXv4SAA52/
PHHR26rdgqqU6eObKt2lFAZ+r45VM3Nag5Wmfy+uG9XItVWfY+o45vp3TcSEhJk29zcXCemdsmgNHbx6927
t4yrXS586wi1s1Ys6wi1FqpevboT27Nnj+yfn5/
vxHzl4dU6QI1L344c6hqkpqbKtmUBvzADAAAAIVgwAwAAACFYMAMAAAAhWDADAAAAIUpl0l/
VqlWdmO8B+Ly8PCd27LHHyrbnnnuuE1MlpH3lS9XD8ioR0Ee19SXtZWRkODH1sL0qS2umE2F8yTUk/
aEoVAninJwc2bZt27ZObPz48ZGPVdSxqhKhSIIqmlatWjkx3zVVn59KeDIzq127thNT830syczqe0TN6762
vu+GWrVqObH169c7MV/S4IYNG5yY77qoktvvv/+
+E2NcF79u3brJuPq+9SW3qXLTKhFQJfmb6VLwKunUV65aJfj5El/V3Kreqy/
BUCW5qk0dzPS4Vps6lCR+YQYAAABCsGAGAAAAQrBgBgAAAEKwYAYAAABCsGAGAAAAQpTKXTIUXya8is+ePV
u2VdmZKuvZdyyVCap2vvBlp6qsZ18mrXpdlR2rMrHN9HvwZV2rDO9169bJtqi4br75Zhm/
6KKLnNiKFStk29atWzsxNdamTp0q+8eyI0YsZe8VtctAr169ZNsPP/
ww8uuWN5mZmU7Mt0ODypr3tVXlftWOGKtXr5b91a5CaocAX1lfVe5Y7V5kpne5UO9V7fxhZrZ8+fLI59W+f
XsZPxi7ZBQ/3/e1+h5Wu32Z+ctQH0ztpmGmx3XUnTfMzJo2berEtm3bJtv6dic7mG/
No+ZrX9uOHTs6MXbJAAAAAMoQFswAAABACBbMAAAAQAgWzAAAAECIUpn0p5LQ1q5dK9uqB+CnTJki26qEDV
/
SnKIe1o+aCGhmVqmSe7kXLVok26rX2LlzpxP77LPPZH+V4OhLeIqlvDdKjirrbFb05J709HQnNm3aNCemkk
XMzP7nf/7HiTVs2FC2VUl/H3zwgRN78803Zf8//OEPTmzp0qWybdQEP998oZJmOnXqJNtW5KS/
3NxcJ+YrtavGmq+srkocVUlzjRs3lv1VGW01B2dnZ8v+qjSxek0zfQ+qpFOVSGgWPUHRTJcQRvFTCW+
+OUUlsvmS29T3bSxzuDqH5ORkJ+b7vlD91X3hE0vitTov3zVUSX8vvvhi5GMdDfzCDAAAAIRgwQwAAACEYM
EMAAAAhGDBDAAAAIQolUl/
6kFxX2JI586dndi9994r21511VVOTD2A7quIpyr3qES8WB72V5UGzczq1KkT6bxWrVol+6uElY0bN0Y+1sq
VK2VbRKMSLlRiRyyJfLEkhqiKVHfeeadsO2zYMCf20EMPObErr7xS9v/9738f+bzU/
aLGmq+i3pIlS5zYRx99JNuOHz/eiQ0ePNiJ1atXT/ZX59WzZ0/
Z1jfnVAQqwdhXKVRVJNu0aZNsq+ZLlaStjm+m52tf9TxFJfj5Ep7UnK+O5buHVaK7b75WSZYofuo6+z4/
NS58Y0V9j6v7wrcOUecQtSKfmU7I9a1ZVFuVIOj7Hou6eYGZP3m3NOEXZgAAACAEC2YAAAAgBAtmAAAAIAQ
LZgAAACAEC2YAAAAgRKncJUOVD/WVCVW7XKiStmY661TFfKWifdmdB/
Pt6KGytn2ZsCprddu2bU5s9uzZsv8ll1zixH788UfZtn79+k7sm2+
+kW0rMvWZ+DKho+5oEcvOFzk5OTL+3HPPObHu3bs7MbVLjJlZ27ZtndiaNWucmNpNw0zvMrFs2TLZVu3eoq
6rb/cXdQ/5dq5QcbUrje9Yareedu3aybYVxfHHH+/
E1Bj2zaELFy50Yr7rr8qQ79q1y4n5dtlQn7U6V19ZYBX33a9Ry7D7zlWNa19b9Z2hynv77kFEU7NmTSfmK/
keyzy+e/
duJ6Z2mfDtXKHKqOfn50eKmZk1a9bMialdOsz0WFPrIN9ONeoe8h2rQYMGMl6a8AszAAAAEIIFMwAAABCCB
TMAAAAQggUzAAAAEKLEk/5UYo1K8PMl3KlSqc2bN5dtU1JSIh1LPVQfi1jKr/qohBVV7tj3YP+cOXOcmK/
UqkoYqSiilrA28yf4HQnXX3+9E/Ml7an3sGDBAif28ccfy/533HGHE7v00kud2IwZM2T/
zMxMJ1a3bl3ZVo1XVdbVV+pVJcx8++23sq1KLlFJh0lJSbK/uo8bNWok2/rmnPJGJeaohLfatWvL/
u+8844T8yUBnXrqqU5M3YO+sry+hOqo1OfvO5b6zlDfLb75Ws3BvmRIlZAbS+ItolFzRSzlrmP5vlDf7WoN
YKbHpUquU+dvpjcl8N0rqq263333hboGvu9XNd7Vddm6davsfzTwCzMAAAAQggUzAAAAEIIFMwAAABCCBTM
AAAAQosST/lRVP/Wwuy8JSFWk8yUcqcRBlZjhq+ajRK0e6OOrMqUebFcJN77Egry8PCemEqbM/
Ne2IlAJCE2aNJFtzz//fCemqi+amR133HGRju+rkNS0aVMn9pe//
EW2Pffcc53Y4MGDnZiv0qO630aNGuXE/
vCHP8j+Y8eOdWIXX3yxbLt27Vonpj6DWCoo+qrKqYRixZekG0v1rqImmJUVqvqZ+qx885qar1X1PjNd6cxX
8VVR87j6vvF9zrGMNXUNVPU+XyKeaqtiZjrBqmPHjk7siy+
+kP0RjRorviq+sSTCqSRj1dY3B6p7QCX4qfM30+PaNweqc1CxLVu2yP6+c1DUHFqnTh0nRtIfAAAAUEqxYA
YAAABCsGAGAAAAQrBgBgAAAEKwYAYAAABClPguGbVq1YrUzpedumHDBifWrl072VbtBpCWlubEVMarmc4OV
ZncPiq7VZXrNjNbvXp1pPNSpSPN9Ln6MlbT09NlvKKaOHGijKtdSh5//
HHZdvbs2U7spJNOcmKqrLSZ2cqVK51Yv379ZNv27ds7sXXr1jkxXwlitStIixYtnJgvw1/
dA76yrtWrV3diKjvat8tCUXdOUDvF+HZJUPebb6cZdb3LIzVfquvv2zVE7SqzefNm2TZq2XrffK36+3YuiN
rft3OBiqvvtmnTpsn+mzZtcmInnniibKvuLd/
OPjh86rvZN9eoMai+w830d7a6L2Ipw63mdt+aSe1045sD1bHUOsK3A5PafUN9B5jpuT0jI8OJLVy4UPY/
GviFGQAAAAjBghkAAAAIwYIZAAAACMGCGQAAAAhR4kl/
qkykerDelwSkHipXr+l7XfWwvu9he1VCWvX3ldZWyRq+c1UP26v+vmNt3LjRidWvX1+29b3fikAl7eXm5sq
2M2fOdGKDBg2SbVXCiBorvvLNvlKliio3rZJAfO9r0aJFTkwlLKmELzNd2thXvlSNV1/
SlxLLWFXHUqWNfQlqKhEmltK45ZGvNPTBfMl1qoSvKgNvpj9rdXxfwpKar1Vb37mq7xzf56/
OVX3f+K5fXl6eE6tdu7Zsq8oQ+5K/
cfhUcpvv+1Z9j6t52Uwneqvva9+GAiq+c+dOJ6YSSc30uIol6U+NtaVLl0bu36VLF9lW3UNqvihJ/
MIMAAAAhGDBDAAAAIRgwQwAAACEYMEMAAAAhGDBDAAAAIQo8V0yopa19WWy5+fnOzFfJrLaZUJl+Pt25Iia
Ie7LmFc7H6jjm+nykevXr498Tr4yyIrKsFWZuOVxN40JEyY4sf79+8u2UUtQm+lsejXWfVnXCQkJTsyXyay
yntVYmzt3ruyvxvuaNWuc2Lx582R/NS5856pEva/M9I4GsZSnj6U0srqPq1atKtu2a9cu8uuWZVE/
a99nsnz5cifWsWNH2VbNjWpc+46lvjNiKZet4r6xqsaKmoN9JazVvRnLfBvLPYRo1Lzo+25Xn5XazcRMl8F
WY0XttGSmv0fUd4BvByb13RTLLmR16tSJdE5meqcQtXuNmb4GvjLaJYVfmAEAAIAQLJgBAACAECyYAQAAgB
AsmAEAAIAQJZ70p8pPqgfIfUl/
KmnKV25alZ+MWprbTCcNqiQgX2JQLOVzVWlilfRXs2ZN2d+XjKb4yr1WBB9+
+KETa9CggWx70003ObGLL75Ytm3btm3RTkyIJQlIjTVfcpRK7FBJICQW+Z1xxhklfQpHhRqDalz5kj7VfNu
oUSPZVs23sYxr9Z0RS2lsxZccpa6Lmld9pX7VfOErbazuw/
KYkF3SVOK1L0lbfa7ffvtt5LaxlDZXn7Waw33rDTV+Ylmb+JL2lAULFjixunXrRm6bnp4e+VhHA78wAwAAA
CFYMAMAAAAhWDADAAAAIVgwAwAAACFKPOlPJbepZAn1ULtZbNW4Vq9e7cTUA/
RpaWmyf15enhNTCSe+5CjV1lfNR10D1d+XLPDuu+86MV81MnUN1MP2sSQSlkf3339/
pJiPqpDUqlUr2VZVP6tXr55sG7Uaku8eUolMURNLzHRFNl8iqYrv3r3bifkSsdR5+RKe1P2i3oMvuUpVqfI
da+rUqU7s5ptvlm3LMvX+VXKdb6ycffbZTsyX2KPGRdSxGst5+caaShD0HUvN+Srmuy5ZWVkyrkQd1ygald
zm23xAfdZbt26N3DZq0qiZ3ihBbQigKviamR177LFObMOGDbKtou6XjIwM2VZV9ozlflOJlyWJX5gBAACAE
CyYAQAAgBAsmAEAAIAQLJgBAACAECyYAQAAgBAlvktG1DLYmzdvlv1VW1/
W9cKFCyOdU35+voyrc1CZ3Dt27JD91bn6sm5Vhq6K+bK2VYatb/
cO9RnEUv4S0ahdVlTMzOzjjz8+wmcDxE7NKyrr3Teue/
fu7cTmz58v20YtTRxLaeuoZeDN9G4UvmOp66LmUF8J4o0bNzoxtauOmZ7za9asKdvi8KndKHzf14pvXKnXU
OPK932t+teoUcOJ+caEKjnvKy+v4mp907x5c9l/
+vTpTsy3+4faJcN3XiWldJ0NAAAAUMqwYAYAAABCsGAGAAAAQrBgBgAAAEKUeNKferBdJUuo5Dqf77//
XsbVw+6qrHBmZqbs36BBAyemztX3oLoqIawS7sz8iYcHS0pKknFVclu9f19bX3lxABWXSsxRMZWcZ6YT+er
Xry/bqqQlNS+q45vppC11Xr752ve6ikrmS01NjXws9T3gSxBUSX+xJD4iGnWdfYl4iq+ss/
q+VcmosSTeq/HjO1c1rnzJjCqukv5q
1ap1qFM8JHVvkPQHAAAAlCEsmAEAAIAQLJgBAACAECyYAQAAgBAsmAEAAIAQJb5LhspaVjs/
qB0mzHQW5aBBg2TblStXOrFVq1Y5MV+56Z07dzoxVS7bl9mpslZ9ZbybNGnixNSOHr7yqQ8//
HDk81LZ3L5rAKDiUrsVqd0kfFn3qizutGnTZNvk5ORI/X07RPh2KTiYbwemWMpoRy1tvH79etn/
5JNPdmINGzaUbdVuR2oHJxTNli1bnJja4cLMbNOmTU6sbdu2kY+l1kG+XVqi7iKmyq2bmTVr1syJqZ0vfNQ
uG75dtRo3buzE8vLyZFs1Z6idbkoSvzADAAAAIVgwAwAAACFYMAMAAAAhWDADAAAAIUo86U8lUagEP19y2x
dffOHELrvsMtlWJb1lZGREPpZ64D+WMpPr1q1zYiqxxEwnEagkhHnz5sn+iq/
U5tatW52YL7kBQMWlkttUwpJvrnn22Wed2L333lv0Eyvj1HfWfffdJ9uq7xGVEI6i2bBhgxNTSadmOkn+lF
NOkW3V97ham/
hKo6vNB6pVq+bEfIl4viRXJer6Rp2TmdlZZ53lxFQZbzOd5Fva8AszAAAAEIIFMwAAABCCBTMAAAAQggUzA
AAAEKLEk/5UlTn1oLlq5/PVV18V6ZzKK1+1RFVtMDMz04l98803xX5OAMoOlVyUn5/
vxHxJaGpe8VGJUL7qZ0XhqxQYy7HUa6jzVwmSZmY5OTmRjx+1qiCKRlXx9V1nVZ34mWeekW1/+9vfOrFatW
o5MV9lXpVgmJaW5sR81ftU9TzfWFMJfuoa+BIJJ02a5MS6desm26qEyv/
+97+ybUnhF2YAAAAgBAtmAAAAIAQLZgAAACAEC2YAAAAgBAtmAAAAIESJ75KhdmhQfNnFsVBluIvjdY8WlT
WrMmZj6R/rawCouKKW8PXNKbGUvz1a81Jx7LxR1NdYv369E/
OVRlalhVesWOHE1M4JZro0M1zLli1zYrF8zm+//XbkePv27Z3YscceK/
vXqFHDidWrV8+JqfWOmdnevXudmK+MthqXH374oRP74osvZH+lS5cuMq5271DHL0n8wgwAAACEYMEMAAAAh
GDBDAAAAIRgwQwAAACEKPGkP0U9/
J2YmFjk1y1LCX5KUZNgfNdQJQeoUp8AKrbOnTs7MZUIqErqmvkTmcojX8ltRSVt+RKxVOKkKlfcq1cv2f/
f//
535POqyHJzc51Yw4YNZdvly5c7MZWcZ6ZLyc+cOTNSrDzwlRdX90DNmjWP9OnEhF+YAQAAgBAsmAEAAIAQL
JgBAACAECyYAQAAgBAsmAEAAIAQJb5Lxrp165yYyqJUWaiIzfz582W8UaNGTmzz5s1H+GwAlDXTpk1zYmrX
hq1bt8r+33zzTbGfU2kVyy4ZTz31lBPzlRFXuxotWrTIib3xxhuRjw/Xe+
+958SaN28u265du9aJqd0wfNROM0erNHwYNYZVLJZznTp1qowvWLDAif3nP/+J/
LpHA78wAwAAACFYMAMAAAAhWDADAAAAIVgwAwAAACHigiAISvokAAAAgNKKX5gBAACAECyYAQAAgBAsmAEA
AIAQLJgBAACAECyYAQAAgBAsmAEAAIAQLJgBAACAECyYI4qLi7M77rij4M/PP/
+8xcXF2dKlS0vsnACfoUOHWkpKyiHbde/
e3bp3715sx+3evbu1adOm2F4P+DXGNfCLuLg4GzFixCHbsVYpPuV2wXxgkBz4X5UqVaxZs2Y2YsQIW7duXU
mfHuB44oknLC4uzjp37lzSp1Im3XPPPfb666+X9GngIIzromFcVzzff/
+9DRw40LKzs61KlSqWlZVlvXv3tscee+yIH5vx5lduF8wH/
PWvf7WxY8fa6NGj7aSTTrInn3zSTjzxRNu5c2dJnxpQyLhx4ywnJ8dmzJhhCxcuLOnTKXOY6EsnxnXRMK4r
lunTp1vHjh1t1qxZdsUVV9jo0aPt8ssvt2OOOcYeeeSRmF/
vkksusV27dll2dnak9ow3v0olfQJH2plnnmkdO3Y0M7PLL7/
catWqZSNHjrQ33njDBg8eXMJnd+Ts2LHDkpOTS/o0ENGSJUts+vTpNmHCBBs+fLiNGzfObr/
99pI+LaBIGNdAbO6++25LS0uzL7/80qpXr17o7/Ly8mJ+vfj4eIuPjw9tEwSB7d6925KSkmJ+/Yqk3P/
CfLCePXua2S8Tue85t6FDh1pOTs5hvf4TTzxhrVu3tsTERMvMzLRrrrnGNm/
eXPD3I0aMsJSUFPkL9+DBgy0jI8P27dtXEJs8ebJ17drVkpOTrVq1ata3b1+bM2eOc74pKSm2aNEiO+uss6
xatWp20UUXHdb5o2SMGzfOatSoYX379rWBAwfauHHjnDZLly61uLg4e/
DBB+2ZZ56x3NxcS0xMtE6dOtmXX355yGPMnDnT0tPTrXv37rZ9+3Zvuz179tjtt99uTZo0scTERGvQoIHdd
NNNtmfPnsjv5+uvv7aTTjrJkpKSrFGjRvbUU085bfLy8uyyyy6zunXrWpUqVaxdu3Y2ZswYp92OHTvsxhtv
tAYNGlhiYqI1b97cHnzwQQuCoKBNXFyc7dixw8aMGVPwGNbQoUMjny+ODMY14xqxWbRokbVu3dpZLJuZ1al
Tx4m9/vrr1qZNG0tMTLTWrVvbu+++W+jv1TPMOTk51q9fP3vvvfesY8eOlpSUZE8//
TTj7RAq3IJ50aJFZmZWq1atYn/tO+64w6655hrLzMy0hx56yAYMGGBPP/
20nX766fbTTz+ZmdmFF15oO3bssHfeeadQ3507d9pbb71lAwcOLPivwbFjx1rfvn0tJSXF7rvvPvvLX/
5iP/zwg51yyinOA/w///
yz9enTx+rUqWMPPvigDRgwoNjfH46ccePG2W9+8xtLSEiwwYMH24IFC7yLhZdeeskeeOABGz58uN111122d
OlS+81vflMwxpQvv/zSevbsaccdd5xNnjzZmzi1f/9+69+/
vz344IN29tln22OPPWbnnnuuPfzww3bhhRdGei/5+fl21llnWYcOHez++++3+vXr21VXXWX//Oc/
C9rs2rXLunfvbmPHjrWLLrrIHnjgAUtLS7OhQ4cW+mfHIAisf//+9vDDD9sZZ5xhI0eOtObNm9uf//xn+
+Mf/
1jQbuzYsZaYmGhdu3a1sWPH2tixY2348OGRzhdHDuOacY3YZGdn29dff22zZ88+ZNvPPvvMrr76ahs0aJDd
f//
9tnv3bhswYIBt3LjxkH3nzZtngwcPtt69e9sjjzxi7du3Z7wdSlBOPffcc4GZBR988EGwfv36YMWKFcH48e
ODWrVqBUlJScHKlSuDbt26Bd26dXP6DhkyJMjOzi4UM7Pg9ttvd15/
yZIlQRAEQV5eXpCQkBCcfvrpwb59+wrajR49OjCz4J///GcQBEGwf//+ICsrKxgwYECh13/llVcCMws+/
fTTIAiCYNu2bUH16tWDK664olC7tWvXBmlpaYXiQ4YMCcwsuPnmm2O9TCgFvvrqq8DMgilTpgRB8MsYqV+/
fnD99dcXardkyZLAzIJatWoFmzZtKoi/8cYbgZkFb731VkFsyJAhQXJychAEQfDZZ58FqampQd+
+fYPdu3cXes2D74GxY8cGxxxzTPCf//
ynULunnnoqMLNg2rRpoe+lW7dugZkFDz30UEFsz549Qfv27YM6deoEe/
fuDYIgCEaNGhWYWfDiiy8WtNu7d29w4oknBikpKcHWrVuDIAiC119/
PTCz4K677ip0nIEDBwZxcXHBwoULC2LJycnBkCFDQs8PRw/j+heMa8Ti/
fffD+Lj44P4+PjgxBNPDG666abgvffeKxhjB5hZkJCQUGiszJo1KzCz4LHHHiuIHbxWCYIgyM7ODswsePfd
d53jM978yv0vzL169bL09HRr0KCBDRo0yFJSUmzixImWlZVVrMf54IMPbO/
evXbDDTfYMcf832W94oorLDU1teAX5bi4ODv//
PNt0qRJhf758OWXX7asrCw75ZRTzMxsypQptnnzZhs8eLBt2LCh4H/
x8fHWuXNnmzp1qnMOV111VbG+Jxwd48aNs7p161qPHj3M7JcxcuGFF9r48eMLPZ5zwIUXXmg1atQo+HPXrl
3NzGzx4sVO26lTp1qfPn3stNNOswkTJlhiYmLoubz66qvWsmVLa9GiRaFxd+BRJjXuDlapUqVCv0okJCTY8
OHDLS8vz77+
+mszM5s0aZJlZGQUyiOoXLmyXXfddbZ9+3b75JNPCtrFx8fbddddV+gYN954owVBYJMnTz7k+aBkMK5/
wbhGLHr37m2ff/659e/f32bNmmX333+/
9enTx7KysuzNN98s1LZXr16Wm5tb8Odjjz3WUlNT5T1zsEaNGlmfPn2K/fzLs3K/
YH788cdtypQpNnXqVPvhhx9s8eLFR2SQLFu2zMzMmjdvXiiekJBgjRs3Lvh7s1+
+GHbt2lUw+Ldv326TJk2y888/3+Li4szMbMGCBWb2yzPX6enphf73/vvvOw//V6pUyerXr1/
s7wtH1r59+2z8+PHWo0cPW7JkiS1cuNAWLlxonTt3tnXr1tmHH37o9GnYsGGhPx9YZOTn5xeK79692/
r27WvHHXecvfLKK5aQkHDI81mwYIHNmTPHGXPNmjUzs2hJJ5mZmU7C6YH+Bx4lWrZsmTVt2rTQf1yambVs2
bLg7w/838zMTKtWrVpoO5QujGvGNQ5fp06dbMKECZafn28zZsywW265xbZt22YDBw60H374oaDdwfeM2S/
3zcH3jNKoUaNiPeeKoNzvknHCCScU7JJxsLi4uEIJFgeoXz+KU5cuXSwnJ8deeeUV+
+1vf2tvvfWW7dq1q9CzdPv37zezX55hy8jIcF6jUqXCH11iYqIzSaP0+
+ijj2zNmjU2fvx4Gz9+vPP348aNs9NPP71QzJfxfPBYTkxMtLPOOsveeOMNe/fdd61fv36HPJ/9+/
db27ZtbeTIkfLvGzRocMjXABjXQNElJCRYp06drFOnTtasWTMbNmyYvfrqqwU7zUS9ZxR2xIhduV8wh6lRo
4b8p4vD+a/7A3sczps3zxo3blwQ37t3ry1ZssR69epVqP0FF1xgjzzyiG3dutVefvlly8nJsS5duhT8/
YF/
ZqlTp47TF+XHuHHjrE6dOvb44487fzdhwgSbOHGiPfXUU4c1ucXFxdm4cePsnHPOsfPPP98mT558yOpnubm
5NmvWLDvttNMK/rUjVqtXr3a2NZw/f76ZWcHuM9nZ2fbdd9/Z/v37C/2H3ty5cwv+/sD//
eCDD2zbtm2Ffo07uN2B94vSgXHNuEbxOvDD35o1a47ocRhvfhX6J8nc3FybO3eurV+/
viA2a9YsmzZtWsyv1atXL0tISLBHH3200H/
dPfvss7Zlyxbr27dvofYXXnih7dmzx8aMGWPvvvuuXXDBBYX+vk+fPpaammr33HOPzBL/
9TmjbNq1a5dNmDDB+vXrZwMHDnT+N2LECNu2bZvz3FosEhISbMKECdapUyc7+
+yzbcaMGaHtL7jgAlu1apX94x//kOe7Y8eOQx7z559/
tqeffrrgz3v37rWnn37a0tPTrUOHDmZmdtZZZ9natWvt5ZdfLtTvscces5SUFOvWrVtBu3379tno0aMLHeP
hhx+2uLg4O/
PMMwtiycnJhbZwRMlgXDOucfimTp0qfyGeNGmSmbmPfRY3xptfhf6F+dJLL7WRI0danz597LLLLrO8vDx76
qmnrHXr1rZ169aYXis9Pd1uueUWu/POO+2MM86w/v3727x58+yJJ56wTp062cUXX1yo/
fHHH29NmjSx2267zfbs2eNsbZSammpPPvmkXXLJJXb88cfboEGDLD093ZYvX27vvPOOnXzyyc5ki7LlzTff
tG3btln//v3l33fp0sXS09Nt3Lhxkbe+UpKSkuztt9+2nj172plnnmmffPKJtWnTRra95JJL7JVXXrHf//
73NnXqVDv55JNt3759NnfuXHvllVcK9u0Mk5mZaffdd58tXbrUmjVrZi+//
LLNnDnTnnnmGatcubKZmV155ZX29NNP29ChQ+3rr7+2nJwce+2112zatGk2atSogl/
dzj77bOvRo4fddttttnTpUmvXrp29//779sYbb9gNN9xQKOGlQ4cO9sEHH9jIkSMtMzPTGjVqRDnmEsC4Zl
zj8F177bW2c+dOO++886xFixa2d+9emz59esG/RA8bNuyIHp/
xFqLkNug4sg5spfLll1+GtnvxxReDxo0bBwkJCUH79u2D995777C2lTtg9OjRQYsWLYLKlSsHdevWDa666q
ogPz9fHvu2224LzCxo0qSJ9/ymTp0a9OnTJ0hLSwuqVKkS5ObmBkOHDg2++uqrgja/
3moJZcfZZ58dVKlSJdixY4e3zdChQ4PKlSsHGzZsKNh+64EHHnDaHTw+1ZjYsGFD0KpVqyAjIyNYsGBBEAT
u9ltB8Ms2WPfdd1/
QunXrIDExMahRo0bQoUOH4M477wy2bNkS+p66desWtG7dOvjqq6+CE088MahSpUqQnZ0djB492mm7bt26YN
iwYUHt2rWDhISEoG3btsFzzz3ntNu2bVvwhz/8IcjMzAwqV64cNG3aNHjggQeC/
fv3F2o3d+7c4NRTTw2SkpICM2NrpBLCuGZc4/
BNnjw5uPTSS4MWLVoEKSkpQUJCQtCkSZPg2muvDdatW1fQzsyCa665xumfnZ1daIz4tpXr27evPD7jzS8uC
CI8HQ4AAABUUBX6GWYAAADgUFgwAwAAACFYMAMAAAAhWDADAAAAIVgwAwAAACFYMAMAAAAhWDADAAAAISJX
+qO+OI6UktwKvDyM6/j4eCe2b9+
+Ir1mpUru1NCsWTPZtkGDBk6sfv36sq0q61qvXj0nlpycLPurths2bJBtP/nkEyf2xBNPOLGdO3fK/
kXFuEZ5xLgufmpeHDx4sGz7ww8/OLGTTjrJic2bN0/2X758uRPr1KmTbPvBBx84sc8+
+0y2LeuijGt+YQYAAABCsGAGAAAAQrBgBgAAAEKwYAYAAABCxAURn+AvrQ/
bx3JeUZMVVBKVmdmrr77qxNQD9JUrV5b9d+3a5cR69eol215wwQVObP78+bKtcswx7n8L+d5/
SSZxlPTxS+u4VtRnama2f/
9+J1alShUndvPNN8v+7dq1c2Lt27d3YjVr1pT9U1NTZbwo1qxZI+Pq3tyyZYtsq+IrV650Yuedd57sr8ZGL
GOVcY3yiHHt6tixoxPLzs6Wbbt06eLE1Hztu86LFy92YikpKU7s+++/l/
2rVq0q40pWVpYTS0tLc2Kffvqp7P/ll186sc2bN0c+/
tFE0h8AAABQRCyYAQAAgBAsmAEAAIAQLJgBAACAECyYAQAAgBBlZpeMWHYIiIXK/
Fflc83MEhISnJi6LtWqVZP9f/75Zye2e/
du2Xbbtm1O7M4773RiCxculP3LErKuo0lMTJTxPXv2OLFBgwY5sbFjx8r+agypcenbjULdFzVq1JBt1T2gd
tnw3UPqvli7dq1sm5GR4cQ2btzoxI4//
njZv6gY1zgSVNl6dV8dKWVlXBd1l5uLLrrIiak5xcwsOTnZifnmpUWLFjkxtUvGvn37Ih9LzcG+MaGuizq+
md6ZS72uKu1tpudx3/
fIpk2bnNh7770n2x4J7JIBAAAAFBELZgAAACAEC2YAAAAgBAtmAAAAIISbPVBKxZLcp8pUmpmdf/
75TiwzM9OJqYfqzfRD+Bs2bHBiKinDzCw/
Pz9yW5WMeN999zkxX7nscePGObHZs2fLtigbfvrpp8htVRLH9u3bZVuVSKfG5axZs2R/
lXDiK6Ndq1atSMf3JWCoeUCV9jbT10C9L19p761bt8o4yi6VPO4ba0VNbjvzzDOdmC/
B9IwzznBiqiyxmf7OueWWW5zYjz/+KPuvXr1axsubWD6/Sy65xImdfvrpTuy1116T/
ZcuXerEkpKSIh9fJcL51jyqtLRax/iS/tS85ttUQV1D9b4WLFgg+6v3oMp4m5l1797diakk7a++
+kr2Pxr4hRkAAAAIwYIZAAAACMGCGQAAAAjBghkAAAAIwYIZAAAACFFmSmP7qHLRzZo1k2337t3rxHbs2OH
EfO9V7Qagyjk2bdpU9l+xYoUT82XSqjLIKsPfVy45Pj7eif3www+yrcqwPprKSqnVkhZLefhHHnnEiQ0cOF
D2nzNnjhOrX7++E/v+++9l/9q1azsx3+4v9erVc2KrVq1yYlWrVpX969at68R8ZbTVbjfqvvjLX/
4i+997770yHhXjuvRR91AsOzCdd955Mv7oo486MXUP+XYTUOPSd15qXFeuXNmJqfvSTH8PdO7cWbZVO+uU5
XGtdukxMzvrrLOcWJ06dZzYvHnzZH+1jlA7PJj5y1AfrKjlzn2ltXfv3h35nNRnrebmXbt2yf5qZye1jjLT
3xnqvBYvXiz7F3X3F0pjAwAAAEXEghkAAAAIwYIZAAAACMGCGQAAAAhRZpL+Lr74YhlXSRjr1q07IuegHlZ
XSXe+krq+RChFva5KAlDJImY6uUUlXJmZzZw504nddNNNhzjD4lOWk0iOJt+5qus3fvx4J+YrGa+SKFq2bO
nEYkn685U/VeeqSv36klBatWrlxFRpbTOzGjVqRH5dpahjg3Fd+qgka1/
C0pVXXunEVJK5mVl+fr4TU3OwL+FJJe2pEshmupS7GmsqEc1MJ76lpaXJtup6leVxfdJJJ8l4bm6uE1Ofn2
+sLF++3In5vu/
VZ62S43xJfyquzlVtcuDjmxfVsdT7Usmhvra+Y6lk1mXLljkxlYxpZjZ9+nQZj4qkPwAAAKCIWDADAAAAIV
gwAwAAACFYMAMAAAAhomehlTBfJSJfwk9UKonA9/C3qry0Z88eJ+ar3qce7PclBqhjqZivv3oPvspDxx9/
vIyjdIkl2UZVefIlrPjG68F8yRaq8pNKbDEz27lzpxNTiVC+Sn+qf/Xq1WXbW2+91Yk9/PDDTuyLL76Q/
YcOHerEnn/+edkWpY+aG9U90KVLF9n/f/7nf5yYLxFP3UOqAqUav2Y6edtXxVXdWyrBKysrS/
ZfunSpE1PfY2ZmI0aMkPGyqkGDBjKuEtFUFV+VcGmmE59VRT1f3DdfRqWS84qayOdrq/
j6q3PwfY+oSn1qzRRLMmNx4xdmAAAAIAQLZgAAACAEC2YAAAAgBAtmAAAAIAQLZgAAACBEmdklw5edqjJOf
RmbUUtd+jJWo+5SEEsmta/Up4qrmMoiNdPv1ZfxWtQMXRS/
WHZvURo2bOjEfFn3vvjBYil37csQV2NQjT+1I4yZzpD27fKxYMECGT/
YWWedJePvvPOOE2OXjPJHlbU20ztH+O4V9Z2jdsTwZfirssC+cR3LzkyK2pEjPT098rHKsnr16sn4li1bnJ
gqIe77/FUZct+uROqz9o1BRY0htQ7wfa/HssuEaqvGuq+0utqtScXM9K4y6viqnZkew+vXr5dtDxe/
MAMAAAAhWDADAAAAIVgwAwAAACFYMAMAAAAhykzSn6/
Mo3rY3fdQ+IYNG5xYLIlUKpFOJWf5HmpXbX3JTeq81PF9yVXqAXjfe1WJWDVq1HBisSQmoGjUZ+VLEFVtVc
LaRRddJPurhCF1r/iSaVVyiRqrZvpcY0nWiFqq1cysT58+Tuztt992YqoErpnZCy+8EPlYKH188/
DB5s2bJ+MqOS6WcsNqrPvOSc2tsYx1dV6+5C6VuOY7r3/84x9O7Jlnnol8XiWpbt26Tsw3h6lrpa6Tr7S2W
ods3LhRtlVx9Vn7Pn/
1HqImbpvFto5QbVVMlVs3M2vUqJETU0mPZvq6qPe6bds22b9Vq1ZO7JNPPpFtDxe/
MAMAAAAhWDADAAAAIVgwAwAAACFYMAMAAAAhWDADAAAAIcrMLhm+Mo9q54CMjAzZNi8vL1J/
384VKmtaZRf7ynirY8Wym4A6vi8TWu1ysXXrVtlWZaJmZ2c7MXb
JKDvuvffeSDEzvUtA9erVnZhv5wpVAthHjWE1/qZOnSr79+rVy4n5xuXQoUOd2LXXXnuIM/w/
Tz75ZOS2KH1i2QFJUTvFqJLzZrq0sm9uV9R3jm83gai2b98u42pHBfXdWNadeOKJTqxbt26y7UsvveTEGjd
u7MT69u0r+z/44INOzLdzhPpcYylXHXVc+dqpHVV8ZbzVDkLqvkhISJD91U4fPXr0kG2//vprJ/
buu+86sQ4dOsj+6juLXTIAAACAo4gFMwAAABCCBTMAAAAQggUzAAAAEKJUJv2lpKQ4MZUsZKYTO3xJd+p1f
YlwijoHlazhK2EcSxKIot6rL0GxTp06TmzHjh2yrTpf1R8lq6hJTD6qNHZaWlqkmJkePyoxxEyP1xkzZjgx
lbRqphM7fEl/
OTk5Tqxfv35OTJXLNoue5IvyaeHChU7suOOOk23XrFnjxKpWrerEfOWOVdyXTKvuodq1azsxlYhoZlarVi0
n9sMPP8i2Zdnrr7/uxHyJeEOGDHFiN9xwgxP78ssvZX/
13Zqeni7bRi337Ev69H2PH8y3NlDrI19pbDWGfa+rqDGcm5sr2/7ud79zYrfeeqsT++9//yv7v/
POO5HP63DxCzMAAAAQggUzAAAAEIIFMwAAABCCBTMAAAAQolQm/
akkpFgq4vmoh+1VsoWvao2qxqMqEPrOSSUM+ZKI1PtVMV8ypEoCWb58uWz7008/
OTH1sD9KJ5V0p8aFL+FIVflSCbLq/vG13bx5s2yrqgWqsdqkSRPZX90DqnKZmU5keeGFF5xYzZo1ZX8S/
Cq2b775xomdf/75sq0aK1GTs8x09TXf/bZx40Ynpr6z9uzZI/urZDRV7bM8mjlzZuS4+r5ctGiR7P/
b3/7WiY0ZM0a29c1XB/
PN12rNoZLrfJsfqO8G35pHUXOwGr9mOvFaJfKZmV144YVO7JFHHol8XkcDvzADAAAAIVgwAwAAACFYMAMAA
AAhWDADAAAAIVgwAwAAACFK5S4ZqiSkb5cMlXHqy9BfsWKFE6tbt64T82U3q+xStSOGL7s1an8fdQ18JS3V
zge+UpuK2vkApVMsY1BZvXq1E2vUqJET27Rpk+yvyurOmTNHtlXltVUZdlW+10xnY/
vuV5U5rl63V69esv8HH3wg4yi71BzqK/WrdqPw7ZyixrXavcZHjUvf3K52dVHz/a5duyIf/
8cff4zctqyI5bNWHn744chtBw4c6MSaNWsm26r1ifqsfOeqymirnTN840f1r1evnmyrXlf1933fZGVlObFX
XnlFtv3qq69k/
GCx3FexrK+i4BdmAAAAIAQLZgAAACAEC2YAAAAgBAtmAAAAIESZSfrzPVSukv58D3qr8p9NmzZ1Ylu3bpX9
VcKRSgJRD8qb6SSEWEpjq/KV69atk/1nz57txJo3by7bqmQuX5Ilyp+oJdt9Y1WNyzZt2si277//
vhObNm2aE7vttttkf5UIo0q7m+n7VSWMqJKsZiT9lUexJH2p7yFfCeG9e/
c6MTXWfOWuVVvfd14srxuV73ukLCvuhK8wvo0GorZV48dX2lyNi6SkJCfmS/pT/
Tds2CDbqmREdayqVavK/r7XLQq1eYKZ//upOLEqAgAAAEKwYAYAAABCsGAGAAAAQrBgBgAAAEKUyqQ/
VWXO90C3SuzZsmWLbLtq1SonppIGfceKmkTg668STmJ5UF0le/
iSUNTD9p07d5ZtVaU334P1KH2iVrTyfaYqQU8lnPgSplSVM1Vlz8wsPT3dibVs2dKJqWpmZrqime99qUQol
XDjS1hBxdatWzcn5kvuUvdL9erVnVi1atVkf5W46kv6iyUhN6r8/Pwi9a/
o1PXzfTeruVHNYbVr15b9fRVXD+abr9UY9CWNqvWVSib0bRLgS1yMSq15fO/
raCR58gszAAAAEIIFMwAAABCCBTMAAAAQggUzAAAAEIIFMwAAABCiVO6SoTIzVelIM7O6des6sUWLFsm2Kr
tT7ZLhyxhVWZgqkzmW8qexZHaqtr7+27Ztc2KqpKWZvrZq5wOUTlHL/
d53330yXqdOHSemSuX6Ssarce0rV926dWsn1q5dOyemxq/
vdX27XKxcudKJqaztopYVRtmmSmCbmXXv3t2JqZ2WzMxSU1OdmPoe8+2coO4t3w4DakcFdQ/
GsvuL2n2mrDuapbHVzhW+71C15lBrE98cqj5/
NYepMWmmx5oaq2b+nVoO5ltb+HYsK4pYytsXN35hBgAAAEKwYAYAAABCsGAGAAAAQrBgBgAAAEKUmWwXXwK
ESmxQZaHNzCpXruzEYkkMUA/W+x7MV1RihjonM/1gv0oW8CV2qCQEXwnhXbt2OTFVnhxlW79+/
WRcJWaoUqe+sTZz5kwntnHjRtm2RYsWTkyVG/YlrCi+hGB1vzRv3tyJ3X///ZGPhaMnapK0audrq/
z1r3+V8Vjme1UGW5VA9pWwjiV53Fd2/mAqkcznpJNOkvFvvvkm8mtUZOr71vf5Rf1cfN/
XUcug+46jxpqvrUoGVPeA717zrW/
KKn5hBgAAAEKwYAYAAABCsGAGAAAAQrBgBgAAAEKwYAYAAABClMpdMlSZRV8mtMrC9O2SUa1aNSemMo59pR
dVdqmK+fqrrFlfJqyislN9pVbXrl3rxOrVqyfbqvcQS1lVHB2x7AZw7rnnOrGsrCzZX5WQVveVun/
MzN59910nNn/+fNn2iiuucGJdunRxYr7dCNTuHb6s8fT0dCe2ZMkSJ/bKK6/I/
hVZUXeeOFLU5x9Lqdxrr73Wif3xj3+Ubb/
99lsnVqtWLdlWXRcV8+1woeK+kt1qvMeye8jq1aud2GmnnSbbjh49WsZxaGqsmunPSn2P+
+41taOF2s3Ct0uHul98bZWo66DyiF+YAQAAgBAsmAEAAIAQLJgBAACAECyYAQAAgBClMukvIyPDicWSBLRw
4ULZViUtqQfgVUld3zmoh/VjSULxvS/1uqoEsC85TyVd+c5LXUOVRICSFUvC1cSJE53Yd999J9uqz79+/
fqRXtNMJ/0df/zxsq16XZX46it3rfiSm1RyzPLlyyO/bnnju06xJB5HTTjzUWPNd15FPVbfvn2d2PXXX+/
EzjnnHNn/2WefdWL5+fmyrUraU2PYl/
SnElR9n4u6hqo0ty8hfNu2bU5MlaxHdFu3bnVidevWlW0zMzMjtd28ebPsrxL01DogagltM7M2bdrIuHpfK
iHcl6Ba1CThkk4yPhi/MAMAAAAhWDADAAAAIVgwAwAAACFYMAMAAAAhykzSn+/h79atWzux//
znP7Lt+eefH+n4sSRbqAfwfVVvYqlSpR6sV/1r1qwp+6tqh77zUskhqtoiip+vGlQsiaObNm1yYqpK2b/
+9S/Z/95773ViM2bMcGK7du2S/YcNG+bEunXrJtuq5BL1ur7KU+p6xXK/
vffee7Jt1P6xfC6ljW8OPZpVuo7E9Rs8eLCM33zzzU6sXbt2TuxPf/
qT7J+SkuLEFi1aJNuq7yyV9KfamenER19CuEpeV7E9e/bI/
uozqFGjhmyLwnwJqhdffLETmzJlimyr5jb1ujt27JD91aYEDRs2dGIbN26U/
bdv3+7EfMmsKplQjUtfgqFK/
p48ebJsWxbmVn5hBgAAAEKwYAYAAABCsGAGAAAAQrBgBgAAAEKwYAYAAABClMpdMtQODb4Mb1X+1pcdmpqa
Gun4sZSQjqWd2n3D19aXIX0wX3bzli1bnJgqiWqmy2D7MqxRmC9rWmVCq3EVS/
nSUaNGybgaA2r3lKlTp8r+agw++uijTkxlV5uZXXrppU7MV1pdXQN1X/pKWKsdXXyfgSpD/
P7778u2FVnt2rWdmG/+UfPKkdKlSxcndssttzixRo0ayf4PP/
ywE7v77rudmCqXbWa2fv16J9a0aVPZVu00osoF++4LdW/5dipSOxeo70ffsdR94SujXVGoOURdU9/
uP8qSJUtk/Mwzz3RiK1ascGJ5eXmyf1ZWlhNT5++7V9V86/
v81TpArbnUrlxmZvXr13diaucMM7OvvvpKxksTfmEGAAAAQrBgBgAAAEKwYAYAAABCsGAGAAAAQpTKpD+VL
KFKRZuZrVq1KvLr1qlTx4mpZA1fIlYsCVqKSnjyJRiqhAOViKMSQMx0gt/
KlStlW5Uw4LveKMyXjBo1aVMlXJnpcr/Dhw+XbVWp0b/85S9OrG3btrL/
1q1bndiVV17pxNauXSv7b9682Yn5EmzT09OdmCrLqhKTzPTc4CvZre6B7777TrZVykKp1lioEuZmZr/
97W+d2OLFi2VbVS5azYsNGjSQ/
VUJ5+zsbNlWjSGVtKnKwJuZ3XHHHZHOyzcvKur8fXGVTOtLvFZtfXPImjVrnJjvfonKlyBYUfjm8YM1adJE
xtVn4vv8atWq5cTUOsa3IUBOTo4TU2XY1b1qppP2fNS9qZJhfXOwSjz0Jc6S9AcAAACUcSyYAQAAgBAsmAE
AAIAQLJgBAACAECyYAQAAgBClcpcMtUODL4t3/vz5kV9XlQtW2Z2qhLWZ3k1CZaf6suvV+/KVpFTHUq/
r66/azp07V7atXr26E1M7JyC6s88+24mlpaU5MV/WtSrB+s0338i2rVq1cmK9evVyYuvWrZP9f/
zxRyemsr59u8Scd955Tsx3v+7YscOJqbLAbdq0kf1VqVa1+42Z2SuvvCLjFdV1110n42rnCDVXmukdTTZt2
uTEfGNVfa6+z0mVfFflsjt16iT7q50j1G4Evp0v1I4svvLwCxcudGKqtLFvNwZ1LDXWzfQuB2rnBd/
OBaptedsR5kjJzMyUcfVZqx24zPTOSO3atXNin3/+uexfr149J6Z2X/
HN92rNoMafmVnnzp2dmFpz+XYfysjIcGLq+8rMrFIldznqu4YlhV+YAQAAgBAsmAEAAIAQLJgBAACAECyYA
QAAgBClMulPJQz5EvFmzZoV+XVVaWCVsOJLllClKqOW1DTTiXwqFovu3bvLuCqZrZK7zHQiDaWxo5kzZ46M
b9y40YmpJCZfIp1K7EhKSpJtfUlLB8vKypLxDRs2OLGWLVs6MV8JY3Vv+pKj1Lhq1KiRE1OJSWZmPXv2dGL
HHnusbKtKIysq2cSs9CWcFNUXX3wh46oEtfr8zXQinfqsW7RoIfura3366afLtio5SZ2Xryy0mgPVd4tvDl
bzpa+0tRrD6r5U95pZ9HM10wmZqjz9ihUrZH+V4HfXXXfJtmWZ+lxj+b5Wn4lvDp45c6YT831+ar7Mzc11Y
mpDATOzxMREJxbL5gXq8/
e9L1VeW93vvvWCuofVWDXT3y+LFi2SbZWift5R8AszAAAAEIIFMwAAABCCBTMAAAAQggUzAAAAEKJUJv2pB
7V9lYh8FaWUyZMnOzH1UPpPP/
0k+6ukKfVgve9cY3kAXVVaU0kIL774ouyvEmEWL14s23bt2tWJUfnJ1b59eyeWk5Mj26prrfpv2bJF9leVk
3wJR77knoMdf/zxMt68eXMnpqqJ+cavGiu+ZEZVVfKNN95wYioZ18zstddeixSLRXlL7vO56qqr/v/
27tclsjAK4/
jZJmgxiGVQg0nFYnGKCAa7yWZRNFgEsSgTJwgmk9Fm9E+wCBoFQcTgDwwqgiCY3bx7nnP2vTur6zjfTzy8d
+bOzJ2ZlwvPObKufleiwOTKyoqrqdCgekwzPb0uChiq32b1WUWhTXW9qsCRCnOb6ZB4RF3D6jsQvdbLy0tX
i6bKqUClCn2dnp7K41UgeW9vT65tZ+rzrxLwVVPqooYAatqo+g010+//
wMBA8XOpyZyqFk32Vd+raGJwX1+fq6nvsJpeGIkC4SrorkJ/
UUj3Xwf8FO4wAwAAAAk2zAAAAECCDTMAAACQYMMMAAAAJNgwAwAAAIkv2SVDdQNQaU0zs+vr6+LHbTQaf31
O30E0ZlIlhKOEbyfb2tpytaibiOqIodZGyV6VcL66upJrVaeWWq3malH3F5V6VqNSo84X6nVFo60fHx9dbX
FxUa5VVJq7Slea6DV8N1GSXFEddTY2NuRaVZ+fn3e19fV1efzExETxeanPr8rrKhV1ZNnZ2XG17e1tufbp6
cnVlpeXXW1ubk4e39/
f72pqBLaZ2eHhoaupjhpjY2Py+N3dXVnvBFU64oyPj7ta1KlIdX6IRlur6011Senu7pbHn52duZr6rkSjsV
WnkOi5bm5uXE118FLjus30+xX956kuZMpndMOIcIcZAAAASLBhBgAAABJsmAEAAIAEG2YAAAAg8SVDfyqYo
0bqmpnd398XP64agdruI6CrjImMQmMqXFLlfe0Ux8fHrjY7OyvXqhCF+qzUSFMzs9XV1eLzUp/
16+urq0WjVlWQTl0T6jHNdJDm5eVFrp2ZmXG15+dnuVaJgjT4VZVgjLouqxx/cHBQVItMT0/
LugoIRkE4RY2MPzo6crVohHCr1Ljpk5MTuVaFo1QY08ysp6fH1VTo7OHh4U+n+K21el2r60+NJTfTv0sXFx
dyrRqDPTg46Gpq3LaZ2ejoqKup/VEUcFRro8Cd+m+Ympoqfq6uri5Xi77D0X/
GV8IdZgAAACDBhhkAAABIsGEGAAAAEmyYAQAAgAQbZgAAACDx470wNvoRI0kja2trrjYyMiLXLi0tFT9up3
fJiOzv77ua6pKxublZfmIV/
M9Rl61e1yrdbGY2OTnpasPDw0U1M7Pe3l5Xi8aXRiOvfxd1mHh7e3M1lfCORqur7iF3d3dF51RVq8n3z9TO
1zUQ6eTremhoSNZV55Lz83O5tlarudrCwoKrNZtNebx6/9UY79vbW3l8vV4vOicz/
RrUaOxovLzqiKH+b8yqdUv6CCXXNXeYAQAAgAQbZgAAACDBhhkAAABIsGEGAAAAEsWhPwAAAKATcYcZAAAA
SLBhBgAAABJsmAEAAIAEG2YAAAAgwYYZAAAASLBhBgAAABJsmAEAAIAEG2YAAAAgwYYZAAAASPwE3tse9QM
U9Z8AAAAASUVORK5CYII=",
"text/plain": [
"<Figure size 900x900 with 16 Axes>"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# Plot more images\n",
"torch.manual_seed(42)\n",
"fig = plt.figure(figsize=(9, 9))\n",
"rows, cols = 4, 4\n",
"for i in range(1, rows * cols + 1):\n",
" random_idx = torch.randint(0, len(train_data), size=[1]).item()\n",
" img, label = train_data[random_idx]\n",
" fig.add_subplot(rows, cols, i)\n",
" plt.imshow(img.squeeze(), cmap=\"gray\")\n",
" plt.title(class_names[label])\n",
" plt.axis(False);"
]
},
{
"cell_type": "markdown",
"id": "f356fbe9-95b1-4f81-a82d-dc15b3adc06a",
"metadata": {
"id": "f356fbe9-95b1-4f81-a82d-dc15b3adc06a"
},
"source": [
"Hmmm, this dataset doesn't look too aesthetic.\n",
"\n",
"But the principles we're going to learn on how to build a model for it will be
similar across a wide range of computer vision problems.\n",
"\n",
"In essence, taking pixel values and building a model to find patterns in them
to use on future pixel values.\n",
"\n",
"Plus, even for this small dataset (yes, even 60,000 images in deep learning is
considered quite small), could you write a program to classify each one of them?\
n",
"\n",
"You probably could.\n",
 "\n",
"But I think coding a model in PyTorch would be faster.\n",
"\n",
"> **Question:** Do you think the above data can be modeled with only straight
(linear) lines? Or do you think you'd also need non-straight (non-linear) lines?"
]
},
{
"cell_type": "markdown",
"id": "43cdd23d-bd1f-4e8c-ba20-22d2b6ac14b1",
"metadata": {
"id": "43cdd23d-bd1f-4e8c-ba20-22d2b6ac14b1"
},
"source": [
"## 2. Prepare DataLoader\n",
"\n",
"Now we've got a dataset ready to go.\n",
"\n",
"The next step is to prepare it with a [`torch.utils.data.DataLoader`]
(https://pytorch.org/docs/stable/data.html#torch.utils.data.Dataset) or
`DataLoader` for short.\n",
"\n",
"The `DataLoader` does what you think it might do.\n",
"\n",
"It helps load data into a model.\n",
"\n",
"For training and for inference.\n",
"\n",
"It turns a large `Dataset` into a Python iterable of smaller chunks.\n",
"\n",
"These smaller chunks are called **batches** or **mini-batches** and can be set
by the `batch_size` parameter.\n",
"\n",
"Why do this?\n",
"\n",
"Because it's more computationally efficient.\n",
"\n",
"In an ideal world you could do the forward pass and backward pass across all
of your data at once.\n",
"\n",
"But once you start using really large datasets, unless you've got infinite
computing power, it's easier to break them up into batches.\n",
"\n",
"It also gives your model more opportunities to improve.\n",
"\n",
"With **mini-batches** (small portions of the data), gradient descent is
performed more often per epoch (once per mini-batch rather than once per epoch).\
n",
"\n",
"What's a good batch size?\n",
"\n",
"[32 is a good place to
start](https://twitter.com/ylecun/status/989610208497360896?s=20&t=N96J_jotN--
PYuJk2WcjMw) for a fair amount of problems.\n",
"\n",
"But since this is a value you can set (a **hyperparameter**) you can try all
different kinds of values, though generally powers of 2 are used most often (e.g.
32, 64, 128, 256, 512).\n",
"\n",
 "![an example of what a batched dataset looks
like](https://raw.githubusercontent.com/mrdbourke/pytorch-deep-learning/main/
images/03-batching-fashionmnist.png)\n",
"*Batching FashionMNIST with a batch size of 32 and shuffle turned on. A
similar batching process will occur for other datasets but will differ depending on
the batch size.*\n",
"\n",
"Let's create `DataLoader`'s for our training and test sets. "
]
},
{
"cell_type": "code",
"execution_count": 10,
"id": "bb2dbf90-a326-43cb-b25b-71af142fafeb",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "bb2dbf90-a326-43cb-b25b-71af142fafeb",
"outputId": "1f563408-3f50-4e8c-a15f-53e2f918b1ac"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Dataloaders: (<torch.utils.data.dataloader.DataLoader object at
0x7fc991463cd0>, <torch.utils.data.dataloader.DataLoader object at
0x7fc991475120>)\n",
"Length of train dataloader: 1875 batches of 32\n",
"Length of test dataloader: 313 batches of 32\n"
]
}
],
"source": [
"from torch.utils.data import DataLoader\n",
"\n",
"# Setup the batch size hyperparameter\n",
"BATCH_SIZE = 32\n",
"\n",
"# Turn datasets into iterables (batches)\n",
"train_dataloader = DataLoader(train_data, # dataset to turn into iterable\n",
" batch_size=BATCH_SIZE, # how many samples per batch? \n",
" shuffle=True # shuffle data every epoch?\n",
")\n",
"\n",
"test_dataloader = DataLoader(test_data,\n",
" batch_size=BATCH_SIZE,\n",
" shuffle=False # don't necessarily have to shuffle the testing data\n",
")\n",
"\n",
"# Let's check out what we've created\n",
"print(f\"Dataloaders: {train_dataloader, test_dataloader}\") \n",
"print(f\"Length of train dataloader: {len(train_dataloader)} batches of
{BATCH_SIZE}\")\n",
"print(f\"Length of test dataloader: {len(test_dataloader)} batches of
{BATCH_SIZE}\")"
]
},
{
"cell_type": "code",
"execution_count": 11,
"id": "7a925ee7-484b-4149-be8f-3ad790172a5f",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "7a925ee7-484b-4149-be8f-3ad790172a5f",
"outputId": "85815bd7-39e9-44ed-b974-9e30fff5b5ad"
},
"outputs": [
{
"data": {
"text/plain": [
"(torch.Size([32, 1, 28, 28]), torch.Size([32]))"
]
},
"execution_count": 11,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Check out what's inside the training dataloader\n",
"train_features_batch, train_labels_batch = next(iter(train_dataloader))\n",
"train_features_batch.shape, train_labels_batch.shape"
]
},
{
"cell_type": "markdown",
"id": "4fee4cf8-ab73-4c81-8e5e-3c81691e799c",
"metadata": {
"id": "4fee4cf8-ab73-4c81-8e5e-3c81691e799c"
},
"source": [
"And we can see that the data remains unchanged by checking a single sample. "
]
},
{
"cell_type": "code",
"execution_count": 12,
"id": "c863d66a-49be-43be-84dc-372a5d6fc2c2",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 463
},
"id": "c863d66a-49be-43be-84dc-372a5d6fc2c2",
"outputId": "1052cbcb-6186-4dfe-b5f0-6968bde9fb21"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Image size: torch.Size([1, 28, 28])\n",
"Label: 6, label size: torch.Size([])\n"
]
 },
{
"data": {
"image/png":
"iVBORw0KGgoAAAANSUhEUgAAAYUAAAGbCAYAAAAr/4yjAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIH
ZlcnNpb24zLjcuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/
bCgiHAAAACXBIWXMAAA9hAAAPYQGoP6dpAAAQuUlEQVR4nO3dX6gfdP3H8fd355ydv9vOGbZl6raT+QcmNm
oqXRitGhJUkC5ICCyCCsu7ugh2mxcSQiRIXim7CDFEulCD6A+EyaJCisniKJktmW7u2DnH8z3/
PL+L4E1Df+28P23f7Zw9Hpd6Xn6/+/
o9PvfV7W1ndXV1NQAgIjZd7CcAwKVDFABIogBAEgUAkigAkEQBgCQKACRRACCJAgBJFLgsdDqd+Pa3v33Or
3v00Uej0+nE3/72twv/pOASJAqse3/+85/j0KFDsXv37hgaGoqrrroqDh48GD/60Y8u+GPff//
98dRTT13wx4Fe6bh9xHr23HPPxYEDB2LXrl1xzz33xPvf//549dVX4/
nnn4+XXnoppqamIuLfnxS+9a1vxUMPPfRf/
3orKyuxtLQUg4OD0el0zvn4Y2NjcejQoXj00UfPxw8HLrr+i/0E4H/x/e9/P7Zt2xa///3vY3x8/Kw/9/
rrr5f/en19fdHX1/dfv2Z1dTW63W4MDw+X//pwqfOvj1jXXnrppdi7d+
+7ghARsWPHjnf9saeeeipuuummGBwcjL1798azzz571p9/r/+msGfPnvjsZz8bP//5z2P//v0xPDwcP/
7xj6PT6cTc3Fw89thj0el0otPpxFe+8pXz/COE3hIF1rXdu3fHH/7wh/jLX/5yzq/97W9/G/
fee2986UtfigceeCC63W7cddddcfr06XNujx8/
HnfffXccPHgwfvjDH8a+ffviyJEjMTg4GLfffnscOXIkjhw5Et/4xjfOxw8LLhr/
+oh17Tvf+U585jOfiX379sWtt94at99+e3zqU5+KAwcOxMDAwFlf++KLL8axY8fi2muvjYiIAwcOxIc//
OH4yU9+cs5fmTQ1NRXPPvts3HHHHWf98W9+85vxwQ9+ML785S+f3x8YXCQ+KbCuHTx4MH73u9/F5z//
+XjhhRfigQceiDvuuCOuuuqq+NnPfnbW137605/OIERE3HzzzbF169Z4+eWXz/
k4k5OT7woCbESiwLp3yy23xJNPPhlnzpyJo0ePxve+972YmZmJQ4cOxbFjx/
Lrdu3a9a7txMREnDlz5pyPMTk5eV6fM1yqRIENY/PmzXHLLbfE/fffHw8//HAsLS3FE088kX/+//
tVRWv5Vdl+pRGXC1FgQ9q/f39ERLz22msX9HHW8nsZYD0RBda1X/3qV+/5M/2nn346IiJuuOGGC/
r4o6OjMT09fUEfA3rJrz5iXbvvvvvi7bffji984Qtx4403xuLiYjz33HPx+OOPx549e+KrX/
3qBX38j370o/GLX/wiHnzwwfjABz4Qk5OTcdttt13Qx4QLSRRY137wgx/EE088EU8//
XQ88sgjsbi4GLt27Yp77703Dh8+/J6/qe18evDBB+PrX/
96HD58OObn5+Oee+4RBdY1t48ASP6bAgBJFABIogBAEgUAkigAkEQBgLTm36fgt/
Nzsezevbu8+fjHP17e/PGPfyxv3ve+95U3v/
71r8ubVi3ft36V+sa1lr+3PikAkEQBgCQKACRRACCJAgBJFABIogBAEgUAkigAkEQBgCQKACRRACCt+f/
R7CBeb7W+3hvxmNnDDz9c3uzdu7e8+elPf1re3HnnneXNQw89VN5EtD2/
jciRv3YO4gFQIgoAJFEAIIkCAEkUAEiiAEASBQCSKACQRAGAJAoAJFEAIIkCAKn/
Yj8B3tulfsBr586d5c0nP/nJpsc6depUeTMyMlLefPe73y1vpqeny5uPfexj5U1ExOnTp8ub48ePlzf//
Oc/y5teutS/
N9Y7nxQASKIAQBIFAJIoAJBEAYAkCgAkUQAgiQIASRQASKIAQBIFAJIoAJBEAYDUWV3jycFOp3Ohnwv/
4aabbmra7du3r7z50Ic+1PRYVZOTk027LVu2lDfXXXddedPymrdccH3+
+efLm4iIbdu2lTfPPPNMedPtdsubf/
zjH+XN0aNHy5uIiFdeeaVpx9ouzPqkAEASBQCSKACQRAGAJAoAJFEAIIkCAEkUAEiiAEASBQCSKACQRAGA5
CBeD9x8883lzRe/+MWmxzp27Fh5s7y8XN688cYb5c3+/fvLm4iIO+
+8s7x57LHHypuvfe1r5U3Lcbarr766vImI+Pvf/
17ePPLII+XN+Ph4eXPFFVeUN9u3by9vItp+TKdPn256rI3GQTwASkQBgCQKACRRACCJAgBJFABIogBAEgUA
kigAkEQBgCQKACRRACA5iNcD9913X3nz5ptvNj1Wy4G2sbGx8qa/
v7+8ef3118ubiIjZ2dnyZuvWreXN3XffXd6cOHGivPnNb35T3kRErKyslDc7d+4sb7rdbnnT8s+HK6+8sry
JiFhcXCxvHn/
88abH2mgcxAOgRBQASKIAQBIFAJIoAJBEAYAkCgAkUQAgiQIASRQASKIAQBIFAFL9qhlle/bsKW/
OnDnT9FgTExNNu17YsWNH027Lli3lzTvvvFPeLC8vlzcvvvhieTMwMFDeRETs2rWrvGk5bjc0NFTetBzr27
Sp7eek119/
fdOOtfFJAYAkCgAkUQAgiQIASRQASKIAQBIFAJIoAJBEAYAkCgAkUQAgiQIAyUG8ohtvvLG8WV1dLW+2bdt
W3kS0HUBrOQQ3Pz9f3nQ6nfImou3Y2vDwcHnTcoTw5MmT5U3Lgb+Itte8v7/+Ld7yfmh5v27durW8iYhYWF
gob2644Yby5vjx4+XNRuCTAgBJFABIogBAEgUAkigAkEQBgCQKACRRACCJAgBJFABIogBAEgUAkigAkFxJL
frEJz5R3rRcW9y8eXN5ExExMTFR3szOzpY309PT5U1fX195ExGxtLRU3oyOjpY3r732WnmzaVPvfl41NzdX
3uzYsaO8GRwcLG927txZ3pw4caK8iWh7j3/
kIx8pb1xJBeCyJwoAJFEAIIkCAEkUAEiiAEASBQCSKACQRAGAJAoAJFEAIIkCAMlBvKLrrruuvPnTn/
5U3kxNTZU3ERG33XZbeTM+Pl7e9PfX3zqnTp0qbyLajgMODAyUN2+++WZ50/
LcxsbGypuIiIWFhfJm69at5U3L+6HlQOIrr7xS3kREXH/99eVNy5G/
y5VPCgAkUQAgiQIASRQASKIAQBIFAJIoAJBEAYAkCgAkUQAgiQIASRQASJf1QbyWw1qzs7PlTV9fX3mzsrJ
S3kREdDqd8mZ5ebm8mZiYKG8WFxfLm4iIbrdb3rQcnWt5zbdt21betBypi2g76tZysK/
lcVr+3o6MjJQ3ERFvvPFGedPy9/
aaa64pb1599dXy5lLjkwIASRQASKIAQBIFAJIoAJBEAYAkCgAkUQAgiQIASRQASKIAQBIFANJlfRDvyiuvL
G9ajrO1HNZqPR539dVXlzdTU1PlzdzcXHnTquU1bzkE12JhYaG8aTmqGNH2OuzcubO8aTke13KAcGBgoLxp
1fI67Nu3r7xxEA+ADUUUAEiiAEASBQCSKACQRAGAJAoAJFEAIIkCAEkUAEiiAEASBQCSKACQOqurq6tr+sJ
O50I/lw1r9+7d5c2WLVuaHutzn/tceTM4OFjenDhxoryZn58vbyIiZmZmypuWK6lr/
FY4S6+u5ka0XRV95513ypvt27eXN9dee21588wzz5Q3EREnT54sb44dO9aTx7nUreU97pMCAEkUAEiiAEAS
BQCSKACQRAGAJAoAJFEAIIkCAEkUAEiiAEASBQCSg3gbzMTERHlz+PDh8uavf/
1refP222+XNxFtR91ajsetrKyUNy3PrWUTETE2NtaTTctr9+STT5Y3U1NT5Q3/
GwfxACgRBQCSKACQRAGAJAoAJFEAIIkCAEkUAEiiAEASBQCSKACQRAGA1H+xn8DF1HLkr1eHAVuPpnW73fJ
mjTcRz9LfX3/
rtGwiIhYXF8ublqNuLcfjTp48Wd4MDQ2VNxERy8vL5U3La9fyOJf6cbuW79uW74uNwCcFAJIoAJBEAYAkCg
AkUQAgiQIASRQASKIAQBIFAJIoAJBEAYAkCgCky/ogXouWQ3W9OqIXETE/
P9+TTcvBuVYtB9pafkwtB9AGBwd78jgREZs3by5vRkdHy5uZmZny5lJ3uR63a+GTAgBJFABIogBAEgUAkig
AkEQBgCQKACRRACCJAgBJFABIogBAEgUA0mV9EK9XR7Iu9WNcCwsL5U1/f/
2t09fXV95ERAwPD5c3Q0ND5U3L4cKWTctRxVYjIyPlzenTpy/
AM2G98EkBgCQKACRRACCJAgBJFABIogBAEgUAkigAkEQBgCQKACRRACCJAgBJFABIl/
WVVP5teXm5vGm5XDo7O1veRLRdcW25XtpyWfVf//
pXebNpU9vPxXp1xXV6erq8YePwSQGAJAoAJFEAIIkCAEkUAEiiAEASBQCSKACQRAGAJAoAJFEAIIkCAMlBP
JqOpvX31986fX195U1E26G6FktLS+VNy3Nreb0j2l7zlsOFLQcS2Th8UgAgiQIASRQASKIAQBIFAJIoAJBE
AYAkCgAkUQAgiQIASRQASKIAQHIQj1hZWSlvNm2q/
3yi9RBcy2ONjo6WNy3H7ebm5sqbxcXF8qZVyxHClvcDG4dPCgAkUQAgiQIASRQASKIAQBIFAJIoAJBEAYAk
CgAkUQAgiQIASRQASA7iEUtLS+XNyMhIedPf3/
Z263a75c3AwEB5s7y8XN5MT0+XN+Pj4+VNRNtxu9bXnMuXTwoAJFEAIIkCAEkUAEiiAEASBQCSKACQRAGAJ
AoAJFEAIIkCAEkUAEiuZdGk0+n0ZBPRdgjuzJkz5c0VV1xR3rQet+uVoaGhnmzYOHxSACCJAgBJFABIogBA
EgUAkigAkEQBgCQKACRRACCJAgBJFABIogBAEgUAkiupxNLSUnnT319/6ywvL5c3EREDAwM92QwPD5c3c3N
z5U232y1vIiIGBwebdlUtrx0bh08KACRRACCJAgBJFABIogBAEgUAkigAkEQBgCQKACRRACCJAgBJFABIDu
IR8/
Pz5U3L8bi+vr7yJiJidna2vOl0Oj15nJmZmfJmZGSkvImIWFlZ6cmm9XAhG4NPCgAkUQAgiQIASRQASKIAQ
BIFAJIoAJBEAYAkCgAkUQAgiQIASRQASA7i0XQ8bvPmzT3ZRLQd3xsfHy9vhoaGyptut9uTx2nV8linTp26
AM/k3VredxERq6ur5/
mZ8J98UgAgiQIASRQASKIAQBIFAJIoAJBEAYAkCgAkUQAgiQIASRQASKIAQHIQj6ajaQsLC+XNli1bypuIi
L6+vvLmrbfeKm9anl8vj9u1GBsbK29aXjs2Dp8UAEiiAEASBQCSKACQRAGAJAoAJFEAIIkCAEkUAEiiAEAS
BQCSKACQRAGA5ErqJarT6TTtVldXy5uZmZny5tZbby1vfvnLX5Y3EREDAwPlTct10NHR0fKm2+2WNy2vd0T
E8PBweTM+Pl7eTE9PlzdsHD4pAJBEAYAkCgAkUQAgiQIASRQASKIAQBIFAJIoAJBEAYAkCgAkUQAgdVbXeE
Gt9UAbG9PevXvLm6WlpabHuuaaa8qbycnJ8mb79u3lzcmTJ8ub1u+lt956q7w5ceJEeXP06NHyhvVhLf+49
0kBgCQKACRRACCJAgBJFABIogBAEgUAkigAkEQBgCQKACRRACCJAgCpf61fuMa7eQCsYz4pAJBEAYAkCgAk
UQAgiQIASRQASKIAQBIFAJIoAJD+DweYWJOnM3TKAAAAAElFTkSuQmCC",
"text/plain": [
"<Figure size 640x480 with 1 Axes>"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# Show a sample\n",
"torch.manual_seed(42)\n",
"random_idx = torch.randint(0, len(train_features_batch), size=[1]).item()\n",
"img, label = train_features_batch[random_idx], train_labels_batch[random_idx]\
n",
"plt.imshow(img.squeeze(), cmap=\"gray\")\n",
"plt.title(class_names[label])\n",
"plt.axis(\"Off\");\n",
"print(f\"Image size: {img.shape}\")\n",
"print(f\"Label: {label}, label size: {label.shape}\")"
]
},
{
"cell_type": "markdown",
 "id": "db1695cf-f53d-4c7c-ad39-dfed76533125",
"metadata": {
"id": "db1695cf-f53d-4c7c-ad39-dfed76533125"
},
"source": [
"## 3. Model 0: Build a baseline model\n",
"\n",
"Data loaded and prepared!\n",
"\n",
"Time to build a **baseline model** by subclassing `nn.Module`.\n",
"\n",
"A **baseline model** is one of the simplest models you can imagine.\n",
"\n",
"You use the baseline as a starting point and try to improve upon it with
subsequent, more complicated models.\n",
"\n",
"Our baseline will consist of two
[`nn.Linear()`](https://pytorch.org/docs/stable/generated/torch.nn.Linear.html)
layers.\n",
"\n",
"We've done this in a previous section but there's going to be one slight
difference.\n",
"\n",
"Because we're working with image data, we're going to use a different layer to
start things off.\n",
"\n",
"And that's the
[`nn.Flatten()`](https://pytorch.org/docs/stable/generated/torch.nn.Flatten.html)
layer.\n",
"\n",
"`nn.Flatten()` compresses the dimensions of a tensor into a single vector.\n",
"\n",
"This is easier to understand when you see it."
]
},
{
"cell_type": "code",
"execution_count": 13,
"id": "405319f1-f242-4bd9-90f5-3abdc50782ac",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "405319f1-f242-4bd9-90f5-3abdc50782ac",
"outputId": "742cd0fe-c95f-4201-a469-f12733625784"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Shape before flattening: torch.Size([1, 28, 28]) -> [color_channels, height,
width]\n",
"Shape after flattening: torch.Size([1, 784]) -> [color_channels,
height*width]\n"
]
}
],
"source": [
 "# Create a flatten layer\n",
"flatten_model = nn.Flatten() # all nn modules function as a model (can do a
forward pass)\n",
"\n",
"# Get a single sample\n",
"x = train_features_batch[0]\n",
"\n",
"# Flatten the sample\n",
"output = flatten_model(x) # perform forward pass\n",
"\n",
"# Print out what happened\n",
"print(f\"Shape before flattening: {x.shape} -> [color_channels, height,
width]\")\n",
"print(f\"Shape after flattening: {output.shape} -> [color_channels,
height*width]\")\n",
"\n",
"# Try uncommenting below and see what happens\n",
"#print(x)\n",
"#print(output)"
]
},
{
"cell_type": "markdown",
"id": "86bb7806-fca6-45af-8111-3e00e38f5be9",
"metadata": {
"id": "86bb7806-fca6-45af-8111-3e00e38f5be9"
},
"source": [
"The `nn.Flatten()` layer took our shape from `[color_channels, height, width]`
to `[color_channels, height*width]`.\n",
"\n",
"Why do this?\n",
"\n",
"Because we've now turned our pixel data from height and width dimensions into
one long **feature vector**.\n",
"\n",
"And `nn.Linear()` layers like their inputs to be in the form of feature
vectors.\n",
"\n",
"Let's create our first model using `nn.Flatten()` as the first layer. "
]
},
{
"cell_type": "code",
"execution_count": 14,
"id": "1449f427-6859-41ae-8133-50b58ffbce72",
"metadata": {
"id": "1449f427-6859-41ae-8133-50b58ffbce72"
},
"outputs": [],
"source": [
"from torch import nn\n",
"class FashionMNISTModelV0(nn.Module):\n",
" def __init__(self, input_shape: int, hidden_units: int, output_shape:
int):\n",
" super().__init__()\n",
" self.layer_stack = nn.Sequential(\n",
" nn.Flatten(), # neural networks like their inputs in vector form\
n",
 " nn.Linear(in_features=input_shape, out_features=hidden_units), #
in_features = number of features in a data sample (784 pixels)\n",
" nn.Linear(in_features=hidden_units, out_features=output_shape)\n",
" )\n",
" \n",
" def forward(self, x):\n",
" return self.layer_stack(x)"
]
},
{
"cell_type": "markdown",
"id": "4d1b50bf-d00b-485c-be00-b3e4de156fab",
"metadata": {
"id": "4d1b50bf-d00b-485c-be00-b3e4de156fab"
},
"source": [
"Wonderful!\n",
"\n",
"We've got a baseline model class we can use, now let's instantiate a model.\
n",
"\n",
"We'll need to set the following parameters:\n",
"* `input_shape=784` - this is how many features you've got going in the model,
in our case, it's one for every pixel in the target image (28 pixels high by 28
pixels wide = 784 features).\n",
"* `hidden_units=10` - number of units/neurons in the hidden layer(s), this
number could be whatever you want but to keep the model small we'll start with
`10`.\n",
"* `output_shape=len(class_names)` - since we're working with a multi-class
classification problem, we need an output neuron per class in our dataset.\n",
"\n",
"Let's create an instance of our model and send to the CPU for now (we'll run a
small test for running `model_0` on CPU vs. a similar model on GPU soon)."
]
},
{
"cell_type": "code",
"execution_count": 15,
"id": "dd18384a-76f9-4b5a-a013-fda077f16865",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "dd18384a-76f9-4b5a-a013-fda077f16865",
"outputId": "e4b63839-d012-40db-a7f7-967a146fe566"
},
"outputs": [
{
"data": {
"text/plain": [
"FashionMNISTModelV0(\n",
" (layer_stack): Sequential(\n",
" (0): Flatten(start_dim=1, end_dim=-1)\n",
" (1): Linear(in_features=784, out_features=10, bias=True)\n",
" (2): Linear(in_features=10, out_features=10, bias=True)\n",
" )\n",
")"
]
},
 "execution_count": 15,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"torch.manual_seed(42)\n",
"\n",
"# Need to setup model with input parameters\n",
"model_0 = FashionMNISTModelV0(input_shape=784, # one for every pixel (28x28)\
n",
" hidden_units=10, # how many units in the hidden layer\n",
" output_shape=len(class_names) # one for every class\n",
")\n",
"model_0.to(\"cpu\") # keep model on CPU to begin with "
]
},
{
"cell_type": "markdown",
"id": "03243179-1cdc-45d9-8b8c-82538ac02e9c",
"metadata": {
"id": "03243179-1cdc-45d9-8b8c-82538ac02e9c"
},
"source": [
"### 3.1 Setup loss, optimizer and evaluation metrics\n",
"\n",
"Since we're working on a classification problem, let's bring in our
[`helper_functions.py`
script](https://github.com/mrdbourke/pytorch-deep-learning/blob/main/
helper_functions.py) and subsequently the `accuracy_fn()` we defined in [notebook
02](https://www.learnpytorch.io/02_pytorch_classification/).\n",
"\n",
"> **Note:** Rather than importing and using our own accuracy function or
evaluation metric(s), you could import various evaluation metrics from the
[TorchMetrics package](https://torchmetrics.readthedocs.io/en/latest/)."
]
},
{
"cell_type": "code",
"execution_count": 16,
"id": "31c91f17-d810-46a4-97c3-c734f93430b1",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "31c91f17-d810-46a4-97c3-c734f93430b1",
"outputId": "d2333811-f5fa-426f-90a7-c884fe2493df"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Downloading helper_functions.py\n"
]
}
],
"source": [
"import requests\n",
 "from pathlib import Path \n",
"\n",
"# Download helper functions from Learn PyTorch repo (if not already
downloaded)\n",
"if Path(\"helper_functions.py\").is_file():\n",
" print(\"helper_functions.py already exists, skipping download\")\n",
"else:\n",
" print(\"Downloading helper_functions.py\")\n",
" # Note: you need the \"raw\" GitHub URL for this to work\n",
" request =
requests.get(\"https://raw.githubusercontent.com/mrdbourke/pytorch-deep-learning/
main/helper_functions.py\")\n",
" with open(\"helper_functions.py\", \"wb\") as f:\n",
" f.write(request.content)"
]
},
{
"cell_type": "code",
"execution_count": 17,
"id": "ce3d13b8-f018-4b44-8bba-375074dc4c5f",
"metadata": {
"id": "ce3d13b8-f018-4b44-8bba-375074dc4c5f"
},
"outputs": [],
"source": [
"# Import accuracy metric\n",
"from helper_functions import accuracy_fn # Note: could also use
torchmetrics.Accuracy(task = 'multiclass',
num_classes=len(class_names)).to(device)\n",
"\n",
"# Setup loss function and optimizer\n",
"loss_fn = nn.CrossEntropyLoss() # this is also called \"criterion\"/\"cost
function\" in some places\n",
"optimizer = torch.optim.SGD(params=model_0.parameters(), lr=0.1)"
]
},
{
"cell_type": "markdown",
"id": "4109f867-83f2-4394-a925-8acdc63ccffe",
"metadata": {
"id": "4109f867-83f2-4394-a925-8acdc63ccffe"
},
"source": [
"### 3.2 Creating a function to time our experiments\n",
"\n",
"Loss function and optimizer ready!\n",
"\n",
"It's time to start training a model.\n",
"\n",
"But how about we do a little experiment while we train.\n",
"\n",
"I mean, let's make a timing function to measure the time it takes our model to
train on CPU versus using a GPU.\n",
"\n",
"We'll train this model on the CPU but the next one on the GPU and see what
happens.\n",
"\n",
"Our timing function will import the [`timeit.default_timer()` function]
(https://docs.python.org/3/library/timeit.html#timeit.default_timer) from the
Python [`timeit` module](https://docs.python.org/3/library/timeit.html)."
]
},
{
"cell_type": "code",
"execution_count": 18,
"id": "31adc3fe-ce90-4b4e-b0d4-3613abae5714",
"metadata": {
"id": "31adc3fe-ce90-4b4e-b0d4-3613abae5714"
},
"outputs": [],
"source": [
"from timeit import default_timer as timer \n",
"def print_train_time(start: float, end: float, device: torch.device = None):\
n",
" \"\"\"Prints difference between start and end time.\n",
"\n",
" Args:\n",
" start (float): Start time of computation (preferred in timeit format).
\n",
" end (float): End time of computation.\n",
" device ([type], optional): Device that compute is running on. Defaults
to None.\n",
"\n",
" Returns:\n",
" float: time between start and end in seconds (higher is longer).\n",
" \"\"\"\n",
" total_time = end - start\n",
" print(f\"Train time on {device}: {total_time:.3f} seconds\")\n",
" return total_time"
]
},
{
"cell_type": "markdown",
"id": "07b9560e-f5dc-45d6-b3b2-ddae17a71b34",
"metadata": {
"id": "07b9560e-f5dc-45d6-b3b2-ddae17a71b34"
},
"source": [
"### 3.3 Creating a training loop and training a model on batches of data\n",
"\n",
"Beautiful!\n",
"\n",
"Looks like we've got all of the pieces of the puzzle ready to go, a timer, a
loss function, an optimizer, a model and most importantly, some data.\n",
"\n",
"Let's now create a training loop and a testing loop to train and evaluate our
model.\n",
"\n",
"We'll be using the same steps as the previous notebook(s), though since our
data is now in batch form, we'll add another loop to loop through our data
batches.\n",
"\n",
"Our data batches are contained within our `DataLoader`s, `train_dataloader`
and `test_dataloader` for the training and test data splits respectively.\n",
"\n",
"A batch is `BATCH_SIZE` samples of `X` (features) and `y` (labels), since
we're using `BATCH_SIZE=32`, our batches have 32 samples of images and targets.\n",
"\n",
 "And since we're computing on batches of data, our loss and evaluation metrics
will be calculated **per batch** rather than across the whole dataset.\n",
"\n",
"This means we'll have to divide our loss and accuracy values by the number of
batches in each dataset's respective dataloader. \n",
"\n",
"Let's step through it: \n",
"1. Loop through epochs.\n",
"2. Loop through training batches, perform training steps, calculate the train
loss *per batch*.\n",
"3. Loop through testing batches, perform testing steps, calculate the test
loss *per batch*.\n",
"4. Print out what's happening.\n",
"5. Time it all (for fun).\n",
"\n",
"A fair few steps but...\n",
"\n",
"...if in doubt, code it out. "
]
},
{
"cell_type": "code",
"execution_count": 19,
"id": "c07bbf10-81e3-47f0-990d-9a4a838276ab",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 587,
"referenced_widgets": [
"0bd8f8b5ff4d4b50b03e3a65cc1446f0",
"430d171cfd584196ad0fa3e1cd0a286c",
"618fb3cf63a94da9ad5f29a3d9a87ac3",
"3524e24faad44aa38926b40b2d590f6b",
"c01ca4def9224135ad367b6f8dbbae62",
"decc1966e6a84973839efc0c65f51790",
"39fc424b6cef4e98a80a342f530be99b",
"e929063168354b018bbf0bb45fdfef1f",
"d62646457b284fcb8aeac382b77eb942",
"5c0883aa74f94568850741dad118cb88",
"e44697d8dd41492e8619a860b3911e19"
]
},
"id": "c07bbf10-81e3-47f0-990d-9a4a838276ab",
"outputId": "3fb70da8-1a65-42bb-a684-85f0d1dd11c0"
},
"outputs": [
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "0bd8f8b5ff4d4b50b03e3a65cc1446f0",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
" 0%| | 0/3 [00:00<?, ?it/s]"
]
},
"metadata": {},
"output_type": "display_data"
 },
{
"name": "stdout",
"output_type": "stream",
"text": [
"Epoch: 0\n",
"-------\n",
"Looked at 0/60000 samples\n",
"Looked at 12800/60000 samples\n",
"Looked at 25600/60000 samples\n",
"Looked at 38400/60000 samples\n",
"Looked at 51200/60000 samples\n",
"\n",
"Train loss: 0.59039 | Test loss: 0.50954, Test acc: 82.04%\n",
"\n",
"Epoch: 1\n",
"-------\n",
"Looked at 0/60000 samples\n",
"Looked at 12800/60000 samples\n",
"Looked at 25600/60000 samples\n",
"Looked at 38400/60000 samples\n",
"Looked at 51200/60000 samples\n",
"\n",
"Train loss: 0.47633 | Test loss: 0.47989, Test acc: 83.20%\n",
"\n",
"Epoch: 2\n",
"-------\n",
"Looked at 0/60000 samples\n",
"Looked at 12800/60000 samples\n",
"Looked at 25600/60000 samples\n",
"Looked at 38400/60000 samples\n",
"Looked at 51200/60000 samples\n",
"\n",
"Train loss: 0.45503 | Test loss: 0.47664, Test acc: 83.43%\n",
"\n",
"Train time on cpu: 32.349 seconds\n"
]
}
],
"source": [
"# Import tqdm for progress bar\n",
"from tqdm.auto import tqdm\n",
"\n",
"# Set the seed and start the timer\n",
"torch.manual_seed(42)\n",
"train_time_start_on_cpu = timer()\n",
"\n",
"# Set the number of epochs (we'll keep this small for faster training times)\
n",
"epochs = 3\n",
"\n",
"# Create training and testing loop\n",
"for epoch in tqdm(range(epochs)):\n",
" print(f\"Epoch: {epoch}\\n-------\")\n",
" ### Training\n",
" train_loss = 0\n",
" # Add a loop to loop through training batches\n",
" for batch, (X, y) in enumerate(train_dataloader):\n",
" model_0.train() \n",
 " # 1. Forward pass\n",
" y_pred = model_0(X)\n",
"\n",
" # 2. Calculate loss (per batch)\n",
" loss = loss_fn(y_pred, y)\n",
" train_loss += loss # accumulatively add up the loss per epoch \n",
"\n",
" # 3. Optimizer zero grad\n",
" optimizer.zero_grad()\n",
"\n",
" # 4. Loss backward\n",
" loss.backward()\n",
"\n",
" # 5. Optimizer step\n",
" optimizer.step()\n",
"\n",
" # Print out how many samples have been seen\n",
" if batch % 400 == 0:\n",
" print(f\"Looked at {batch *
len(X)}/{len(train_dataloader.dataset)} samples\")\n",
"\n",
" # Divide total train loss by length of train dataloader (average loss per
batch per epoch)\n",
" train_loss /= len(train_dataloader)\n",
" \n",
" ### Testing\n",
" # Setup variables for accumulatively adding up loss and accuracy \n",
" test_loss, test_acc = 0, 0 \n",
" model_0.eval()\n",
" with torch.inference_mode():\n",
" for X, y in test_dataloader:\n",
" # 1. Forward pass\n",
" test_pred = model_0(X)\n",
" \n",
" # 2. Calculate loss (accumulatively)\n",
" test_loss += loss_fn(test_pred, y) # accumulatively add up the
loss per epoch\n",
"\n",
" # 3. Calculate accuracy (preds need to be same as y_true)\n",
" test_acc += accuracy_fn(y_true=y, y_pred=test_pred.argmax(dim=1))\
n",
" \n",
" # Calculations on test metrics need to happen inside
torch.inference_mode()\n",
" # Divide total test loss by length of test dataloader (per batch)\n",
" test_loss /= len(test_dataloader)\n",
"\n",
" # Divide total accuracy by length of test dataloader (per batch)\n",
" test_acc /= len(test_dataloader)\n",
"\n",
" ## Print out what's happening\n",
" print(f\"\\nTrain loss: {train_loss:.5f} | Test loss: {test_loss:.5f},
Test acc: {test_acc:.2f}%\\n\")\n",
"\n",
"# Calculate training time \n",
"train_time_end_on_cpu = timer()\n",
"total_train_time_model_0 = print_train_time(start=train_time_start_on_cpu, \
n",
" end=train_time_end_on_cpu,\n",
 "
device=str(next(model_0.parameters()).device))"
]
},
{
"cell_type": "markdown",
"id": "7b02a939-a3a1-4a9d-bb9d-62928def2ded",
"metadata": {
"id": "7b02a939-a3a1-4a9d-bb9d-62928def2ded"
},
"source": [
"Nice! Looks like our baseline model did fairly well.\n",
"\n",
"It didn't take too long to train either, even just on the CPU, I wonder if
it'll speed up on the GPU?\n",
"\n",
"Let's write some code to evaluate our model."
]
},
{
"cell_type": "markdown",
"id": "7442511b-bfe9-4ec7-9f5b-9c808f8e560b",
"metadata": {
"id": "7442511b-bfe9-4ec7-9f5b-9c808f8e560b"
},
"source": [
"## 4. Make predictions and get Model 0 results\n",
"\n",
"Since we're going to be building a few models, it's a good idea to write some
code to evaluate them all in similar ways.\n",
"\n",
"Namely, let's create a function that takes in a trained model, a `DataLoader`,
a loss function and an accuracy function.\n",
"\n",
"The function will use the model to make predictions on the data in the
`DataLoader` and then we can evaluate those predictions using the loss function and
accuracy function."
]
},
{
"cell_type": "code",
"execution_count": 20,
"id": "8317dd04-9de2-4fd7-97bd-1e202621397d",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "8317dd04-9de2-4fd7-97bd-1e202621397d",
"outputId": "7132105e-6f51-43d3-efa7-f0231a21addc"
},
"outputs": [
{
"data": {
"text/plain": [
"{'model_name': 'FashionMNISTModelV0',\n",
" 'model_loss': 0.47663894295692444,\n",
" 'model_acc': 83.42651757188499}"
]
},
 "execution_count": 20,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"torch.manual_seed(42)\n",
"def eval_model(model: torch.nn.Module, \n",
" data_loader: torch.utils.data.DataLoader, \n",
" loss_fn: torch.nn.Module, \n",
" accuracy_fn):\n",
" \"\"\"Returns a dictionary containing the results of model predicting on
data_loader.\n",
"\n",
" Args:\n",
" model (torch.nn.Module): A PyTorch model capable of making predictions
on data_loader.\n",
" data_loader (torch.utils.data.DataLoader): The target dataset to
predict on.\n",
" loss_fn (torch.nn.Module): The loss function of model.\n",
" accuracy_fn: An accuracy function to compare the models predictions to
the truth labels.\n",
"\n",
" Returns:\n",
" (dict): Results of model making predictions on data_loader.\n",
" \"\"\"\n",
" loss, acc = 0, 0\n",
" model.eval()\n",
" with torch.inference_mode():\n",
" for X, y in data_loader:\n",
" # Make predictions with the model\n",
" y_pred = model(X)\n",
" \n",
" # Accumulate the loss and accuracy values per batch\n",
" loss += loss_fn(y_pred, y)\n",
" acc += accuracy_fn(y_true=y, \n",
" y_pred=y_pred.argmax(dim=1)) # For accuracy,
need the prediction labels (logits -> pred_prob -> pred_labels)\n",
" \n",
" # Scale loss and acc to find the average loss/acc per batch\n",
" loss /= len(data_loader)\n",
" acc /= len(data_loader)\n",
" \n",
" return {\"model_name\": model.__class__.__name__, # only works when model
was created with a class\n",
" \"model_loss\": loss.item(),\n",
" \"model_acc\": acc}\n",
"\n",
"# Calculate model 0 results on test dataset\n",
"model_0_results = eval_model(model=model_0, data_loader=test_dataloader,\n",
" loss_fn=loss_fn, accuracy_fn=accuracy_fn\n",
")\n",
"model_0_results"
]
},
{
"cell_type": "markdown",
"id": "a39c3042-1262-4d1f-b33e-c8e2ba6781d3",
"metadata": {
 "id": "a39c3042-1262-4d1f-b33e-c8e2ba6781d3"
},
"source": [
"Looking good!\n",
"\n",
"We can use this dictionary to compare the baseline model results to other
models later on."
]
},
{
"cell_type": "markdown",
"id": "3b76784d-4cdb-43d2-a6da-8e4da9a812a9",
"metadata": {
"id": "3b76784d-4cdb-43d2-a6da-8e4da9a812a9"
},
"source": [
"## 5. Setup device agnostic-code (for using a GPU if there is one)\n",
"We've seen how long it takes to train ma PyTorch model on 60,000 samples on
CPU.\n",
"\n",
"> **Note:** Model training time is dependent on hardware used. Generally, more
processors means faster training and smaller models on smaller datasets will often
train faster than large models and large datasets.\n",
"\n",
"Now let's setup some [device-agnostic
code](https://pytorch.org/docs/stable/notes/cuda.html#best-practices) for our
models and data to run on GPU if it's available.\n",
"\n",
"If you're running this notebook on Google Colab, and you don't have a GPU
turned on yet, it's now time to turn one on via `Runtime -> Change runtime type ->
Hardware accelerator -> GPU`. If you do this, your runtime will likely reset and
you'll have to run all of the cells above by going `Runtime -> Run before`."
]
},
{
"cell_type": "code",
"execution_count": 21,
"id": "17b69fe9-f974-4538-922c-20c5cc8220cc",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 35
},
"id": "17b69fe9-f974-4538-922c-20c5cc8220cc",
"outputId": "10c3b74b-4db7-4a30-8c3a-5a259d1f54b8"
},
"outputs": [
{
"data": {
"application/vnd.google.colaboratory.intrinsic+json": {
"type": "string"
},
"text/plain": [
"'cuda'"
]
},
"execution_count": 21,
"metadata": {},
"output_type": "execute_result"
 }
],
"source": [
"# Setup device agnostic code\n",
"import torch\n",
"device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
"device"
]
},
{
"cell_type": "markdown",
"id": "514021a8-d6f2-47f3-ab50-55f844e42310",
"metadata": {
"id": "514021a8-d6f2-47f3-ab50-55f844e42310"
},
"source": [
"Beautiful!\n",
"\n",
"Let's build another model."
]
},
{
"cell_type": "markdown",
"id": "d7893907-5f82-4c5e-8fde-fa542a9f25af",
"metadata": {
"id": "d7893907-5f82-4c5e-8fde-fa542a9f25af"
},
"source": [
"## 6. Model 1: Building a better model with non-linearity\n",
"\n",
"We learned about [the power of non-linearity in notebook
02](https://www.learnpytorch.io/02_pytorch_classification/#6-the-missing-piece-non-
linearity).\n",
"\n",
"Seeing the data we've been working with, do you think it needs non-linear
functions?\n",
"\n",
"And remember, linear means straight and non-linear means non-straight.\n",
"\n",
"Let's find out.\n",
"\n",
"We'll do so by recreating a similar model to before, except this time we'll
put non-linear functions (`nn.ReLU()`) in between each linear layer."
]
},
{
"cell_type": "code",
"execution_count": 22,
"id": "2ccce5f2-b1e5-47a6-a7f3-6bc096b35ffb",
"metadata": {
"id": "2ccce5f2-b1e5-47a6-a7f3-6bc096b35ffb"
},
"outputs": [],
"source": [
"# Create a model with non-linear and linear layers\n",
"class FashionMNISTModelV1(nn.Module):\n",
" def __init__(self, input_shape: int, hidden_units: int, output_shape:
int):\n",
" super().__init__()\n",
 " self.layer_stack = nn.Sequential(\n",
" nn.Flatten(), # flatten inputs into single vector\n",
" nn.Linear(in_features=input_shape, out_features=hidden_units),\n",
" nn.ReLU(),\n",
" nn.Linear(in_features=hidden_units, out_features=output_shape),\
n",
" nn.ReLU()\n",
" )\n",
" \n",
" def forward(self, x: torch.Tensor):\n",
" return self.layer_stack(x)"
]
},
{
"cell_type": "markdown",
"id": "4b4b7a2f-4834-4aa1-a8e2-b6e3e2b49224",
"metadata": {
"id": "4b4b7a2f-4834-4aa1-a8e2-b6e3e2b49224"
},
"source": [
"That looks good.\n",
"\n",
"Now let's instantiate it with the same settings we used before.\n",
"\n",
"We'll need `input_shape=784` (equal to the number of features of our image
data), `hidden_units=10` (starting small and the same as our baseline model) and
`output_shape=len(class_names)` (one output unit per class).\n",
"\n",
"> **Note:** Notice how we kept most of the settings of our model the same
except for one change: adding non-linear layers. This is a standard practice for
running a series of machine learning experiments, change one thing and see what
happens, then do it again, again, again."
]
},
{
"cell_type": "code",
"execution_count": 23,
"id": "907091ec-7e46-470b-a305-788a3009b837",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "907091ec-7e46-470b-a305-788a3009b837",
"outputId": "4cecd2df-2918-4368-fa33-7aea8f958d8f"
},
"outputs": [
{
"data": {
"text/plain": [
"device(type='cuda', index=0)"
]
},
"execution_count": 23,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"torch.manual_seed(42)\n",
 "model_1 = FashionMNISTModelV1(input_shape=784, # number of input features\n",
" hidden_units=10,\n",
" output_shape=len(class_names) # number of output classes desired\n",
").to(device) # send model to GPU if it's available\n",
"next(model_1.parameters()).device # check model device"
]
},
{
"cell_type": "markdown",
"id": "b54a4e9d-a7ad-404c-920f-485fcff18a92",
"metadata": {
"id": "b54a4e9d-a7ad-404c-920f-485fcff18a92"
},
"source": [
"### 6.1 Setup loss, optimizer and evaluation metrics\n",
"\n",
"As usual, we'll setup a loss function, an optimizer and an evaluation metric
(we could do multiple evaluation metrics but we'll stick with accuracy for now)."
]
},
{
"cell_type": "code",
"execution_count": 24,
"id": "fe7e463b-d46c-4f00-853c-fdf0a28d74c8",
"metadata": {
"id": "fe7e463b-d46c-4f00-853c-fdf0a28d74c8"
},
"outputs": [],
"source": [
"from helper_functions import accuracy_fn\n",
"loss_fn = nn.CrossEntropyLoss()\n",
"optimizer = torch.optim.SGD(params=model_1.parameters(), \n",
" lr=0.1)"
]
},
{
"cell_type": "markdown",
"id": "1eb30af6-a355-49a2-a59f-25169fd27a6e",
"metadata": {
"id": "1eb30af6-a355-49a2-a59f-25169fd27a6e"
},
"source": [
"### 6.2 Functionizing training and test loops\n",
"\n",
"So far we've been writing train and test loops over and over. \n",
"\n",
"Let's write them again but this time we'll put them in functions so they can
be called again and again.\n",
"\n",
"And because we're using device-agnostic code now, we'll be sure to call
`.to(device)` on our feature (`X`) and target (`y`) tensors.\n",
"\n",
"For the training loop we'll create a function called `train_step()` which
takes in a model, a `DataLoader` a loss function and an optimizer.\n",
"\n",
"The testing loop will be similar but it'll be called `test_step()` and it'll
take in a model, a `DataLoader`, a loss function and an evaluation function.\n",
"\n",
"> **Note:** Since these are functions, you can customize them in any way you
like. What we're making here can be considered barebones training and testing
functions for our specific classification use case."
]
},
{
"cell_type": "code",
"execution_count": 25,
"id": "3d239ed2-4028-4603-8db3-ffca2b727819",
"metadata": {
"id": "3d239ed2-4028-4603-8db3-ffca2b727819"
},
"outputs": [],
"source": [
"def train_step(model: torch.nn.Module,\n",
" data_loader: torch.utils.data.DataLoader,\n",
" loss_fn: torch.nn.Module,\n",
" optimizer: torch.optim.Optimizer,\n",
" accuracy_fn,\n",
" device: torch.device = device):\n",
" train_loss, train_acc = 0, 0\n",
" model.to(device)\n",
" for batch, (X, y) in enumerate(data_loader):\n",
" # Send data to GPU\n",
" X, y = X.to(device), y.to(device)\n",
"\n",
" # 1. Forward pass\n",
" y_pred = model(X)\n",
"\n",
" # 2. Calculate loss\n",
" loss = loss_fn(y_pred, y)\n",
" train_loss += loss\n",
" train_acc += accuracy_fn(y_true=y,\n",
" y_pred=y_pred.argmax(dim=1)) # Go from logits
-> pred labels\n",
"\n",
" # 3. Optimizer zero grad\n",
" optimizer.zero_grad()\n",
"\n",
" # 4. Loss backward\n",
" loss.backward()\n",
"\n",
" # 5. Optimizer step\n",
" optimizer.step()\n",
"\n",
" # Calculate loss and accuracy per epoch and print out what's happening\n",
" train_loss /= len(data_loader)\n",
" train_acc /= len(data_loader)\n",
" print(f\"Train loss: {train_loss:.5f} | Train accuracy: {train_acc:.2f}
%\")\n",
"\n",
"def test_step(data_loader: torch.utils.data.DataLoader,\n",
" model: torch.nn.Module,\n",
" loss_fn: torch.nn.Module,\n",
" accuracy_fn,\n",
" device: torch.device = device):\n",
" test_loss, test_acc = 0, 0\n",
" model.to(device)\n",
" model.eval() # put model in eval mode\n",
" # Turn on inference context manager\n",
 " with torch.inference_mode(): \n",
" for X, y in data_loader:\n",
" # Send data to GPU\n",
" X, y = X.to(device), y.to(device)\n",
" \n",
" # 1. Forward pass\n",
" test_pred = model(X)\n",
" \n",
" # 2. Calculate loss and accuracy\n",
" test_loss += loss_fn(test_pred, y)\n",
" test_acc += accuracy_fn(y_true=y,\n",
" y_pred=test_pred.argmax(dim=1) # Go from logits -> pred
labels\n",
" )\n",
" \n",
" # Adjust metrics and print out\n",
" test_loss /= len(data_loader)\n",
" test_acc /= len(data_loader)\n",
" print(f\"Test loss: {test_loss:.5f} | Test accuracy: {test_acc:.2f}%\\
n\")"
]
},
{
"cell_type": "markdown",
"id": "e44121b6-c4be-4909-9175-dc9bd8dc6273",
"metadata": {
"id": "e44121b6-c4be-4909-9175-dc9bd8dc6273"
},
"source": [
"Woohoo!\n",
"\n",
"Now we've got some functions for training and testing our model, let's run
them.\n",
"\n",
"We'll do so inside another loop for each epoch.\n",
"\n",
"That way, for each epoch, we're going through a training step and a testing
step.\n",
"\n",
"> **Note:** You can customize how often you do a testing step. Sometimes
people do them every five epochs or 10 epochs or in our case, every epoch.\n",
"\n",
"Let's also time things to see how long our code takes to run on the GPU."
]
},
{
"cell_type": "code",
"execution_count": 26,
"id": "2bb8094b-01a0-4b84-9526-ba8888d04901",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 327,
"referenced_widgets": [
"3ee8f4a32dae40a2954869aa28d511af",
"9bdbfed6aaa64648ac9624541a719134",
"a7e31e6725a0417495bb5d8d9bb0eedb",
"8a07bf3a83cf44b09ebec23372699dd4",
"4da7f6dcecfc44928a784709a2f85c67",
 "85241944b82749bda4b5b6ff50f484b2",
"b139c87d10be44229d2f65d356912c25",
"b684374f8a3c41cb887142dd2c4a0c94",
"325e5b7b95db4289b3ee1bd6dbfc4a6c",
"987db9e4bab746ff9d393aa1409cf628",
"dd5dcc8d0c424965ba5a329efbf725cc"
]
},
"id": "2bb8094b-01a0-4b84-9526-ba8888d04901",
"outputId": "83769d2d-6f3b-4704-e443-cfc4ef52cc81"
},
"outputs": [
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "3ee8f4a32dae40a2954869aa28d511af",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
" 0%| | 0/3 [00:00<?, ?it/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Epoch: 0\n",
"---------\n",
"Train loss: 1.09199 | Train accuracy: 61.34%\n",
"Test loss: 0.95636 | Test accuracy: 65.00%\n",
"\n",
"Epoch: 1\n",
"---------\n",
"Train loss: 0.78101 | Train accuracy: 71.93%\n",
"Test loss: 0.72227 | Test accuracy: 73.91%\n",
"\n",
"Epoch: 2\n",
"---------\n",
"Train loss: 0.67027 | Train accuracy: 75.94%\n",
"Test loss: 0.68500 | Test accuracy: 75.02%\n",
"\n",
"Train time on cuda: 36.878 seconds\n"
]
}
],
"source": [
"torch.manual_seed(42)\n",
"\n",
"# Measure time\n",
"from timeit import default_timer as timer\n",
"train_time_start_on_gpu = timer()\n",
"\n",
"epochs = 3\n",
"for epoch in tqdm(range(epochs)):\n",
" print(f\"Epoch: {epoch}\\n---------\")\n",
 " train_step(data_loader=train_dataloader, \n",
" model=model_1, \n",
" loss_fn=loss_fn,\n",
" optimizer=optimizer,\n",
" accuracy_fn=accuracy_fn\n",
" )\n",
" test_step(data_loader=test_dataloader,\n",
" model=model_1,\n",
" loss_fn=loss_fn,\n",
" accuracy_fn=accuracy_fn\n",
" )\n",
"\n",
"train_time_end_on_gpu = timer()\n",
"total_train_time_model_1 = print_train_time(start=train_time_start_on_gpu,\n",
" end=train_time_end_on_gpu,\n",
" device=device)"
]
},
{
"cell_type": "markdown",
"id": "719b8eb9-9a7f-42ed-a49f-5eedc6fdd720",
"metadata": {
"id": "719b8eb9-9a7f-42ed-a49f-5eedc6fdd720"
},
"source": [
"Excellent!\n",
"\n",
"Our model trained but the training time took longer?\n",
"\n",
"> **Note:** The training time on CUDA vs CPU will depend largely on the
quality of the CPU/GPU you're using. Read on for a more explained answer.\n",
"\n",
"> **Question:** \"I used a GPU but my model didn't train faster, why might
that be?\"\n",
">\n",
"> **Answer:** Well, one reason could be because your dataset and model are
both so small (like the dataset and model we're working with) the benefits of using
a GPU are outweighed by the time it actually takes to transfer the data there.\n",
"> \n",
"> There's a small bottleneck between copying data from the CPU memory
(default) to the GPU memory.\n",
">\n",
"> So for smaller models and datasets, the CPU might actually be the optimal
place to compute on.\n",
">\n",
"> But for larger datasets and models, the speed of computing the GPU can offer
usually far outweighs the cost of getting the data there.\n",
">\n",
"> However, this is largely dependent on the hardware you're using. With
practice, you will get used to where the best place to train your models is. \n",
"\n",
"Let's evaluate our trained `model_1` using our `eval_model()` function and see
how it went."
]
},
{
"cell_type": "code",
"execution_count": 27,
"id": "32a544e3-9dbe-4aa1-b074-22e28b8f2f2a",
 "metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 381
},
"id": "32a544e3-9dbe-4aa1-b074-22e28b8f2f2a",
"outputId": "bab29648-1e35-4f01-9efe-fa4d2030cddb"
},
"outputs": [
{
"ename": "RuntimeError",
"evalue": "ignored",
"output_type": "error",
"traceback": [
"\
u001b[0;31m------------------------------------------------------------------------
---\u001b[0m",
"\u001b[0;31mRuntimeError\u001b[0m Traceback
(most recent call last)",
"\u001b[0;32m<ipython-input-27-93fed76e63a5>\u001b[0m in \u001b[0;36m<cell
line: 4>\u001b[0;34m()\u001b[0m\n\u001b[1;32m 2\u001b[0m \u001b[0;34m\
u001b[0m\u001b[0m\n\u001b[1;32m 3\u001b[0m \u001b[0;31m# Note: This will error
due to `eval_model()` not using device agnostic code\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 4\u001b[0;31m model_1_results =
eval_model(model=model_1, \n\u001b[0m\u001b[1;32m 5\u001b[0m \
u001b[0mdata_loader\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mtest_dataloader\
u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\
n\u001b[1;32m 6\u001b[0m \u001b[0mloss_fn\u001b[0m\u001b[0;34m=\u001b[0m\
u001b[0mloss_fn\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\
u001b[0m\u001b[0m\n",
"\u001b[0;32m<ipython-input-20-885bc9be9cde>\u001b[0m in \
u001b[0;36meval_model\u001b[0;34m(model, data_loader, loss_fn, accuracy_fn)\
u001b[0m\n\u001b[1;32m 20\u001b[0m \u001b[0;32mfor\u001b[0m \u001b[0mX\
u001b[0m\u001b[0;34m,\u001b[0m \u001b[0my\u001b[0m \u001b[0;32min\u001b[0m \
u001b[0mdata_loader\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 21\u001b[0m \
u001b[0;31m# Make predictions with the model\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 22\u001b[0;31m \
u001b[0my_pred\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mmodel\u001b[0m\
u001b[0;34m(\u001b[0m\u001b[0mX\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\
u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 23\u001b[0m \
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 24\u001b[0m \
u001b[0;31m# Accumulate the loss and accuracy values per batch\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
"\u001b[0;32m/usr/local/lib/python3.10/dist-packages/torch/nn/modules/
module.py\u001b[0m in \u001b[0;36m_call_impl\u001b[0;34m(self, *args, **kwargs)\
u001b[0m\n\u001b[1;32m 1499\u001b[0m \u001b[0;32mor\u001b[0m \
u001b[0m_global_backward_pre_hooks\u001b[0m \u001b[0;32mor\u001b[0m \
u001b[0m_global_backward_hooks\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0m\n\u001b[1;32m 1500\u001b[0m or _global_forward_hooks or
_global_forward_pre_hooks):\n\u001b[0;32m-> 1501\u001b[0;31m \
u001b[0;32mreturn\u001b[0m \u001b[0mforward_call\u001b[0m\u001b[0;34m(\u001b[0m\
u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\
u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 1502\u001b[0m \
u001b[0;31m# Do not call functions when jit is used\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1503\u001b[0m \
u001b[0mfull_backward_hooks\u001b[0m\u001b[0;34m,\u001b[0m \
u001b[0mnon_full_backward_hooks\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m[\
u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m[\u001b[0m\
u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
"\u001b[0;32m<ipython-input-22-a46e692b8bdd>\u001b[0m in \u001b[0;36mforward\
u001b[0;34m(self, x)\u001b[0m\n\u001b[1;32m 12\u001b[0m \u001b[0;34m\u001b[0m\
u001b[0m\n\u001b[1;32m 13\u001b[0m \u001b[0;32mdef\u001b[0m \
u001b[0mforward\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\
u001b[0m \u001b[0mx\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mtorch\u001b[0m\
u001b[0;34m.\u001b[0m\u001b[0mTensor\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\
u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 14\
u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0mself\u001b[0m\
u001b[0;34m.\u001b[0m\u001b[0mlayer_stack\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\
u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\
n\u001b[0m",
"\u001b[0;32m/usr/local/lib/python3.10/dist-packages/torch/nn/modules/
module.py\u001b[0m in \u001b[0;36m_call_impl\u001b[0;34m(self, *args, **kwargs)\
u001b[0m\n\u001b[1;32m 1499\u001b[0m \u001b[0;32mor\u001b[0m \
u001b[0m_global_backward_pre_hooks\u001b[0m \u001b[0;32mor\u001b[0m \
u001b[0m_global_backward_hooks\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0m\n\u001b[1;32m 1500\u001b[0m or _global_forward_hooks or
_global_forward_pre_hooks):\n\u001b[0;32m-> 1501\u001b[0;31m \
u001b[0;32mreturn\u001b[0m \u001b[0mforward_call\u001b[0m\u001b[0;34m(\u001b[0m\
u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\
u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 1502\u001b[0m \
u001b[0;31m# Do not call functions when jit is used\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1503\u001b[0m \
u001b[0mfull_backward_hooks\u001b[0m\u001b[0;34m,\u001b[0m \
u001b[0mnon_full_backward_hooks\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m[\
u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m[\u001b[0m\
u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
"\u001b[0;32m/usr/local/lib/python3.10/dist-packages/torch/nn/modules/
container.py\u001b[0m in \u001b[0;36mforward\u001b[0;34m(self, input)\u001b[0m\n\
u001b[1;32m 215\u001b[0m \u001b[0;32mdef\u001b[0m \u001b[0mforward\u001b[0m\
u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0minput\
u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 216\u001b[0m \
u001b[0;32mfor\u001b[0m \u001b[0mmodule\u001b[0m \u001b[0;32min\u001b[0m \
u001b[0mself\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\
u001b[0m\u001b[0m\n\u001b[0;32m--> 217\u001b[0;31m \u001b[0minput\
u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mmodule\u001b[0m\u001b[0;34m(\u001b[0m\
u001b[0minput\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\
u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 218\u001b[0m \
u001b[0;32mreturn\u001b[0m \u001b[0minput\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 219\u001b[0m \u001b[0;34m\u001b[0m\
u001b[0m\n",
"\u001b[0;32m/usr/local/lib/python3.10/dist-packages/torch/nn/modules/
module.py\u001b[0m in \u001b[0;36m_call_impl\u001b[0;34m(self, *args, **kwargs)\
u001b[0m\n\u001b[1;32m 1499\u001b[0m \u001b[0;32mor\u001b[0m \
u001b[0m_global_backward_pre_hooks\u001b[0m \u001b[0;32mor\u001b[0m \
u001b[0m_global_backward_hooks\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0m\n\u001b[1;32m 1500\u001b[0m or _global_forward_hooks or
_global_forward_pre_hooks):\n\u001b[0;32m-> 1501\u001b[0;31m \
u001b[0;32mreturn\u001b[0m \u001b[0mforward_call\u001b[0m\u001b[0;34m(\u001b[0m\
u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\
u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 1502\u001b[0m \
u001b[0;31m# Do not call functions when jit is used\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1503\u001b[0m \
u001b[0mfull_backward_hooks\u001b[0m\u001b[0;34m,\u001b[0m \
u001b[0mnon_full_backward_hooks\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m[\
u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m[\u001b[0m\
u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
"\u001b[0;32m/usr/local/lib/python3.10/dist-packages/torch/nn/modules/
linear.py\u001b[0m in \u001b[0;36mforward\u001b[0;34m(self, input)\u001b[0m\n\
u001b[1;32m 112\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 113\
u001b[0m \u001b[0;32mdef\u001b[0m \u001b[0mforward\u001b[0m\u001b[0;34m(\
u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0minput\u001b[0m\
u001b[0;34m:\u001b[0m \u001b[0mTensor\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;34m-
>\u001b[0m \u001b[0mTensor\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 114\u001b[0;31m \
u001b[0;32mreturn\u001b[0m \u001b[0mF\u001b[0m\u001b[0;34m.\u001b[0m\
u001b[0mlinear\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0minput\u001b[0m\u001b[0;34m,\
u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mweight\u001b[0m\
u001b[0;34m,\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mbias\
u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\
n\u001b[0m\u001b[1;32m 115\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m
116\u001b[0m \u001b[0;32mdef\u001b[0m \u001b[0mextra_repr\u001b[0m\
u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;34m->\
u001b[0m \u001b[0mstr\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\
u001b[0;34m\u001b[0m\u001b[0m\n",
"\u001b[0;31mRuntimeError\u001b[0m: Expected all tensors to be on the same
device, but found at least two devices, cuda:0 and cpu! (when checking argument for
argument mat1 in method wrapper_CUDA_addmm)"
]
}
],
"source": [
"torch.manual_seed(42)\n",
"\n",
"# Note: This will error due to `eval_model()` not using device agnostic code \
n",
"model_1_results = eval_model(model=model_1, \n",
" data_loader=test_dataloader,\n",
" loss_fn=loss_fn, \n",
" accuracy_fn=accuracy_fn) \n",
"model_1_results "
]
},
{
"cell_type": "markdown",
"id": "6a3481a5-489d-4db9-ac95-c3ce385978b7",
"metadata": {
"id": "6a3481a5-489d-4db9-ac95-c3ce385978b7"
},
"source": [
"Oh no! \n",
"\n",
"It looks like our `eval_model()` function errors out with:\n",
"\n",
"> `RuntimeError: Expected all tensors to be on the same device, but found at
least two devices, cuda:0 and cpu! (when checking argument for argument mat1 in
method wrapper_addmm)`\n",
"\n",
"It's because we've setup our data and model to use device-agnostic code but
not our evaluation function.\n",
"\n",
"How about we fix that by passing a target `device` parameter to our
`eval_model()` function?\n",
 "\n",
"Then we'll try calculating the results again."
]
},
{
"cell_type": "code",
"execution_count": 28,
"id": "f3665d99-1adc-4d9f-bfc6-e5601a80691c",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "f3665d99-1adc-4d9f-bfc6-e5601a80691c",
"outputId": "05312922-d30b-4c09-9989-963a4a579bf8"
},
"outputs": [
{
"data": {
"text/plain": [
"{'model_name': 'FashionMNISTModelV1',\n",
" 'model_loss': 0.6850008964538574,\n",
" 'model_acc': 75.01996805111821}"
]
},
"execution_count": 28,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Move values to device\n",
"torch.manual_seed(42)\n",
"def eval_model(model: torch.nn.Module, \n",
" data_loader: torch.utils.data.DataLoader, \n",
" loss_fn: torch.nn.Module, \n",
" accuracy_fn, \n",
" device: torch.device = device):\n",
" \"\"\"Evaluates a given model on a given dataset.\n",
"\n",
" Args:\n",
" model (torch.nn.Module): A PyTorch model capable of making predictions
on data_loader.\n",
" data_loader (torch.utils.data.DataLoader): The target dataset to
predict on.\n",
" loss_fn (torch.nn.Module): The loss function of model.\n",
" accuracy_fn: An accuracy function to compare the models predictions to
the truth labels.\n",
" device (str, optional): Target device to compute on. Defaults to
device.\n",
"\n",
" Returns:\n",
" (dict): Results of model making predictions on data_loader.\n",
" \"\"\"\n",
" loss, acc = 0, 0\n",
" model.eval()\n",
" with torch.inference_mode():\n",
" for X, y in data_loader:\n",
" # Send data to the target device\n",
" X, y = X.to(device), y.to(device)\n",
 " y_pred = model(X)\n",
" loss += loss_fn(y_pred, y)\n",
" acc += accuracy_fn(y_true=y, y_pred=y_pred.argmax(dim=1))\n",
" \n",
" # Scale loss and acc\n",
" loss /= len(data_loader)\n",
" acc /= len(data_loader)\n",
" return {\"model_name\": model.__class__.__name__, # only works when model
was created with a class\n",
" \"model_loss\": loss.item(),\n",
" \"model_acc\": acc}\n",
"\n",
"# Calculate model 1 results with device-agnostic code \n",
"model_1_results = eval_model(model=model_1, data_loader=test_dataloader,\n",
" loss_fn=loss_fn, accuracy_fn=accuracy_fn,\n",
" device=device\n",
")\n",
"model_1_results"
]
},
{
"cell_type": "code",
"execution_count": 29,
"id": "a9e916cf-f873-4481-a983-bac26ce4cac2",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "a9e916cf-f873-4481-a983-bac26ce4cac2",
"outputId": "5cdb9f7f-366c-4c14-9afa-f2d1d4e0267d"
},
"outputs": [
{
"data": {
"text/plain": [
"{'model_name': 'FashionMNISTModelV0',\n",
" 'model_loss': 0.47663894295692444,\n",
" 'model_acc': 83.42651757188499}"
]
},
"execution_count": 29,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Check baseline results\n",
"model_0_results"
]
},
{
"cell_type": "markdown",
"id": "340cbf14-e83f-4981-8a93-5fedb6b51418",
"metadata": {
"id": "340cbf14-e83f-4981-8a93-5fedb6b51418"
},
"source": [
"Woah, in this case, it looks like adding non-linearities to our model made it
perform worse than the baseline.\n",
 "\n",
"That's a thing to note in machine learning, sometimes the thing you thought
should work doesn't. \n",
"\n",
"And then the thing you thought might not work does.\n",
"\n",
"It's part science, part art.\n",
"\n",
"From the looks of things, it seems like our model is **overfitting** on the
training data.\n",
"\n",
"Overfitting means our model is learning the training data well but those
patterns aren't generalizing to the testing data.\n",
"\n",
"Two of the main ways to fix overfitting include:\n",
"1. Using a smaller or different model (some models fit certain kinds of data
better than others).\n",
"2. Using a larger dataset (the more data, the more chance a model has to learn
generalizable patterns).\n",
"\n",
"There are more, but I'm going to leave that as a challenge for you to
explore.\n",
"\n",
"Try searching online, \"ways to prevent overfitting in machine learning\" and
see what comes up.\n",
"\n",
"In the meantime, let's take a look at number 1: using a different model."
]
},
{
"cell_type": "markdown",
"id": "ac22d685-1b8d-4215-90de-c0476cb0fbdf",
"metadata": {
"id": "ac22d685-1b8d-4215-90de-c0476cb0fbdf"
},
"source": [
"## 7. Model 2: Building a Convolutional Neural Network (CNN)\n",
"\n",
"Alright, time to step things up a notch.\n",
"\n",
"It's time to create a [Convolutional Neural
Network](https://en.wikipedia.org/wiki/Convolutional_neural_network) (CNN or
ConvNet).\n",
"\n",
"CNN's are known for their capabilities to find patterns in visual data.\n",
"\n",
"And since we're dealing with visual data, let's see if using a CNN model can
improve upon our baseline.\n",
"\n",
"The CNN model we're going to be using is known as TinyVGG from the [CNN
Explainer](https://poloclub.github.io/cnn-explainer/) website.\n",
"\n",
"It follows the typical structure of a convolutional neural network:\n",
"\n",
"`Input layer -> [Convolutional layer -> activation layer -> pooling layer] ->
Output layer`\n",
"\n",
"Where the contents of `[Convolutional layer -> activation layer -> pooling
layer]` can be upscaled and repeated multiple times, depending on requirements. "
 ]
},
{
"cell_type": "markdown",
"id": "9c358955-1d20-4903-b872-a239d2753d88",
"metadata": {
"id": "9c358955-1d20-4903-b872-a239d2753d88"
},
"source": [
"### What model should I use?\n",
"\n",
"> **Question:** Wait, you say CNN's are good for images, are there any other
model types I should be aware of?\n",
"\n",
"Good question.\n",
"\n",
"This table is a good general guide for which model to use (though there are
exceptions).\n",
"\n",
"| **Problem type** | **Model to use (generally)** | **Code example** |\n",
"| ----- | ----- | ----- |\n",
"| Structured data (Excel spreadsheets, row and column data) | Gradient boosted
models, Random Forests, XGBoost |
[`sklearn.ensemble`](https://scikit-learn.org/stable/modules/classes.html#module-
sklearn.ensemble), [XGBoost library](https://xgboost.readthedocs.io/en/stable/) |\
n",
"| Unstructured data (images, audio, language) | Convolutional Neural Networks,
Transformers |
[`torchvision.models`](https://pytorch.org/vision/stable/models.html), [HuggingFace
Transformers](https://huggingface.co/docs/transformers/index) | \n",
"\n",
"> **Note:** The table above is only for reference, the model you end up using
will be highly dependent on the problem you're working on and the constraints you
have (amount of data, latency requirements).\n",
"\n",
"Enough talking about models, let's now build a CNN that replicates the model
on the [CNN Explainer website](https://poloclub.github.io/cnn-explainer/).\n",
"\n",
"![TinyVGG architecture, as setup by CNN explainer
website](https://raw.githubusercontent.com/mrdbourke/pytorch-deep-learning/main/
images/03-cnn-explainer-model.png)\n",
"\n",
"To do so, we'll leverage the
[`nn.Conv2d()`](https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html) and
[`nn.MaxPool2d()`](https://pytorch.org/docs/stable/generated/torch.nn.MaxPool2d.htm
l) layers from `torch.nn`.\n"
]
},
{
"cell_type": "code",
"execution_count": 30,
"id": "dce60214-63fd-46e2-89ba-125445ac76b7",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "dce60214-63fd-46e2-89ba-125445ac76b7",
"outputId": "5ae97191-bb41-4e58-e7f1-914b612cbb60"
},
 "outputs": [
{
"data": {
"text/plain": [
"FashionMNISTModelV2(\n",
" (block_1): Sequential(\n",
" (0): Conv2d(1, 10, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\
n",
" (1): ReLU()\n",
" (2): Conv2d(10, 10, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\
n",
" (3): ReLU()\n",
" (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1,
ceil_mode=False)\n",
" )\n",
" (block_2): Sequential(\n",
" (0): Conv2d(10, 10, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\
n",
" (1): ReLU()\n",
" (2): Conv2d(10, 10, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\
n",
" (3): ReLU()\n",
" (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1,
ceil_mode=False)\n",
" )\n",
" (classifier): Sequential(\n",
" (0): Flatten(start_dim=1, end_dim=-1)\n",
" (1): Linear(in_features=490, out_features=10, bias=True)\n",
" )\n",
")"
]
},
"execution_count": 30,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Create a convolutional neural network \n",
"class FashionMNISTModelV2(nn.Module):\n",
" \"\"\"\n",
" Model architecture copying TinyVGG from: \n",
" https://poloclub.github.io/cnn-explainer/\n",
" \"\"\"\n",
" def __init__(self, input_shape: int, hidden_units: int, output_shape:
int):\n",
" super().__init__()\n",
" self.block_1 = nn.Sequential(\n",
" nn.Conv2d(in_channels=input_shape, \n",
" out_channels=hidden_units, \n",
" kernel_size=3, # how big is the square that's going over
the image?\n",
" stride=1, # default\n",
" padding=1),# options = \"valid\" (no padding)
or \"same\" (output has same shape as input) or int for specific number \n",
" nn.ReLU(),\n",
" nn.Conv2d(in_channels=hidden_units, \n",
" out_channels=hidden_units,\n",
" kernel_size=3,\n",
 " stride=1,\n",
" padding=1),\n",
" nn.ReLU(),\n",
" nn.MaxPool2d(kernel_size=2,\n",
" stride=2) # default stride value is same as
kernel_size\n",
" )\n",
" self.block_2 = nn.Sequential(\n",
" nn.Conv2d(hidden_units, hidden_units, 3, padding=1),\n",
" nn.ReLU(),\n",
" nn.Conv2d(hidden_units, hidden_units, 3, padding=1),\n",
" nn.ReLU(),\n",
" nn.MaxPool2d(2)\n",
" )\n",
" self.classifier = nn.Sequential(\n",
" nn.Flatten(),\n",
" # Where did this in_features shape come from? \n",
" # It's because each layer of our network compresses and changes
the shape of our input data.\n",
" nn.Linear(in_features=hidden_units*7*7, \n",
" out_features=output_shape)\n",
" )\n",
" \n",
" def forward(self, x: torch.Tensor):\n",
" x = self.block_1(x)\n",
" # print(x.shape)\n",
" x = self.block_2(x)\n",
" # print(x.shape)\n",
" x = self.classifier(x)\n",
" # print(x.shape)\n",
" return x\n",
"\n",
"torch.manual_seed(42)\n",
"model_2 = FashionMNISTModelV2(input_shape=1, \n",
" hidden_units=10, \n",
" output_shape=len(class_names)).to(device)\n",
"model_2"
]
},
{
"cell_type": "markdown",
"id": "0a20f25e-cc16-4f85-a69b-62008c01d0ed",
"metadata": {
"id": "0a20f25e-cc16-4f85-a69b-62008c01d0ed"
},
"source": [
"Nice!\n",
"\n",
"Our biggest model yet!\n",
"\n",
"What we've done is a common practice in machine learning.\n",
"\n",
"Find a model architecture somewhere and replicate it with code. "
]
},
{
"cell_type": "markdown",
"id": "6478cc5a-7b33-425d-9ab3-6d40168a1aee",
"metadata": {
 "id": "6478cc5a-7b33-425d-9ab3-6d40168a1aee"
},
"source": [
"### 7.1 Stepping through `nn.Conv2d()`\n",
"\n",
"We could start using our model above and see what happens but let's first step
through the two new layers we've added:\n",
"*
[`nn.Conv2d()`](https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html),
also known as a convolutional layer.\n",
"*
[`nn.MaxPool2d()`](https://pytorch.org/docs/stable/generated/torch.nn.MaxPool2d.htm
l), also known as a max pooling layer.\n",
"\n",
"> **Question:** What does the \"2d\" in `nn.Conv2d()` stand for?\n",
">\n",
"> The 2d is for 2-dimensional data. As in, our images have two dimensions:
height and width. Yes, there's color channel dimension but each of the color
channel dimensions have two dimensions too: height and width.\n",
">\n",
"> For other dimensional data (such as 1D for text or 3D for 3D objects)
there's also `nn.Conv1d()` and `nn.Conv3d()`. \n",
"\n",
"To test the layers out, let's create some toy data just like the data used on
CNN Explainer."
]
},
{
"cell_type": "code",
"execution_count": 31,
"id": "058b01ac-3f6a-4472-bcbf-3377974e3254",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "058b01ac-3f6a-4472-bcbf-3377974e3254",
"outputId": "c404a8dd-d804-4993-bc2b-e4fdbf02b62d"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Image batch shape: torch.Size([32, 3, 64, 64]) -> [batch_size,
color_channels, height, width]\n",
"Single image shape: torch.Size([3, 64, 64]) -> [color_channels, height,
width]\n",
"Single image pixel values:\n",
"tensor([[[ 1.9269, 1.4873, 0.9007, ..., 1.8446, -1.1845, 1.3835],\n",
" [ 1.4451, 0.8564, 2.2181, ..., 0.3399, 0.7200, 0.4114],\n",
" [ 1.9312, 1.0119, -1.4364, ..., -0.5558, 0.7043, 0.7099],\n",
" ...,\n",
" [-0.5610, -0.4830, 0.4770, ..., -0.2713, -0.9537, -0.6737],\n",
" [ 0.3076, -0.1277, 0.0366, ..., -2.0060, 0.2824, -0.8111],\n",
" [-1.5486, 0.0485, -0.7712, ..., -0.1403, 0.9416, -0.0118]],\n",
"\n",
" [[-0.5197, 1.8524, 1.8365, ..., 0.8935, -1.5114, -0.8515],\n",
" [ 2.0818, 1.0677, -1.4277, ..., 1.6612, -2.6223, -0.4319],\n",
" [-0.1010, -0.4388, -1.9775, ..., 0.2106, 0.2536, -0.7318],\n",
 " ...,\n",
" [ 0.2779, 0.7342, -0.3736, ..., -0.4601, 0.1815, 0.1850],\n",
" [ 0.7205, -0.2833, 0.0937, ..., -0.1002, -2.3609, 2.2465],\n",
" [-1.3242, -0.1973, 0.2920, ..., 0.5409, 0.6940, 1.8563]],\n",
"\n",
" [[-0.7978, 1.0261, 1.1465, ..., 1.2134, 0.9354, -0.0780],\n",
" [-1.4647, -1.9571, 0.1017, ..., -1.9986, -0.7409, 0.7011],\n",
" [-1.3938, 0.8466, -1.7191, ..., -1.1867, 0.1320, 0.3407],\n",
" ...,\n",
" [ 0.8206, -0.3745, 1.2499, ..., -0.0676, 0.0385, 0.6335],\n",
" [-0.5589, -0.3393, 0.2347, ..., 2.1181, 2.4569, 1.3083],\n",
" [-0.4092, 1.5199, 0.2401, ..., -0.2558, 0.7870, 0.9924]]])\n"
]
}
],
"source": [
"torch.manual_seed(42)\n",
"\n",
"# Create sample batch of random numbers with same size as image batch\n",
"images = torch.randn(size=(32, 3, 64, 64)) # [batch_size, color_channels,
height, width]\n",
"test_image = images[0] # get a single image for testing\n",
"print(f\"Image batch shape: {images.shape} -> [batch_size, color_channels,
height, width]\")\n",
"print(f\"Single image shape: {test_image.shape} -> [color_channels, height,
width]\") \n",
"print(f\"Single image pixel values:\\n{test_image}\")"
]
},
{
"cell_type": "markdown",
"id": "bd3291c2-854e-4d0c-97b9-8bf46085fc43",
"metadata": {
"id": "bd3291c2-854e-4d0c-97b9-8bf46085fc43"
},
"source": [
"Let's create an example `nn.Conv2d()` with various parameters:\n",
"* `in_channels` (int) - Number of channels in the input image.\n",
"* `out_channels` (int) - Number of channels produced by the convolution.\n",
"* `kernel_size` (int or tuple) - Size of the convolving kernel/filter.\n",
"* `stride` (int or tuple, optional) - How big of a step the convolving kernel
takes at a time. Default: 1.\n",
"* `padding` (int, tuple, str) - Padding added to all four sides of input.
Default: 0.\n",
"\n",
"![example of going through the different parameters of a Conv2d layer]
(https://raw.githubusercontent.com/mrdbourke/pytorch-deep-learning/main/images/03-
conv2d-layer.gif)\n",
"\n",
"*Example of what happens when you change the hyperparameters of a
`nn.Conv2d()` layer.*"
]
},
{
"cell_type": "code",
"execution_count": 32,
"id": "ebd39562-1dad-40e3-90f5-750a5dac24e2",
"metadata": {
"colab": {
 "base_uri": "https://localhost:8080/"
},
"id": "ebd39562-1dad-40e3-90f5-750a5dac24e2",
"outputId": "b61154fb-c5f7-4c3f-c619-4bde6edb4d16"
},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[[ 1.5396, 0.0516, 0.6454, ..., -0.3673, 0.8711, 0.4256],\n",
" [ 0.3662, 1.0114, -0.5997, ..., 0.8983, 0.2809, -0.2741],\n",
" [ 1.2664, -1.4054, 0.3727, ..., -0.3409, 1.2191, -0.0463],\n",
" ...,\n",
" [-0.1541, 0.5132, -0.3624, ..., -0.2360, -0.4609, -0.0035],\n",
" [ 0.2981, -0.2432, 1.5012, ..., -0.6289, -0.7283, -0.5767],\n",
" [-0.0386, -0.0781, -0.0388, ..., 0.2842, 0.4228, -0.1802]],\n",
"\n",
" [[-0.2840, -0.0319, -0.4455, ..., -0.7956, 1.5599, -1.2449],\n",
" [ 0.2753, -0.1262, -0.6541, ..., -0.2211, 0.1999, -0.8856],\n",
" [-0.5404, -1.5489, 0.0249, ..., -0.5932, -1.0913, -0.3849],\n",
" ...,\n",
" [ 0.3870, -0.4064, -0.8236, ..., 0.1734, -0.4330, -0.4951],\n",
" [-0.1984, -0.6386, 1.0263, ..., -0.9401, -0.0585, -0.7833],\n",
" [-0.6306, -0.2052, -0.3694, ..., -1.3248, 0.2456, -0.7134]],\n",
"\n",
" [[ 0.4414, 0.5100, 0.4846, ..., -0.8484, 0.2638, 1.1258],\n",
" [ 0.8117, 0.3191, -0.0157, ..., 1.2686, 0.2319, 0.5003],\n",
" [ 0.3212, 0.0485, -0.2581, ..., 0.2258, 0.2587, -0.8804],\n",
" ...,\n",
" [-0.1144, -0.1869, 0.0160, ..., -0.8346, 0.0974, 0.8421],\n",
" [ 0.2941, 0.4417, 0.5866, ..., -0.1224, 0.4814, -0.4799],\n",
" [ 0.6059, -0.0415, -0.2028, ..., 0.1170, 0.2521, -0.4372]],\n",
"\n",
" ...,\n",
"\n",
" [[-0.2560, -0.0477, 0.6380, ..., 0.6436, 0.7553, -0.7055],\n",
" [ 1.5595, -0.2209, -0.9486, ..., -0.4876, 0.7754, 0.0750],\n",
" [-0.0797, 0.2471, 1.1300, ..., 0.1505, 0.2354, 0.9576],\n",
" ...,\n",
" [ 1.1065, 0.6839, 1.2183, ..., 0.3015, -0.1910, -0.1902],\n",
" [-0.3486, -0.7173, -0.3582, ..., 0.4917, 0.7219, 0.1513],\n",
" [ 0.0119, 0.1017, 0.7839, ..., -0.3752, -0.8127, -0.1257]],\n",
"\n",
" [[ 0.3841, 1.1322, 0.1620, ..., 0.7010, 0.0109, 0.6058],\n",
" [ 0.1664, 0.1873, 1.5924, ..., 0.3733, 0.9096, -0.5399],\n",
" [ 0.4094, -0.0861, -0.7935, ..., -0.1285, -0.9932, -0.3013],\n",
" ...,\n",
" [ 0.2688, -0.5630, -1.1902, ..., 0.4493, 0.5404, -0.0103],\n",
" [ 0.0535, 0.4411, 0.5313, ..., 0.0148, -1.0056, 0.3759],\n",
" [ 0.3031, -0.1590, -0.1316, ..., -0.5384, -0.4271, -0.4876]],\n",
"\n",
" [[-1.1865, -0.7280, -1.2331, ..., -0.9013, -0.0542, -1.5949],\n",
" [-0.6345, -0.5920, 0.5326, ..., -1.0395, -0.7963, -0.0647],\n",
" [-0.1132, 0.5166, 0.2569, ..., 0.5595, -1.6881, 0.9485],\n",
" ...,\n",
" [-0.0254, -0.2669, 0.1927, ..., -0.2917, 0.1088, -0.4807],\n",
" [-0.2609, -0.2328, 0.1404, ..., -0.1325, -0.8436, -0.7524],\n",
" [-1.1399, -0.1751, -0.8705, ..., 0.1589, 0.3377, 0.3493]]],\
n",
 " grad_fn=<SqueezeBackward1>)"
]
},
"execution_count": 32,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"torch.manual_seed(42)\n",
"\n",
"# Create a convolutional layer with same dimensions as TinyVGG \n",
"# (try changing any of the parameters and see what happens)\n",
"conv_layer = nn.Conv2d(in_channels=3,\n",
" out_channels=10,\n",
" kernel_size=3,\n",
" stride=1,\n",
" padding=0) # also try using \"valid\" or \"same\"
here \n",
"\n",
"# Pass the data through the convolutional layer\n",
"conv_layer(test_image) # Note: If running PyTorch <1.11.0, this will error
because of shape issues (nn.Conv.2d() expects a 4d tensor as input) "
]
},
{
"cell_type": "markdown",
"id": "cb0184ad-5c16-4e1c-bcfa-70ecf15377da",
"metadata": {
"id": "cb0184ad-5c16-4e1c-bcfa-70ecf15377da"
},
"source": [
"If we try to pass a single image in, we get a shape mismatch error:\n",
"\n",
"> `RuntimeError: Expected 4-dimensional input for 4-dimensional weight [10, 3,
3, 3], but got 3-dimensional input of size [3, 64, 64] instead`\n",
">\n",
"> **Note:** If you're running PyTorch 1.11.0+, this error won't occur.\n",
"\n",
"This is because our `nn.Conv2d()` layer expects a 4-dimensional tensor as
input with size `(N, C, H, W)` or `[batch_size, color_channels, height, width]`.\
n",
"\n",
"Right now our single image `test_image` only has a shape of `[color_channels,
height, width]` or `[3, 64, 64]`.\n",
"\n",
"We can fix this for a single image using `test_image.unsqueeze(dim=0)` to add
an extra dimension for `N`."
]
},
{
"cell_type": "code",
"execution_count": 33,
"id": "abba741d-a1ed-44ed-ba53-41d589433a2c",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "abba741d-a1ed-44ed-ba53-41d589433a2c",
 "outputId": "9dd8151d-376c-4342-c379-91fcb6468706"
},
"outputs": [
{
"data": {
"text/plain": [
"torch.Size([1, 3, 64, 64])"
]
},
"execution_count": 33,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Add extra dimension to test image\n",
"test_image.unsqueeze(dim=0).shape"
]
},
{
"cell_type": "code",
"execution_count": 34,
"id": "c7280a49-4ee0-452b-a514-61115b6a444c",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "c7280a49-4ee0-452b-a514-61115b6a444c",
"outputId": "87bf7e37-c1a7-44a4-eef2-a02eb0489147"
},
"outputs": [
{
"data": {
"text/plain": [
"torch.Size([1, 10, 62, 62])"
]
},
"execution_count": 34,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Pass test image with extra dimension through conv_layer\n",
"conv_layer(test_image.unsqueeze(dim=0)).shape"
]
},
{
"cell_type": "markdown",
"id": "181df81b-7c5a-46cc-b8d5-a592bf755a13",
"metadata": {
"id": "181df81b-7c5a-46cc-b8d5-a592bf755a13"
},
"source": [
"Hmm, notice what happens to our shape (the same shape as the first layer of
TinyVGG on [CNN Explainer](https://poloclub.github.io/cnn-explainer/)), we get
different channel sizes as well as different pixel sizes.\n",
"\n",
"What if we changed the values of `conv_layer`?"
 ]
},
{
"cell_type": "code",
"execution_count": 35,
"id": "04445d45-cf2f-4c1d-b215-bc50865a207a",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "04445d45-cf2f-4c1d-b215-bc50865a207a",
"outputId": "eaa97fb8-52c0-493d-eac3-f2e23df9b01c"
},
"outputs": [
{
"data": {
"text/plain": [
"torch.Size([1, 10, 30, 30])"
]
},
"execution_count": 35,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"torch.manual_seed(42)\n",
"# Create a new conv_layer with different values (try setting these to whatever
you like)\n",
"conv_layer_2 = nn.Conv2d(in_channels=3, # same number of color channels as our
input image\n",
" out_channels=10,\n",
" kernel_size=(5, 5), # kernel is usually a square so a
tuple also works\n",
" stride=2,\n",
" padding=0)\n",
"\n",
"# Pass single image through new conv_layer_2 (this calls nn.Conv2d()'s
forward() method on the input)\n",
"conv_layer_2(test_image.unsqueeze(dim=0)).shape"
]
},
{
"cell_type": "markdown",
"id": "b27dbdbb-3e32-4ffa-803e-cf943d96c72b",
"metadata": {
"id": "b27dbdbb-3e32-4ffa-803e-cf943d96c72b"
},
"source": [
"Woah, we get another shape change.\n",
"\n",
"Now our image is of shape `[1, 10, 30, 30]` (it will be different if you use
different values) or `[batch_size=1, color_channels=10, height=30, width=30]`.\n",
"\n",
"What's going on here?\n",
"\n",
"Behind the scenes, our `nn.Conv2d()` is compressing the information stored in
the image.\n",
"\n",
 "It does this by performing operations on the input (our test image) against
its internal parameters.\n",
"\n",
"The goal of this is similar to all of the other neural networks we've been
building.\n",
"\n",
"Data goes in and the layers try to update their internal parameters (patterns)
to lower the loss function thanks to some help of the optimizer.\n",
"\n",
"The only difference is *how* the different layers calculate their parameter
updates or in PyTorch terms, the operation present in the layer `forward()`
method.\n",
"\n",
"If we check out our `conv_layer_2.state_dict()` we'll find a similar weight
and bias setup as we've seen before."
]
},
{
"cell_type": "code",
"execution_count": 36,
"id": "46027ed1-c3a7-46bd-bab7-17f8c20e354b",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "46027ed1-c3a7-46bd-bab7-17f8c20e354b",
"outputId": "bc493b18-1ef3-41c6-9de9-5dfeaa4c259e"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"OrderedDict([('weight', tensor([[[[ 0.0883, 0.0958, -0.0271, 0.1061, -
0.0253],\n",
" [ 0.0233, -0.0562, 0.0678, 0.1018, -0.0847],\n",
" [ 0.1004, 0.0216, 0.0853, 0.0156, 0.0557],\n",
" [-0.0163, 0.0890, 0.0171, -0.0539, 0.0294],\n",
" [-0.0532, -0.0135, -0.0469, 0.0766, -0.0911]],\n",
"\n",
" [[-0.0532, -0.0326, -0.0694, 0.0109, -0.1140],\n",
" [ 0.1043, -0.0981, 0.0891, 0.0192, -0.0375],\n",
" [ 0.0714, 0.0180, 0.0933, 0.0126, -0.0364],\n",
" [ 0.0310, -0.0313, 0.0486, 0.1031, 0.0667],\n",
" [-0.0505, 0.0667, 0.0207, 0.0586, -0.0704]],\n",
"\n",
" [[-0.1143, -0.0446, -0.0886, 0.0947, 0.0333],\n",
" [ 0.0478, 0.0365, -0.0020, 0.0904, -0.0820],\n",
" [ 0.0073, -0.0788, 0.0356, -0.0398, 0.0354],\n",
" [-0.0241, 0.0958, -0.0684, -0.0689, -0.0689],\n",
" [ 0.1039, 0.0385, 0.1111, -0.0953, -0.1145]]],\n",
"\n",
"\n",
" [[[-0.0903, -0.0777, 0.0468, 0.0413, 0.0959],\n",
" [-0.0596, -0.0787, 0.0613, -0.0467, 0.0701],\n",
" [-0.0274, 0.0661, -0.0897, -0.0583, 0.0352],\n",
" [ 0.0244, -0.0294, 0.0688, 0.0785, -0.0837],\n",
" [-0.0616, 0.1057, -0.0390, -0.0409, -0.1117]],\n",
"\n",
" [[-0.0661, 0.0288, -0.0152, -0.0838, 0.0027],\n",
" [-0.0789, -0.0980, -0.0636, -0.1011, -0.0735],\n",
" [ 0.1154, 0.0218, 0.0356, -0.1077, -0.0758],\n",
" [-0.0384, 0.0181, -0.1016, -0.0498, -0.0691],\n",
" [ 0.0003, -0.0430, -0.0080, -0.0782, -0.0793]],\n",
"\n",
" [[-0.0674, -0.0395, -0.0911, 0.0968, -0.0229],\n",
" [ 0.0994, 0.0360, -0.0978, 0.0799, -0.0318],\n",
" [-0.0443, -0.0958, -0.1148, 0.0330, -0.0252],\n",
" [ 0.0450, -0.0948, 0.0857, -0.0848, -0.0199],\n",
" [ 0.0241, 0.0596, 0.0932, 0.1052, -0.0916]]],\n",
"\n",
"\n",
" [[[ 0.0291, -0.0497, -0.0127, -0.0864, 0.1052],\n",
" [-0.0847, 0.0617, 0.0406, 0.0375, -0.0624],\n",
" [ 0.1050, 0.0254, 0.0149, -0.1018, 0.0485],\n",
" [-0.0173, -0.0529, 0.0992, 0.0257, -0.0639],\n",
" [-0.0584, -0.0055, 0.0645, -0.0295, -0.0659]],\n",
"\n",
" [[-0.0395, -0.0863, 0.0412, 0.0894, -0.1087],\n",
" [ 0.0268, 0.0597, 0.0209, -0.0411, 0.0603],\n",
" [ 0.0607, 0.0432, -0.0203, -0.0306, 0.0124],\n",
" [-0.0204, -0.0344, 0.0738, 0.0992, -0.0114],\n",
" [-0.0259, 0.0017, -0.0069, 0.0278, 0.0324]],\n",
"\n",
" [[-0.1049, -0.0426, 0.0972, 0.0450, -0.0057],\n",
" [-0.0696, -0.0706, -0.1034, -0.0376, 0.0390],\n",
" [ 0.0736, 0.0533, -0.1021, -0.0694, -0.0182],\n",
" [ 0.1117, 0.0167, -0.0299, 0.0478, -0.0440],\n",
" [-0.0747, 0.0843, -0.0525, -0.0231, -0.1149]]],\n",
"\n",
"\n",
" [[[ 0.0773, 0.0875, 0.0421, -0.0805, -0.1140],\n",
" [-0.0938, 0.0861, 0.0554, 0.0972, 0.0605],\n",
" [ 0.0292, -0.0011, -0.0878, -0.0989, -0.1080],\n",
" [ 0.0473, -0.0567, -0.0232, -0.0665, -0.0210],\n",
" [-0.0813, -0.0754, 0.0383, -0.0343, 0.0713]],\n",
"\n",
" [[-0.0370, -0.0847, -0.0204, -0.0560, -0.0353],\n",
" [-0.1099, 0.0646, -0.0804, 0.0580, 0.0524],\n",
" [ 0.0825, -0.0886, 0.0830, -0.0546, 0.0428],\n",
" [ 0.1084, -0.0163, -0.0009, -0.0266, -0.0964],\n",
" [ 0.0554, -0.1146, 0.0717, 0.0864, 0.1092]],\n",
"\n",
" [[-0.0272, -0.0949, 0.0260, 0.0638, -0.1149],\n",
" [-0.0262, -0.0692, -0.0101, -0.0568, -0.0472],\n",
" [-0.0367, -0.1097, 0.0947, 0.0968, -0.0181],\n",
" [-0.0131, -0.0471, -0.1043, -0.1124, 0.0429],\n",
" [-0.0634, -0.0742, -0.0090, -0.0385, -0.0374]]],\n",
"\n",
"\n",
" [[[ 0.0037, -0.0245, -0.0398, -0.0553, -0.0940],\n",
" [ 0.0968, -0.0462, 0.0306, -0.0401, 0.0094],\n",
" [ 0.1077, 0.0532, -0.1001, 0.0458, 0.1096],\n",
" [ 0.0304, 0.0774, 0.1138, -0.0177, 0.0240],\n",
" [-0.0803, -0.0238, 0.0855, 0.0592, -0.0731]],\n",
"\n",
" [[-0.0926, -0.0789, -0.1140, -0.0891, -0.0286],\n",
" [ 0.0779, 0.0193, -0.0878, -0.0926, 0.0574],\n",
" [-0.0859, -0.0142, 0.0554, -0.0534, -0.0126],\n",
" [-0.0101, -0.0273, -0.0585, -0.1029, -0.0933],\n",
" [-0.0618, 0.1115, -0.0558, -0.0775, 0.0280]],\n",
"\n",
" [[ 0.0318, 0.0633, 0.0878, 0.0643, -0.1145],\n",
" [ 0.0102, 0.0699, -0.0107, -0.0680, 0.1101],\n",
" [-0.0432, -0.0657, -0.1041, 0.0052, 0.0512],\n",
" [ 0.0256, 0.0228, -0.0876, -0.1078, 0.0020],\n",
" [ 0.1053, 0.0666, -0.0672, -0.0150, -0.0851]]],\n",
"\n",
"\n",
" [[[-0.0557, 0.0209, 0.0629, 0.0957, -0.1060],\n",
" [ 0.0772, -0.0814, 0.0432, 0.0977, 0.0016],\n",
" [ 0.1051, -0.0984, -0.0441, 0.0673, -0.0252],\n",
" [-0.0236, -0.0481, 0.0796, 0.0566, 0.0370],\n",
" [-0.0649, -0.0937, 0.0125, 0.0342, -0.0533]],\n",
"\n",
" [[-0.0323, 0.0780, 0.0092, 0.0052, -0.0284],\n",
" [-0.1046, -0.1086, -0.0552, -0.0587, 0.0360],\n",
" [-0.0336, -0.0452, 0.1101, 0.0402, 0.0823],\n",
" [-0.0559, -0.0472, 0.0424, -0.0769, -0.0755],\n",
" [-0.0056, -0.0422, -0.0866, 0.0685, 0.0929]],\n",
"\n",
" [[ 0.0187, -0.0201, -0.1070, -0.0421, 0.0294],\n",
" [ 0.0544, -0.0146, -0.0457, 0.0643, -0.0920],\n",
" [ 0.0730, -0.0448, 0.0018, -0.0228, 0.0140],\n",
" [-0.0349, 0.0840, -0.0030, 0.0901, 0.1110],\n",
" [-0.0563, -0.0842, 0.0926, 0.0905, -0.0882]]],\n",
"\n",
"\n",
" [[[-0.0089, -0.1139, -0.0945, 0.0223, 0.0307],\n",
" [ 0.0245, -0.0314, 0.1065, 0.0165, -0.0681],\n",
" [-0.0065, 0.0277, 0.0404, -0.0816, 0.0433],\n",
" [-0.0590, -0.0959, -0.0631, 0.1114, 0.0987],\n",
" [ 0.1034, 0.0678, 0.0872, -0.0155, -0.0635]],\n",
"\n",
" [[ 0.0577, -0.0598, -0.0779, -0.0369, 0.0242],\n",
" [ 0.0594, -0.0448, -0.0680, 0.0156, -0.0681],\n",
" [-0.0752, 0.0602, -0.0194, 0.1055, 0.1123],\n",
" [ 0.0345, 0.0397, 0.0266, 0.0018, -0.0084],\n",
" [ 0.0016, 0.0431, 0.1074, -0.0299, -0.0488]],\n",
"\n",
" [[-0.0280, -0.0558, 0.0196, 0.0862, 0.0903],\n",
" [ 0.0530, -0.0850, -0.0620, -0.0254, -0.0213],\n",
" [ 0.0095, -0.1060, 0.0359, -0.0881, -0.0731],\n",
" [-0.0960, 0.1006, -0.1093, 0.0871, -0.0039],\n",
" [-0.0134, 0.0722, -0.0107, 0.0724, 0.0835]]],\n",
"\n",
"\n",
" [[[-0.1003, 0.0444, 0.0218, 0.0248, 0.0169],\n",
" [ 0.0316, -0.0555, -0.0148, 0.1097, 0.0776],\n",
" [-0.0043, -0.1086, 0.0051, -0.0786, 0.0939],\n",
" [-0.0701, -0.0083, -0.0256, 0.0205, 0.1087],\n",
" [ 0.0110, 0.0669, 0.0896, 0.0932, -0.0399]],\n",
"\n",
" [[-0.0258, 0.0556, -0.0315, 0.0541, -0.0252],\n",
" [-0.0783, 0.0470, 0.0177, 0.0515, 0.1147],\n",
" [ 0.0788, 0.1095, 0.0062, -0.0993, -0.0810],\n",
" [-0.0717, -0.1018, -0.0579, -0.1063, -0.1065],\n",
 " [-0.0690, -0.1138, -0.0709, 0.0440, 0.0963]],\n",
"\n",
" [[-0.0343, -0.0336, 0.0617, -0.0570, -0.0546],\n",
" [ 0.0711, -0.1006, 0.0141, 0.1020, 0.0198],\n",
" [ 0.0314, -0.0672, -0.0016, 0.0063, 0.0283],\n",
" [ 0.0449, 0.1003, -0.0881, 0.0035, -0.0577],\n",
" [-0.0913, -0.0092, -0.1016, 0.0806, 0.0134]]],\n",
"\n",
"\n",
" [[[-0.0622, 0.0603, -0.1093, -0.0447, -0.0225],\n",
" [-0.0981, -0.0734, -0.0188, 0.0876, 0.1115],\n",
" [ 0.0735, -0.0689, -0.0755, 0.1008, 0.0408],\n",
" [ 0.0031, 0.0156, -0.0928, -0.0386, 0.1112],\n",
" [-0.0285, -0.0058, -0.0959, -0.0646, -0.0024]],\n",
"\n",
" [[-0.0717, -0.0143, 0.0470, -0.1130, 0.0343],\n",
" [-0.0763, -0.0564, 0.0443, 0.0918, -0.0316],\n",
" [-0.0474, -0.1044, -0.0595, -0.1011, -0.0264],\n",
" [ 0.0236, -0.1082, 0.1008, 0.0724, -0.1130],\n",
" [-0.0552, 0.0377, -0.0237, -0.0126, -0.0521]],\n",
"\n",
" [[ 0.0927, -0.0645, 0.0958, 0.0075, 0.0232],\n",
" [ 0.0901, -0.0190, -0.0657, -0.0187, 0.0937],\n",
" [-0.0857, 0.0262, -0.1135, 0.0605, 0.0427],\n",
" [ 0.0049, 0.0496, 0.0001, 0.0639, -0.0914],\n",
" [-0.0170, 0.0512, 0.1150, 0.0588, -0.0840]]],\n",
"\n",
"\n",
" [[[ 0.0888, -0.0257, -0.0247, -0.1050, -0.0182],\n",
" [ 0.0817, 0.0161, -0.0673, 0.0355, -0.0370],\n",
" [ 0.1054, -0.1002, -0.0365, -0.1115, -0.0455],\n",
" [ 0.0364, 0.1112, 0.0194, 0.1132, 0.0226],\n",
" [ 0.0667, 0.0926, 0.0965, -0.0646, 0.1062]],\n",
"\n",
" [[ 0.0699, -0.0540, -0.0551, -0.0969, 0.0290],\n",
" [-0.0936, 0.0488, 0.0365, -0.1003, 0.0315],\n",
" [-0.0094, 0.0527, 0.0663, -0.1148, 0.1059],\n",
" [ 0.0968, 0.0459, -0.1055, -0.0412, -0.0335],\n",
" [-0.0297, 0.0651, 0.0420, 0.0915, -0.0432]],\n",
"\n",
" [[ 0.0389, 0.0411, -0.0961, -0.1120, -0.0599],\n",
" [ 0.0790, -0.1087, -0.1005, 0.0647, 0.0623],\n",
" [ 0.0950, -0.0872, -0.0845, 0.0592, 0.1004],\n",
" [ 0.0691, 0.0181, 0.0381, 0.1096, -0.0745],\n",
" [-0.0524, 0.0808, -0.0790, -0.0637, 0.0843]]]])), ('bias',
tensor([ 0.0364, 0.0373, -0.0489, -0.0016, 0.1057, -0.0693, 0.0009, 0.0549,\n",
" -0.0797, 0.1121]))])\n"
]
}
],
"source": [
"# Check out the conv_layer_2 internal parameters\n",
"print(conv_layer_2.state_dict())"
]
},
{
"cell_type": "markdown",
"id": "8b708eb6-ae46-4d8c-a8a4-1392827d3e37",
"metadata": {
 "id": "8b708eb6-ae46-4d8c-a8a4-1392827d3e37"
},
"source": [
"Look at that! A bunch of random numbers for a weight and bias tensor.\n",
"\n",
"The shapes of these are manipulated by the inputs we passed to `nn.Conv2d()`
when we set it up.\n",
"\n",
"Let's check them out."
]
},
{
"cell_type": "code",
"execution_count": 37,
"id": "e5518d61-c0b7-4351-b5ea-4d6b6144291a",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "e5518d61-c0b7-4351-b5ea-4d6b6144291a",
"outputId": "14ef701e-c82d-4fae-9dd6-e18352817cf2"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"conv_layer_2 weight shape: \n",
"torch.Size([10, 3, 5, 5]) -> [out_channels=10, in_channels=3, kernel_size=5,
kernel_size=5]\n",
"\n",
"conv_layer_2 bias shape: \n",
"torch.Size([10]) -> [out_channels=10]\n"
]
}
],
"source": [
"# Get shapes of weight and bias tensors within conv_layer_2\n",
"print(f\"conv_layer_2 weight shape: \\n{conv_layer_2.weight.shape} ->
[out_channels=10, in_channels=3, kernel_size=5, kernel_size=5]\")\n",
"print(f\"\\nconv_layer_2 bias shape: \\n{conv_layer_2.bias.shape} ->
[out_channels=10]\")"
]
},
{
"cell_type": "markdown",
"id": "f0de23c7-4501-4156-80a4-ac889a636a42",
"metadata": {
"id": "f0de23c7-4501-4156-80a4-ac889a636a42"
},
"source": [
"> **Question:** What should we set the parameters of our `nn.Conv2d()`
layers?\n",
">\n",
"> That's a good one. But similar to many other things in machine learning, the
values of these aren't set in stone (and recall, because these values are ones we
can set ourselves, they're referred to as \"**hyperparameters**\"). \n",
">\n",
"> The best way to find out is to try out different values and see how they
effect your model's performance.\n",
">\n",
"> Or even better, find a working example on a problem similar to yours (like
we've done with TinyVGG) and copy it. \n",
"\n",
"We're working with a different of layer here to what we've seen before.\n",
"\n",
"But the premise remains the same: start with random numbers and update them to
better represent the data."
]
},
{
"cell_type": "markdown",
"id": "6370d45d-ca44-4fa0-a2d7-efaf0a207b91",
"metadata": {
"id": "6370d45d-ca44-4fa0-a2d7-efaf0a207b91"
},
"source": [
"### 7.2 Stepping through `nn.MaxPool2d()`\n",
"Now let's check out what happens when we move data through `nn.MaxPool2d()`."
]
},
{
"cell_type": "code",
"execution_count": 38,
"id": "1164c753-19d9-43b7-a04f-017d0f7188c3",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "1164c753-19d9-43b7-a04f-017d0f7188c3",
"outputId": "9c46f08e-928d-4ee4-e43c-b402113fc2b4"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Test image original shape: torch.Size([3, 64, 64])\n",
"Test image with unsqueezed dimension: torch.Size([1, 3, 64, 64])\n",
"Shape after going through conv_layer(): torch.Size([1, 10, 62, 62])\n",
"Shape after going through conv_layer() and max_pool_layer(): torch.Size([1,
10, 31, 31])\n"
]
}
],
"source": [
"# Print out original image shape without and with unsqueezed dimension\n",
"print(f\"Test image original shape: {test_image.shape}\")\n",
"print(f\"Test image with unsqueezed dimension:
{test_image.unsqueeze(dim=0).shape}\")\n",
"\n",
"# Create a sample nn.MaxPoo2d() layer\n",
"max_pool_layer = nn.MaxPool2d(kernel_size=2)\n",
"\n",
"# Pass data through just the conv_layer\n",
"test_image_through_conv = conv_layer(test_image.unsqueeze(dim=0))\n",
"print(f\"Shape after going through conv_layer():
{test_image_through_conv.shape}\")\n",
 "\n",
"# Pass data through the max pool layer\n",
"test_image_through_conv_and_max_pool =
max_pool_layer(test_image_through_conv)\n",
"print(f\"Shape after going through conv_layer() and max_pool_layer():
{test_image_through_conv_and_max_pool.shape}\")"
]
},
{
"cell_type": "markdown",
"id": "de029abd-6674-4bfa-99ab-322f339f89f4",
"metadata": {
"id": "de029abd-6674-4bfa-99ab-322f339f89f4"
},
"source": [
"Notice the change in the shapes of what's happening in and out of a
`nn.MaxPool2d()` layer.\n",
"\n",
"The `kernel_size` of the `nn.MaxPool2d()` layer will affect the size of the
output shape.\n",
"\n",
"In our case, the shape halves from a `62x62` image to `31x31` image.\n",
"\n",
"Let's see this work with a smaller tensor."
]
},
{
"cell_type": "code",
"execution_count": 39,
"id": "e6a2b196-4845-4b40-9212-e75406e88875",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "e6a2b196-4845-4b40-9212-e75406e88875",
"outputId": "5a5e5df1-8e25-4061-d223-398ecfc7ef4c"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Random tensor:\n",
"tensor([[[[0.3367, 0.1288],\n",
" [0.2345, 0.2303]]]])\n",
"Random tensor shape: torch.Size([1, 1, 2, 2])\n",
"\n",
"Max pool tensor:\n",
"tensor([[[[0.3367]]]]) <- this is the maximum value from random_tensor\n",
"Max pool tensor shape: torch.Size([1, 1, 1, 1])\n"
]
}
],
"source": [
"torch.manual_seed(42)\n",
"# Create a random tensor with a similar number of dimensions to our images\n",
"random_tensor = torch.randn(size=(1, 1, 2, 2))\n",
"print(f\"Random tensor:\\n{random_tensor}\")\n",
"print(f\"Random tensor shape: {random_tensor.shape}\")\n",
 "\n",
"# Create a max pool layer\n",
"max_pool_layer = nn.MaxPool2d(kernel_size=2) # see what happens when you
change the kernel_size value \n",
"\n",
"# Pass the random tensor through the max pool layer\n",
"max_pool_tensor = max_pool_layer(random_tensor)\n",
"print(f\"\\nMax pool tensor:\\n{max_pool_tensor} <- this is the maximum value
from random_tensor\")\n",
"print(f\"Max pool tensor shape: {max_pool_tensor.shape}\")"
]
},
{
"cell_type": "markdown",
"id": "002e586e-dcb3-40fe-a7dd-a1c18a3b8da0",
"metadata": {
"id": "002e586e-dcb3-40fe-a7dd-a1c18a3b8da0"
},
"source": [
"Notice the final two dimensions between `random_tensor` and `max_pool_tensor`,
they go from `[2, 2]` to `[1, 1]`.\n",
"\n",
"In essence, they get halved.\n",
"\n",
"And the change would be different for different values of `kernel_size` for
`nn.MaxPool2d()`.\n",
"\n",
"Also notice the value leftover in `max_pool_tensor` is the **maximum** value
from `random_tensor`.\n",
"\n",
"What's happening here?\n",
"\n",
"This is another important piece of the puzzle of neural networks.\n",
"\n",
"Essentially, **every layer in a neural network is trying to compress data from
higher dimensional space to lower dimensional space**. \n",
"\n",
"In other words, take a lot of numbers (raw data) and learn patterns in those
numbers, patterns that are predictive whilst also being *smaller* in size than the
original values.\n",
"\n",
"From an artificial intelligence perspective, you could consider the whole goal
of a neural network to *compress* information.\n",
"\n",
"![each layer of a neural network compresses the original input data into a
smaller representation that is (hopefully) capable of making predictions on future
input data](https://raw.githubusercontent.com/mrdbourke/pytorch-deep-learning/
main/images/03-conv-net-as-compression.png)\n",
"\n",
"This means, that from the point of view of a neural network, intelligence is
compression.\n",
"\n",
"This is the idea of the use of a `nn.MaxPool2d()` layer: take the maximum
value from a portion of a tensor and disregard the rest.\n",
"\n",
"In essence, lowering the dimensionality of a tensor whilst still retaining a
(hopefully) significant portion of the information.\n",
"\n",
"It is the same story for a `nn.Conv2d()` layer.\n",
 "\n",
"Except instead of just taking the maximum, the `nn.Conv2d()` performs a
convolutional operation on the data (see this in action on the [CNN Explainer
webpage](https://poloclub.github.io/cnn-explainer/)).\n",
"\n",
"> **Exercise:** What do you think the
[`nn.AvgPool2d()`](https://pytorch.org/docs/stable/generated/torch.nn.AvgPool2d.htm
l) layer does? Try making a random tensor like we did above and passing it through.
Check the input and output shapes as well as the input and output values.\n",
"\n",
"> **Extra-curriculum:** Lookup \"most common convolutional neural networks\",
what architectures do you find? Are any of them contained within the
[`torchvision.models`](https://pytorch.org/vision/stable/models.html) library? What
do you think you could do with these?"
]
},
{
"cell_type": "markdown",
"id": "39a3c646-52f0-4f4b-8527-2fc33d0dfb13",
"metadata": {
"id": "39a3c646-52f0-4f4b-8527-2fc33d0dfb13"
},
"source": [
"### 7.3 Setup a loss function and optimizer for `model_2`\n",
"\n",
"We've stepped through the layers in our first CNN enough.\n",
"\n",
"But remember, if something still isn't clear, try starting small.\n",
"\n",
"Pick a single layer of a model, pass some data through it and see what
happens.\n",
"\n",
"Now it's time to move forward and get to training!\n",
"\n",
"Let's setup a loss function and an optimizer.\n",
"\n",
"We'll use the functions as before, `nn.CrossEntropyLoss()` as the loss
function (since we're working with multi-class classification data).\n",
"\n",
"And `torch.optim.SGD()` as the optimizer to optimize `model_2.parameters()`
with a learning rate of `0.1`."
]
},
{
"cell_type": "code",
"execution_count": 40,
"id": "06a76a1b-5f6f-4018-bf7b-8388b385476f",
"metadata": {
"id": "06a76a1b-5f6f-4018-bf7b-8388b385476f"
},
"outputs": [],
"source": [
"# Setup loss and optimizer\n",
"loss_fn = nn.CrossEntropyLoss()\n",
"optimizer = torch.optim.SGD(params=model_2.parameters(), \n",
" lr=0.1)"
]
},
{
 "cell_type": "markdown",
"id": "758bc223-a244-4604-a07a-e2fc2f96c2f6",
"metadata": {
"id": "758bc223-a244-4604-a07a-e2fc2f96c2f6"
},
"source": [
"### 7.4 Training and testing `model_2` using our training and test functions\
n",
"\n",
"Loss and optimizer ready!\n",
"\n",
"Time to train and test.\n",
"\n",
"We'll use our `train_step()` and `test_step()` functions we created before.\
n",
"\n",
"We'll also measure the time to compare it to our other models."
]
},
{
"cell_type": "code",
"execution_count": 41,
"id": "861d126e-d876-40b3-9b7a-66cfc2f1bf05",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 327,
"referenced_widgets": [
"2b9c90ceb8554eaaaaf33acacecbcc11",
"9051de473b1b456592576115260f0c48",
"a315b7b535e4461ca11a6d96ca74411c",
"4cd01ed6c2534d4e80a0af2e9da02052",
"eb3ca30526e24fff9194d4e82436df99",
"6117713ba6c9490fab837cfbad7d442b",
"d426a4e9fff447ea95c7256372e272b7",
"fdcf4db7208d42b5acb35bf56c72fd82",
"606df8221adf48308f48c86ac54475e8",
"ebc6218d07f54005b0e46553b83464e1",
"03435ac81db84ae6bb6cb99bb78167c2"
]
},
"id": "861d126e-d876-40b3-9b7a-66cfc2f1bf05",
"outputId": "77ae1601-bf72-4239-b3f6-9fb99a1c62c0"
},
"outputs": [
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "2b9c90ceb8554eaaaaf33acacecbcc11",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
" 0%| | 0/3 [00:00<?, ?it/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
 {
"name": "stdout",
"output_type": "stream",
"text": [
"Epoch: 0\n",
"---------\n",
"Train loss: 0.59302 | Train accuracy: 78.41%\n",
"Test loss: 0.39771 | Test accuracy: 86.01%\n",
"\n",
"Epoch: 1\n",
"---------\n",
"Train loss: 0.36149 | Train accuracy: 87.00%\n",
"Test loss: 0.35713 | Test accuracy: 87.00%\n",
"\n",
"Epoch: 2\n",
"---------\n",
"Train loss: 0.32354 | Train accuracy: 88.28%\n",
"Test loss: 0.32857 | Test accuracy: 88.38%\n",
"\n",
"Train time on cuda: 44.250 seconds\n"
]
}
],
"source": [
"torch.manual_seed(42)\n",
"\n",
"# Measure time\n",
"from timeit import default_timer as timer\n",
"train_time_start_model_2 = timer()\n",
"\n",
"# Train and test model \n",
"epochs = 3\n",
"for epoch in tqdm(range(epochs)):\n",
" print(f\"Epoch: {epoch}\\n---------\")\n",
" train_step(data_loader=train_dataloader, \n",
" model=model_2, \n",
" loss_fn=loss_fn,\n",
" optimizer=optimizer,\n",
" accuracy_fn=accuracy_fn,\n",
" device=device\n",
" )\n",
" test_step(data_loader=test_dataloader,\n",
" model=model_2,\n",
" loss_fn=loss_fn,\n",
" accuracy_fn=accuracy_fn,\n",
" device=device\n",
" )\n",
"\n",
"train_time_end_model_2 = timer()\n",
"total_train_time_model_2 = print_train_time(start=train_time_start_model_2,\
n",
" end=train_time_end_model_2,\n",
" device=device)"
]
},
{
"cell_type": "markdown",
"id": "cfec7b7e-4dba-4016-957a-a29c6c10fde0",
"metadata": {
 "id": "cfec7b7e-4dba-4016-957a-a29c6c10fde0"
},
"source": [
"Woah! Looks like the convolutional and max pooling layers helped improve
performance a little.\n",
"\n",
"Let's evaluate `model_2`'s results with our `eval_model()` function."
]
},
{
"cell_type": "code",
"execution_count": 42,
"id": "c1bf8b89-1389-4395-a1c4-9c6e94d9e71c",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "c1bf8b89-1389-4395-a1c4-9c6e94d9e71c",
"outputId": "6ccddfc5-a86b-409d-d329-0d140f5b2f10"
},
"outputs": [
{
"data": {
"text/plain": [
"{'model_name': 'FashionMNISTModelV2',\n",
" 'model_loss': 0.3285697102546692,\n",
" 'model_acc': 88.37859424920129}"
]
},
"execution_count": 42,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Get model_2 results \n",
"model_2_results = eval_model(\n",
" model=model_2,\n",
" data_loader=test_dataloader,\n",
" loss_fn=loss_fn,\n",
" accuracy_fn=accuracy_fn\n",
")\n",
"model_2_results"
]
},
{
"cell_type": "markdown",
"id": "24c5ff68-b0bc-4b09-9da6-696736bc262d",
"metadata": {
"id": "24c5ff68-b0bc-4b09-9da6-696736bc262d"
},
"source": [
"## 8. Compare model results and training time\n",
"\n",
"We've trained three different models.\n",
"\n",
"1. `model_0` - our baseline model with two `nn.Linear()` layers.\n",
"2. `model_1` - the same setup as our baseline model except with `nn.ReLU()`
layers in between the `nn.Linear()` layers.\n",
 "3. `model_2` - our first CNN model that mimics the TinyVGG architecture on the
CNN Explainer website.\n",
"\n",
"This is a regular practice in machine learning.\n",
"\n",
"Building multiple models and performing multiple training experiments to see
which performs best.\n",
"\n",
"Let's combine our model results dictionaries into a DataFrame and find out."
]
},
{
"cell_type": "code",
"execution_count": 43,
"id": "52d84ee1-1ad4-4860-b147-f8912c1febc7",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 143
},
"id": "52d84ee1-1ad4-4860-b147-f8912c1febc7",
"outputId": "6eebb1ea-293a-4ae1-c054-f6154868426c"
},
"outputs": [
{
"data": {
"text/html": [
"\n",
" <div id=\"df-49abf68d-2959-4c68-b57b-6aeb2b48c2aa\">\n",
" <div class=\"colab-df-container\">\n",
" <div>\n",
"<style scoped>\n",
" .dataframe tbody tr th:only-of-type {\n",
" vertical-align: middle;\n",
" }\n",
"\n",
" .dataframe tbody tr th {\n",
" vertical-align: top;\n",
" }\n",
"\n",
" .dataframe thead th {\n",
" text-align: right;\n",
" }\n",
"</style>\n",
"<table border=\"1\" class=\"dataframe\">\n",
" <thead>\n",
" <tr style=\"text-align: right;\">\n",
" <th></th>\n",
" <th>model_name</th>\n",
" <th>model_loss</th>\n",
" <th>model_acc</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <th>0</th>\n",
" <td>FashionMNISTModelV0</td>\n",
" <td>0.476639</td>\n",
" <td>83.426518</td>\n",
 " </tr>\n",
" <tr>\n",
" <th>1</th>\n",
" <td>FashionMNISTModelV1</td>\n",
" <td>0.685001</td>\n",
" <td>75.019968</td>\n",
" </tr>\n",
" <tr>\n",
" <th>2</th>\n",
" <td>FashionMNISTModelV2</td>\n",
" <td>0.328570</td>\n",
" <td>88.378594</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>\n",
"</div>\n",
" <button class=\"colab-df-convert\"
onclick=\"convertToInteractive('df-49abf68d-2959-4c68-b57b-6aeb2b48c2aa')\"\n",
" title=\"Convert this dataframe to an interactive table.\"\n",
" style=\"display:none;\">\n",
" \n",
" <svg xmlns=\"http://www.w3.org/2000/svg\" height=\"24px\"viewBox=\"0 0 24
24\"\n",
" width=\"24px\">\n",
" <path d=\"M0 0h24v24H0V0z\" fill=\"none\"/>\n",
" <path d=\"M18.56 5.44l.94 2.06.94-2.06 2.06-.94-2.06-.94-.94-2.06-.94
2.06-2.06.94zm-11 1L8.5 8.5l.94-2.06 2.06-.94-2.06-.94L8.5 2.5l-.94 2.06-
2.06.94zm10 10l.94 2.06.94-2.06 2.06-.94-2.06-.94-.94-2.06-.94
2.06-2.06.94z\"/><path d=\"M17.41 7.96l-1.37-1.37c-.4-.4-.92-.59-1.43-.59-.52 0-
1.04.2-1.43.59L10.3 9.45l-7.72 7.72c-.78.78-.78 2.05 0 2.83L4 21.41c.39.39.9.59
1.41.59.51 0 1.02-.2 1.41-.59l7.78-7.78 2.81-2.81c.8-.78.8-2.07 0-2.86zM5.41 20L4
18.59l7.72-7.72 1.47 1.35L5.41 20z\"/>\n",
" </svg>\n",
" </button>\n",
" \n",
" <style>\n",
" .colab-df-container {\n",
" display:flex;\n",
" flex-wrap:wrap;\n",
" gap: 12px;\n",
" }\n",
"\n",
" .colab-df-convert {\n",
" background-color: #E8F0FE;\n",
" border: none;\n",
" border-radius: 50%;\n",
" cursor: pointer;\n",
" display: none;\n",
" fill: #1967D2;\n",
" height: 32px;\n",
" padding: 0 0 0 0;\n",
" width: 32px;\n",
" }\n",
"\n",
" .colab-df-convert:hover {\n",
" background-color: #E2EBFA;\n",
" box-shadow: 0px 1px 2px rgba(60, 64, 67, 0.3), 0px 1px 3px 1px
rgba(60, 64, 67, 0.15);\n",
" fill: #174EA6;\n",
 " }\n",
"\n",
" [theme=dark] .colab-df-convert {\n",
" background-color: #3B4455;\n",
" fill: #D2E3FC;\n",
" }\n",
"\n",
" [theme=dark] .colab-df-convert:hover {\n",
" background-color: #434B5C;\n",
" box-shadow: 0px 1px 3px 1px rgba(0, 0, 0, 0.15);\n",
" filter: drop-shadow(0px 1px 2px rgba(0, 0, 0, 0.3));\n",
" fill: #FFFFFF;\n",
" }\n",
" </style>\n",
"\n",
" <script>\n",
" const buttonEl =\n",
" document.querySelector('#df-49abf68d-2959-4c68-b57b-6aeb2b48c2aa
button.colab-df-convert');\n",
" buttonEl.style.display =\n",
" google.colab.kernel.accessAllowed ? 'block' : 'none';\n",
"\n",
" async function convertToInteractive(key) {\n",
" const element = document.querySelector('#df-49abf68d-2959-4c68-
b57b-6aeb2b48c2aa');\n",
" const dataTable =\n",
" await
google.colab.kernel.invokeFunction('convertToInteractive',\n",
" [key], {});\n",
" if (!dataTable) return;\n",
"\n",
" const docLinkHtml = 'Like what you see? Visit the ' +\n",
" '<a target=\"_blank\"
href=https://colab.research.google.com/notebooks/data_table.ipynb>data table
notebook</a>'\n",
" + ' to learn more about interactive tables.';\n",
" element.innerHTML = '';\n",
" dataTable['output_type'] = 'display_data';\n",
" await google.colab.output.renderOutput(dataTable, element);\n",
" const docLink = document.createElement('div');\n",
" docLink.innerHTML = docLinkHtml;\n",
" element.appendChild(docLink);\n",
" }\n",
" </script>\n",
" </div>\n",
" </div>\n",
" "
],
"text/plain": [
" model_name model_loss model_acc\n",
"0 FashionMNISTModelV0 0.476639 83.426518\n",
"1 FashionMNISTModelV1 0.685001 75.019968\n",
"2 FashionMNISTModelV2 0.328570 88.378594"
]
},
"execution_count": 43,
"metadata": {},
"output_type": "execute_result"
}
 ],
"source": [
"import pandas as pd\n",
"compare_results = pd.DataFrame([model_0_results, model_1_results,
model_2_results])\n",
"compare_results"
]
},
{
"cell_type": "markdown",
"id": "c67f3fb5-ce7b-40b8-86a0-2797492de0ef",
"metadata": {
"id": "c67f3fb5-ce7b-40b8-86a0-2797492de0ef"
},
"source": [
"Nice!\n",
"\n",
"We can add the training time values too."
]
},
{
"cell_type": "code",
"execution_count": 44,
"id": "297af38f-e69f-4c6f-9027-fcaf0482a55c",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 143
},
"id": "297af38f-e69f-4c6f-9027-fcaf0482a55c",
"outputId": "67c01781-78c2-47e8-a1ee-452869bea5e2"
},
"outputs": [
{
"data": {
"text/html": [
"\n",
" <div id=\"df-961e0788-f5fc-4668-9303-bbfe454246f9\">\n",
" <div class=\"colab-df-container\">\n",
" <div>\n",
"<style scoped>\n",
" .dataframe tbody tr th:only-of-type {\n",
" vertical-align: middle;\n",
" }\n",
"\n",
" .dataframe tbody tr th {\n",
" vertical-align: top;\n",
" }\n",
"\n",
" .dataframe thead th {\n",
" text-align: right;\n",
" }\n",
"</style>\n",
"<table border=\"1\" class=\"dataframe\">\n",
" <thead>\n",
" <tr style=\"text-align: right;\">\n",
" <th></th>\n",
" <th>model_name</th>\n",
" <th>model_loss</th>\n",
 " <th>model_acc</th>\n",
" <th>training_time</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <th>0</th>\n",
" <td>FashionMNISTModelV0</td>\n",
" <td>0.476639</td>\n",
" <td>83.426518</td>\n",
" <td>32.348722</td>\n",
" </tr>\n",
" <tr>\n",
" <th>1</th>\n",
" <td>FashionMNISTModelV1</td>\n",
" <td>0.685001</td>\n",
" <td>75.019968</td>\n",
" <td>36.877976</td>\n",
" </tr>\n",
" <tr>\n",
" <th>2</th>\n",
" <td>FashionMNISTModelV2</td>\n",
" <td>0.328570</td>\n",
" <td>88.378594</td>\n",
" <td>44.249765</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>\n",
"</div>\n",
" <button class=\"colab-df-convert\"
onclick=\"convertToInteractive('df-961e0788-f5fc-4668-9303-bbfe454246f9')\"\n",
" title=\"Convert this dataframe to an interactive table.\"\n",
" style=\"display:none;\">\n",
" \n",
" <svg xmlns=\"http://www.w3.org/2000/svg\" height=\"24px\"viewBox=\"0 0 24
24\"\n",
" width=\"24px\">\n",
" <path d=\"M0 0h24v24H0V0z\" fill=\"none\"/>\n",
" <path d=\"M18.56 5.44l.94 2.06.94-2.06 2.06-.94-2.06-.94-.94-2.06-.94
2.06-2.06.94zm-11 1L8.5 8.5l.94-2.06 2.06-.94-2.06-.94L8.5 2.5l-.94 2.06-
2.06.94zm10 10l.94 2.06.94-2.06 2.06-.94-2.06-.94-.94-2.06-.94
2.06-2.06.94z\"/><path d=\"M17.41 7.96l-1.37-1.37c-.4-.4-.92-.59-1.43-.59-.52 0-
1.04.2-1.43.59L10.3 9.45l-7.72 7.72c-.78.78-.78 2.05 0 2.83L4 21.41c.39.39.9.59
1.41.59.51 0 1.02-.2 1.41-.59l7.78-7.78 2.81-2.81c.8-.78.8-2.07 0-2.86zM5.41 20L4
18.59l7.72-7.72 1.47 1.35L5.41 20z\"/>\n",
" </svg>\n",
" </button>\n",
" \n",
" <style>\n",
" .colab-df-container {\n",
" display:flex;\n",
" flex-wrap:wrap;\n",
" gap: 12px;\n",
" }\n",
"\n",
" .colab-df-convert {\n",
" background-color: #E8F0FE;\n",
" border: none;\n",
" border-radius: 50%;\n",
 " cursor: pointer;\n",
" display: none;\n",
" fill: #1967D2;\n",
" height: 32px;\n",
" padding: 0 0 0 0;\n",
" width: 32px;\n",
" }\n",
"\n",
" .colab-df-convert:hover {\n",
" background-color: #E2EBFA;\n",
" box-shadow: 0px 1px 2px rgba(60, 64, 67, 0.3), 0px 1px 3px 1px
rgba(60, 64, 67, 0.15);\n",
" fill: #174EA6;\n",
" }\n",
"\n",
" [theme=dark] .colab-df-convert {\n",
" background-color: #3B4455;\n",
" fill: #D2E3FC;\n",
" }\n",
"\n",
" [theme=dark] .colab-df-convert:hover {\n",
" background-color: #434B5C;\n",
" box-shadow: 0px 1px 3px 1px rgba(0, 0, 0, 0.15);\n",
" filter: drop-shadow(0px 1px 2px rgba(0, 0, 0, 0.3));\n",
" fill: #FFFFFF;\n",
" }\n",
" </style>\n",
"\n",
" <script>\n",
" const buttonEl =\n",
" document.querySelector('#df-961e0788-f5fc-4668-9303-bbfe454246f9
button.colab-df-convert');\n",
" buttonEl.style.display =\n",
" google.colab.kernel.accessAllowed ? 'block' : 'none';\n",
"\n",
" async function convertToInteractive(key) {\n",
" const element = document.querySelector('#df-961e0788-f5fc-4668-
9303-bbfe454246f9');\n",
" const dataTable =\n",
" await
google.colab.kernel.invokeFunction('convertToInteractive',\n",
" [key], {});\n",
" if (!dataTable) return;\n",
"\n",
" const docLinkHtml = 'Like what you see? Visit the ' +\n",
" '<a target=\"_blank\"
href=https://colab.research.google.com/notebooks/data_table.ipynb>data table
notebook</a>'\n",
" + ' to learn more about interactive tables.';\n",
" element.innerHTML = '';\n",
" dataTable['output_type'] = 'display_data';\n",
" await google.colab.output.renderOutput(dataTable, element);\n",
" const docLink = document.createElement('div');\n",
" docLink.innerHTML = docLinkHtml;\n",
" element.appendChild(docLink);\n",
" }\n",
" </script>\n",
" </div>\n",
" </div>\n",
 " "
],
"text/plain": [
" model_name model_loss model_acc training_time\n",
"0 FashionMNISTModelV0 0.476639 83.426518 32.348722\n",
"1 FashionMNISTModelV1 0.685001 75.019968 36.877976\n",
"2 FashionMNISTModelV2 0.328570 88.378594 44.249765"
]
},
"execution_count": 44,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Add training times to results comparison\n",
"compare_results[\"training_time\"] = [total_train_time_model_0,\n",
" total_train_time_model_1,\n",
" total_train_time_model_2]\n",
"compare_results"
]
},
{
"cell_type": "markdown",
"id": "fbbe5832-1081-4c76-8d5b-06c7a06da7b9",
"metadata": {
"id": "fbbe5832-1081-4c76-8d5b-06c7a06da7b9"
},
"source": [
"It looks like our CNN (`FashionMNISTModelV2`) model performed the best (lowest
loss, highest accuracy) but had the longest training time.\n",
"\n",
"And our baseline model (`FashionMNISTModelV0`) performed better than `model_1`
(`FashionMNISTModelV1`).\n",
"\n",
"### Performance-speed tradeoff\n",
"\n",
"Something to be aware of in machine learning is the **performance-speed**
tradeoff.\n",
"\n",
"Generally, you get better performance out of a larger, more complex model
(like we did with `model_2`).\n",
"\n",
"However, this performance increase often comes at a sacrifice of training
speed and inference speed.\n",
"\n",
"> **Note:** The training times you get will be very dependent on the hardware
you use. \n",
">\n",
"> Generally, the more CPU cores you have, the faster your models will train on
CPU. And similar for GPUs.\n",
"> \n",
"> Newer hardware (in terms of age) will also often train models faster due to
incorporating technological advances.\n",
"\n",
"How about we get visual?"
]
},
{
 "cell_type": "code",
"execution_count": 45,
"id": "5eb0df60-9318-47d0-adce-f8788ed3999e",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 449
},
"id": "5eb0df60-9318-47d0-adce-f8788ed3999e",
"outputId": "d904e039-1544-49f5-bed7-b17555d03b5a"
},
"outputs": [
{
"data": {
"image/png":
"iVBORw0KGgoAAAANSUhEUgAAAr0AAAGwCAYAAACkUt2bAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIH
ZlcnNpb24zLjcuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/
bCgiHAAAACXBIWXMAAA9hAAAPYQGoP6dpAAA9IklEQVR4nO3de1wWZd7H8S/
KQQ6iIB5CQQkkRUFwzVRKH9RdZAnbLEtTsIDdLMxYz50WTQ1IqTyUbB7AntXUAjuYq2uaFqilGIrJqnkuj6
sGHhFhnj96vLc7UBE1dPy8X695xT1zzTW/ueaVfu/xmsHGMAxDAAAAgInVqukCAAAAgJuN0AsAAADTI/
QCAADA9Ai9AAAAMD1CLwAAAEyP0AsAAADTI/
QCAADA9GxrugDgVlBeXq6DBw+qbt26srGxqelyAABAFRiGoVOnTsnT01O1al35Xi6hF5B08OBBeXl51XQZA
ACgGg4cOKBmzZpdsQ2hF5BUt25dST//T+Pq6lrD1QAAgKooLi6Wl5eX5e/
xKyH0ApJlSoOrqyuhFwCA20xVpibyIBsAAABMj9ALAAAA0yP0AgAAwPQIvQAAADA9Qi8AAABMj9ALAAAA0y
P0AgAAwPQIvQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwPQIvQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwPQIv
QAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwPQIvQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwPQIvQAAADA9Qi8A
AABMj9ALAAAA0yP0AgAAwPQIvQAAADA925ouALiVtE1arloOTjVdBgAAprI3JbKmS+BOLwAAAMyP0AsAAAD
TI/QCAADA9Ai9AAAAMD1CLwAAAEyP0AsAAADTI/QCAADA9Ai9AAAAMD1CLwAAAEyP0AsAAADTI/
QCAADA9Ai9AAAAMD1CLwAAAEyP0AsAAADTI/QCAADA9Ai9AAAAMD1CLwAAAEyP0AsAAADTI/
QCAADA9Ai9AAAAMD1CLwAAAEyP0AsAAADTI/QCAADA9Ai9AAAAMD1CLwAAAEzPdKF39erVsrGx0U8//
XTZNmPHjlVwcPBvVtOdqirX4tdatGiht95666bVBAAA7kw1GnqffPJJ2djYVFi+//
77m3rcESNGaOXKlTe0z0sBz83NTefPn7fatmHDBsu5/
bp9mzZtVFZWZtW+fv36yszMtHz+dRDcvHmzevfurUaNGqlOnTpq0aKFHn/8cR09elRjx46tdEx/uUj/
HfvBgwdXOJeEhATZ2NjoySefvP6BqaYLFy7Iw8NDKSkplW4fP368GjdurNLSUmVnZ+v3v/
+9GjZsKFdXV3Xu3FnLly//jSsGAAC3shq/
09urVy8dOnTIavHx8bmpx3RxcVGDBg1uSt9169bV4sWLrdbNnj1b3t7elbbfvXu33nvvvSr3f+zYMfXo0UP
u7u5avny5CgsLlZGRIU9PT505c0YjRoywGstmzZrp1VdftVp3iZeXlxYsWKBz585Z1p0/f17z58+/bL2/
FXt7ew0cOFAZGRkVthmGoczMTMXExMjOzk5ffvmlfv/
732vp0qXKy8tTWFiYoqKi9O2339ZA5QAA4FZU46HXwcFBTZo0sVqmTJmiwMBAOTs7y8vLS88+
+6xOnz5t2Wffvn2KioqSm5ubnJ2d1aZNGy1dutSq37y8PHXo0EFOTk7q0qWLtm/
fbtn26+kN5eXlevXVV9WsWTM5ODgoODhYy5Yts2zfu3evbGxslJ2drbCwMDk5Oaldu3Zat25dhfMZNGiQ5s
yZY/l87tw5LViwQIMGDar0/
J977jklJSWppKSkSuOVm5uroqIizZo1SyEhIfLx8VFYWJjefPNN+fj4yMXFxWosa9eurbp161qtu6R9+/
by8vJSdna2ZV12dra8vb0VEhJiddySkhINHTrUcnf5/vvv14YNG6zaLF26VP7+/
nJ0dFRYWJj27t1bof6cnBw98MADcnR0lJeXl4YOHaozZ85Ueq5xcXHasWOHcnJyrNavWbNGu3fvVlxcnCTp
rbfe0qhRo3TvvfeqZcuWeu2119SyZUt9+umnVRpTAABgfjUeeitTq1YtTZ06Vd99953mzp2rVatWadSoUZb
tCQkJKikp0ZdffqmCggKlpqbKxcXFqo+XXnpJaWlp2rhxo2xtbRUbG3vZ402ZMkVpaWmaPHmytmzZovDwcP
Xu3Vs7d+6s0OeIESOUn58vf39/9e/fXxcvXrRqEx0dra+++kr79+
+XJGVlZalFixZq3759pcdOTEzUxYsXNW3atCqNTZMmTXTx4kUtXrxYhmFUaZ8riY2NtbqbOmfOHD311FMV2
o0aNUpZWVmaO3euNm3aJD8/
P4WHh+vEiROSpAMHDqhPnz6KiopSfn6+4uPjNWbMGKs+du3apV69eumRRx7Rli1btHDhQuXk5GjIkCGV1hY
YGKh7773X6kuEJGVkZKhLly5q1apVpfuVl5fr1KlTcnd3v+x5l5SUqLi42GoBAADmVeOhd8mSJXJxcbEsff
v2VWJiosLCwtSiRQt1795dEyZM0KJFiyz77N+/
X6GhoQoMDNTdd9+tBx98UF27drXqd+LEierWrZsCAgI0ZswYrV27tsJc20smT56s0aNHq1+/
frrnnnuUmpqq4ODgCg9UjRgxQpGRkfL399e4ceO0b9+
+CvOPGzVqpIiICMuc3Dlz5lwxcDs5OSkpKUnJyckqKiq66nh16tRJL774op544gl5eHgoIiJCkyZN0pEjR6
66b2UGDhyonJwc7du3T/
v27VNubq4GDhxo1ebMmTOaMWOGJk2apIiICAUEBGjmzJlydHTU7NmzJUkzZsyQr6+v0tLSdM8992jAgAEV5
gQnJydrwIABSkxMVMuWLdWlSxdNnTpV77333mWvTVxcnD744APLnf5Tp07pww8/
vOKYTp48WadPn9Zjjz122TbJycmqV6+eZfHy8qrKcAEAgNtUjYfesLAw5efnW5apU6fq888/
V48ePdS0aVPVrVtX0dHROn78uM6ePStJGjp0qCZMmKDQ0FAlJSVpy5YtFfoNCgqy/
HzXXXdJko4ePVqhXXFxsQ4ePKjQ0FCr9aGhoSosLKxWn7GxscrMzNTu3bu1bt06DRgw4IpjEBcXpwYNGig1
NfWK7S6ZOHGiDh8+rPT0dLVp00bp6elq1aqVCgoKqrT/
LzVs2FCRkZHKzMxURkaGIiMj5eHhYdVm165dKi0ttRojOzs7dezY0TJGhYWFuu++
+6z269y5s9XnzZs3KzMz0+pLTnh4uMrLy7Vnz55K6+vfv7/KysosX3oWLlyoWrVq6fHHH6+0/
fz58zVu3DgtWrRIjRo1uux5v/DCCyoqKrIsBw4cuGxbAABw+6vx0Ovs7Cw/Pz/
LUlJSogcffFBBQUHKyspSXl6e3n77bUk/P9EvSfHx8dq9e7eio6NVUFCgDh06VJgeYGdnZ/
n50hsLysvLr6vWqvYZERGhc+fOKS4uTlFRUVd9aM7W1lYTJ07UlClTdPDgwSrV0qBBA/
Xt21eTJ09WYWGhPD09NXny5Gs4m/
+6FNLnzp17xTuo1+v06dN6+umnrb7kbN68WTt37pSvr2+l+7i6uurRRx+1TMHIyMjQY489VmE6iyQtWLBA8
fHxWrRokXr27HnFWhwcHOTq6mq1AAAA86rx0PtreXl5Ki8vV1pamjp16iR/f/
9Kg6CXl5cGDx6s7OxsDR8+XDNnzqzW8VxdXeXp6anc3Fyr9bm5uQoICKhWn7a2toqJidHq1aurHCL79u2rN
m3aaNy4cdd8PHt7e/
n6+l72gbCr6dWrly5cuKDS0lKFh4dX2O7r6yt7e3urMSotLdWGDRssY9S6dWt98803VvutX7/
e6nP79u21bds2qy85lxZ7e/vL1hcXF6ecnBwtWbJEa9eutTzA9kvvv/++nnrqKb3//
vuKjIy8pvMHAADmZ1vTBfyan5+fSktLNW3aNEVFRSk3N1fp6elWbRITExURESF/
f3+dPHlSX3zxhVq3bl3tY44cOVJJSUny9fVVcHCwMjIylJ+fr3nz5lW7z/
Hjx2vkyJHX9Gq0lJSUSkPnLy1ZskQLFixQv3795O/
vL8Mw9Omnn2rp0qWVvt6rKmrXrm2ZplC7du0K252dnfXMM89o5MiRcnd3l7e3t15//
XWdPXvWEkAHDx6stLQ0jRw5UvHx8crLy7N617AkjR49Wp06ddKQIUMUHx8vZ2dnbdu2TStWrND06dMvW1/
Xrl3l5+enmJgYtWrVSl26dLHaPn/
+fA0aNEhTpkzRfffdp8OHD0uSHB0dVa9evWqNCQAAMJdb7k5vu3bt9MYbbyg1NVVt27bVvHnzlJycbNWmrK
xMCQkJat26tXr16iV/
f3+988471T7m0KFDNWzYMA0fPlyBgYFatmyZPvnkE7Vs2bLafdrb28vDw8PqF1JcTffu3dW9e/
cKb4T4pYCAADk5OWn48OEKDg5Wp06dtGjRIs2aNUvR0dHVrvdq/8SfkpKiRx55RNHR0Wrfvr2+//57LV+
+XG5ubpIkb29vZWVl6aOPPlK7du2Unp6u1157zaqPoKAgrVmzRjt27NADDzygkJAQ/
e1vf5Onp+cVa7OxsVFsbKxOnjxZ6Z3zd999VxcvXlRCQoLuuusuy/
L8889XYyQAAIAZ2Rg34r1XwG2uuLj457c4JC5SLQenmi4HAABT2Ztyc6YeXvr7u6io6KrP59xyd3oBAACAG
43QCwAAANMj9AIAAMD0CL0AAAAwPUIvAAAATI/QCwAAANMj9AIAAMD0CL0AAAAwPUIvAAAATI/
QCwAAANMj9AIAAMD0CL0AAAAwPUIvAAAATI/QCwAAANMj9AIAAMD0CL0AAAAwPUIvAAAATI/
QCwAAANMj9AIAAMD0CL0AAAAwPUIvAAAATI/
QCwAAANMj9AIAAMD0CL0AAAAwPduaLgC4lWwdFy5XV9eaLgMAANxg3OkFAACA6RF6AQAAYHqEXgAAAJgeoR
cAAACmR+gFAACA6RF6AQAAYHqEXgAAAJgeoRcAAACmR+gFAACA6RF6AQAAYHqEXgAAAJgeoRcAAACmR+gFA
ACA6RF6AQAAYHqEXgAAAJgeoRcAAACmR+gFAACA6RF6AQAAYHqEXgAAAJgeoRcAAACmR+gFAACA6RF6AQAA
YHqEXgAAAJgeoRcAAACmR+gFAACA6RF6AQAAYHqEXgAAAJgeoRcAAACmR+gFAACA6RF6AQAAYHqEXgAAAJg
eoRcAAACmR+gFAACA6RF6AQAAYHqEXgAAAJgeoRcAAACmR+gFAACA6RF6AQAAYHqEXgAAAJgeoRcAAACmR+
gFAACA6RF6AQAAYHqEXgAAAJgeoRcAAACmR+gFAACA6RF6AQAAYHqEXgAAAJgeoRcAAACmR+gFAACA6RF6A
QAAYHqEXgAAAJgeoRcAAACmR+gFAACA6RF6AQAAYHqEXgAAAJgeoRcAAACmR+gFAACA6RF6AQAAYHqEXgAA
AJgeoRcAAACmZ1vTBQC3krZJy1XLwammywAA3Gb2pkTWdAm4Cu70AgAAwPQIvQAAADA9Qi8AAABMj9ALAAA
A0yP0AgAAwPQIvQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwPQIvQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwP
Rsq9rQzc1NNjY2VWp74sSJahcEAAAA3GhVDr1vvfXWTSwDAAAAuHmqHHoHDRp0M+sAAAAAbppqz+ndtWuXX
n75ZfXv319Hjx6VJP3zn//Ud999d8OKAwAAAG6EaoXeNWvWKDAwUF9//
bWys7N1+vRpSdLmzZuVlJR0QwsEAAAArle1Qu+YMWM0YcIErVixQvb29pb13bt31/
r1629YcQAAAMCNUK3QW1BQoIcffrjC+kaNGuk///nPdRcFAAAA3EjVCr3169fXoUOHKqz/
9ttv1bRp0+suCgAAALiRqhV6+/
Xrp9GjR+vw4cOysbFReXm5cnNzNWLECMXExNzoGgEAAIDrUq3Q+9prr6lVq1by8vLS6dOnFRAQoK5du6pLl
y56+eWXb3SNAAAAwHWp8nt6f8ne3l4zZ87UK6+8oq1bt+r06dMKCQlRy5Ytb3R9AAAAwHWrVui9xNvbW97e
3jeqFgAAAOCmqHLoHTZsWJU7feONN6pVzI2wevVqhYWF6eTJk6pfv36lbcaOHauPPvpI+fn5v2ltd5qqXIt
fa9GihRITE5WYmHhTawMAAHeWKs/p/fbbb62W2bNn6+9//7tWr16t1atX691339Xs2bOvKUg+
+eSTsrGxqbB8//331TmXKhsxYoRWrlx5Q/
tcvXq1bGxs5ObmpvPnz1tt27Bhg+Xcft2+TZs2Kisrs2pfv359ZWZmWj63aNFCb731luXz5s2b1bt3bzVq1
Eh16tRRixYt9Pjjj+vo0aMaO3ZspWP6y0X679gPHjy4wrkkJCTIxsZGTz755PUPTDVduHBBHh4eSklJqXT7
+PHj1bhxY5WWlurQoUN64okn5O/
vr1q1ahGYAQBABVUOvV988YVliYqKUrdu3fTDDz9o06ZN2rRpkw4cOKCwsDBFRkZeUwG9evXSoUOHrBYfH5
9rPpFr4eLiogYNGtyUvuvWravFixdbrZs9e/
Zlp4Hs3r1b7733XpX7P3bsmHr06CF3d3ctX75chYWFysjIkKenp86cOaMRI0ZYjWWzZs306quvWq27xMvLS
wsWLNC5c+cs686fP6/58+fX+LQVe3t7DRw4UBkZGRW2GYahzMxMxcTEyM7OTiUlJWrYsKFefvlltWvXrgaq
BQAAt7pqvb0hLS1NycnJcnNzs6xzc3PThAkTlJaWdk19OTg4qEmTJlbLlClTFBgYKGdnZ3l5eenZZ5+1/
KpjSdq3b5+ioqLk5uYmZ2dntWnTRkuXLrXqNy8vTx06dJCTk5O6dOmi7du3W7aNHTtWwcHBls/
l5eV69dVX1axZMzk4OCg4OFjLli2zbN+7d69sbGyUnZ2tsLAwOTk5qV27dlq3bl2F8xk0aJDmzJlj+Xzu3D
ktWLBAgwYNqvT8n3vuOSUlJamkpKRK45Wbm6uioiLNmjVLISEh8vHxUVhYmN588035+PjIxcXFaixr166tu
nXrWq27pH379vLy8lJ2drZlXXZ2try9vRUSEmJ13JKSEg0dOtRyd/n+++/Xhg0brNosXbpU/
v7+cnR0VFhYmPbu3Vuh/
pycHD3wwANydHSUl5eXhg4dqjNnzlR6rnFxcdqxY4dycnKs1q9Zs0a7d+9WXFycpJ/
vhE+ZMkUxMTGqV69elcYRAADcWaoVeouLi3Xs2LEK648dO6ZTp05df1G1amnq1Kn67rvvNHfuXK1atUqjRo
2ybE9ISFBJSYm+/
PJLFRQUKDU1VS4uLlZ9vPTSS0pLS9PGjRtla2ur2NjYyx5vypQpSktL0+TJk7VlyxaFh4erd+/
e2rlzZ4U+R4wYofz8fPn7+6t///66ePGiVZvo6Gh99dVX2r9/
vyQpKytLLVq0UPv27Ss9dmJioi5evKhp06ZVaWyaNGmiixcvavHixTIMo0r7XElsbKzV3dQ5c+boqaeeqtB
u1KhRysrK0ty5c7Vp0yb5+fkpPDxcJ06ckCQdOHBAffr0UVRUlPLz8xUfH68xY8ZY9bFr1y716tVLjzzyiL
Zs2aKFCxcqJydHQ4YMqbS2wMBA3XvvvVZfIiQpIyNDXbp0UatWrap93iUlJSouLrZaAACAeVUr9D788MN66
qmnlJ2drR9++EE//PCDsrKyFBcXpz59+lxTX0uWLJGLi4tl6du3rxITExUWFqYWLVqoe/
fumjBhghYtWmTZZ//+/
QoNDVVgYKDuvvtuPfjgg+ratatVvxMnTlS3bt0UEBCgMWPGaO3atRXm2l4yefJkjR49Wv369dM999yj1NRU
BQcHW82jlX6eCxwZGSl/f3+NGzdO+/btqzD/
uFGjRoqIiLDMyZ0zZ84VA7eTk5OSkpKUnJysoqKiq45Xp06d9OKLL+qJJ56Qh4eHIiIiNGnSJB05cuSq+1Z
m4MCBysnJ0b59+7Rv3z7l5uZq4MCBVm3OnDmjGTNmaNKkSYqIiFBAQIBmzpwpR0dHzZ49W5I0Y8YM+fr6Ki
0tTffcc48GDBhQYU5wcnKyBgwYoMTERLVs2VJdunTR1KlT9d5771322sTFxemDDz6w3Ok/
deqUPvzwwyuOaVUkJyerXr16lsXLy+u6+gMAALe2aoXe9PR0RURE6IknnlDz5s3VvHlzPfHEE+rVq5feeee
da+orLCxM+fn5lmXq1Kn6/PPP1aNHDzVt2lR169ZVdHS0jh8/
rrNnz0qShg4dqgkTJig0NFRJSUnasmVLhX6DgoIsP991112SpKNHj1ZoV1xcrIMHDyo0NNRqfWhoqAoLC6v
VZ2xsrDIzM7V7926tW7dOAwYMuOIYxMXFqUGDBkpNTb1iu0smTpyow4cPKz09XW3atFF6erpatWqlgoKCKu
3/
Sw0bNlRkZKQyMzOVkZGhyMhIeXh4WLXZtWuXSktLrcbIzs5OHTt2tIxRYWGh7rvvPqv9OnfubPV58+bNysz
MtPqSEx4ervLycu3Zs6fS+vr376+ysjLLl56FCxeqVq1aevzxx6/5XH/
phRdeUFFRkWU5cODAdfUHAABubdUKvU5OTnrnnXd0/Phxy9scTpw4oXfeeUfOzs7X1Jezs7P8/
PwsS0lJiR588EEFBQUpKytLeXl5evvttyX9/ES/
JMXHx2v37t2Kjo5WQUGBOnToUGF6gJ2dneXnS28sKC8vr87pXnOfEREROnfunOLi4hQVFXXVh+ZsbW01ceJ
ETZkyRQcPHqxSLQ0aNFDfvn01efJkFRYWytPTU5MnT76Gs/
mvSyF97ty5130H9UpOnz6tp59+2upLzubNm7Vz5075+vpWuo+rq6seffRRyxSMjIwMPfbYYxWms1wrBwcHu
bq6Wi0AAMC8qhV6L3F2dpa7u7vc3d2vOexeTl5ensrLy5WWlqZOnTrJ39+/
0iDo5eWlwYMHKzs7W8OHD9fMmTOrdTxXV1d5enoqNzfXan1ubq4CAgKq1aetra1iYmK0evXqKofIvn37qk2
bNho3btw1H8/
e3l6+vr6XfSDsanr16qULFy6otLRU4eHhFbb7+vrK3t7eaoxKS0u1YcMGyxi1bt1a33zzjdV+69evt/
rcvn17bdu2zepLzqXF3t7+svXFxcUpJydHS5Ys0dq1ay0PsAEAAFRVtULvpbcd1KtXzzK9oX79+ho/
fvx130318/NTaWmppk2bpt27d+t///d/
lZ6ebtUmMTFRy5cv1549e7Rp0yZ98cUXat26dbWPOXLkSKWmpmrhwoXavn27xowZo/z8fD3//
PPV7nP8+PE6duxYpSHyclJSUjRnzpwrhtclS5Zo4MCBWrJkiXbs2KHt27dr8uTJWrp0qR566KFq1Vq7dm0V
FhZq27Ztql27doXtzs7OeuaZZzRy5EgtW7ZM27Zt05///
GedPXvWEkAHDx6snTt3auTIkdq+fbvmz59v9a5hSRo9erTWrl2rIUOGKD8/
Xzt37tTHH3982QfZLunatav8/
PwUExOjVq1aqUuXLhXaXLpzfPr0aR07dkz5+fnatm1btcYDAACYT7V+DfFLL72k2bNnKyUlxTLPMycnR2PH
jtX58+c1ceLEahfUrl07vfHGG0pNTdULL7ygrl27Kjk5WTExMZY2ZWVlSkhI0A8//
CBXV1f16tVLb775ZrWPOXToUBUVFWn48OE6evSoAgIC9Mknn6hly5bV7tPe3r7C3Nir6d69u7p3765//
etfl20TEBAgJycnDR8+XAcOHJCDg4NatmypWbNmKTo6utr1Xu2f91NSUlReXq7o6GidOnVKHTp00PLlyy2v
rfP29lZWVpb++te/atq0aerYsaNee+01qzvdQUFBWrNmjV566SU98MADMgxDvr6
+V52fa2Njo9jYWL344ot64YUXKm3zy1es5eXlaf78+WrevHmlr00DAAB3HhujGu+98vT0VHp6unr37m21/
uOPP9azzz6rH3/88YYVCPwWiouLf36LQ+Ii1XJwqulyAAC3mb0p1/
bLuXBjXPr7u6io6Ko38Ko1veHEiROVviO1VatWlve2AgAAALeKaoXedu3aafr06RXWT58+nV8DCwAAgFtOt
eb0vv7664qMjNTnn39ueRfrunXrtH//fv3zn/+8oQUCAAAA16tad3q7deum7du3q0+fPvrpp5/0008/
qU+fPtqxY4ceeOCBG10jAAAAcF2qdadX+vmXI/
Tu3VudOnWyvKZs48aNklThATcAAACgJlUr9C5btkwxMTE6fvy4fv3yBxsbG5WVld2Q4gAAAIAboVrTG5577
jn17dtXBw8eVHl5udVC4AUAAMCtplqh98iRIxo2bJgaN258o+sBAAAAbrhqhd5HH31Uq1evvsGlAAAAADdH
teb0Tp8+XX379tVXX32lwMBA2dnZWW0fOnToDSkOAAAAuBGqFXrff/99/etf/
1KdOnW0evVq2djYWLbZ2NgQegEAAHBLqVbofemllzRu3DiNGTNGtWpVa4YEAAAA8JupVmK9cOGCHn/
8cQIvAAAAbgvVSq2DBg3SwoULb3QtAAAAwE1RrekNZWVlev3117V8+XIFBQVVeJDtjTfeuCHFAQAAADdCtU
JvQUGBQkJCJElbt2612vbLh9oAAACAW0G1Qu8XX3xxo+sAAAAAbhqeRAMAAIDpEXoBAABgeoReAAAAmB6hF
wAAAKZH6AUAAIDpEXoBAABgeoReAAAAmF613tMLmNXWceFydXWt6TIAAMANxp1eAAAAmB6hFwAAAKZH6AUA
AIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAAB
geoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB
6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6
AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoB
AABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAA
AmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAK
ZH6AUAAIDpEXoBAABgerY1XQBwK2mbtFy1HJxqugwAAG66vSmRNV3Cb4o7vQAAADA9Qi8AAABMj9ALAAAA0
yP0AgAAwPQIvQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwPQIvQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwPQI
vQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwPQIvQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwPQIvQAAADA9Qi8
AAABMj9ALAAAA0yP0AgAAwPQIvQAAADA9Qi8AAABMz3Shd/Xq1bKxsdFPP/
102TZjx45VcHDwb1bTnaoq1+LXWrRoobfeeuum1QQAAO5MNRp6n3zySdnY2FRYvv/+
+5t63BEjRmjlypU3tM9LAc/NzU3nz5+32rZhwwbLuf26fZs2bVRWVmbVvn79+srMzLR8/
nUQ3Lx5s3r37q1GjRqpTp06atGihR5//HEdPXpUY8eOrXRMf7lI/
x37wYMHVziXhIQE2djY6Mknn7z+gammCxcuyMPDQykpKZVuHz9+vBo3bqzS0lJJP49n+/bt5eDgID8/
P6vxAwAAqPE7vb169dKhQ4esFh8fn5t6TBcXFzVo0OCm9F23bl0tXrzYat3s2bPl7e1dafvdu3frvffeq3L
/
x44dU48ePeTu7q7ly5ersLBQGRkZ8vT01JkzZzRixAirsWzWrJleffVVq3WXeHl5acGCBTp37pxl3fnz5zV
//
vzL1vtbsbe318CBA5WRkVFhm2EYyszMVExMjOzs7LRnzx5FRkYqLCxM+fn5SkxMVHx8vJYvX14DlQMAgFtR
jYdeBwcHNWnSxGqZMmWKAgMD5ezsLC8vLz377LM6ffq0ZZ99+/
YpKipKbm5ucnZ2Vps2bbR06VKrfvPy8tShQwc5OTmpS5cu2r59u2Xbr6c3lJeX69VXX1WzZs3k4OCg4OBgL
Vu2zLJ97969srGxUXZ2tsLCwuTk5KR27dpp3bp1Fc5n0KBBmjNnjuXzuXPntGDBAg0aNKjS83/
uueeUlJSkkpKSKo1Xbm6uioqKNGvWLIWEhMjHx0dhYWF688035ePjIxcXF6uxrF27turWrWu17pL27dvLy8
tL2dnZlnXZ2dny9vZWSEiI1XFLSko0dOhQy93l+++/
Xxs2bLBqs3TpUvn7+8vR0VFhYWHau3dvhfpzcnL0wAMPyNHRUV5eXho6dKjOnDlT6bnGxcVpx44dysnJsVq
/
Zs0a7d69W3FxcZKk9PR0+fj4KC0tTa1bt9aQIUP06KOP6s0337zsOJaUlKi4uNhqAQAA5lXjobcytWrV0tS
pU/
Xdd99p7ty5WrVqlUaNGmXZnpCQoJKSEn355ZcqKChQamqqXFxcrPp46aWXlJaWpo0bN8rW1laxsbGXPd6UK
VOUlpamyZMna8uWLQoPD1fv3r21c+fOCn2OGDFC+fn58vf3V//+/XXx4kWrNtHR0frqq6+0f/
9+SVJWVpZatGih9u3bV3rsxMREXbx4UdOmTavS2DRp0kQXL17U4sWLZRhGlfa5ktjYWKu7qXPmzNFTTz1Vo
d2oUaOUlZWluXPnatOmTfLz81N4eLhOnDghSTpw4ID69OmjqKgo5efnKz4+XmPGjLHqY9euXerVq5ceeeQR
bdmyRQsXLlROTo6GDBlSaW2BgYG69957rb5ESFJGRoa6dOmiVq1aSZLWrVunnj17WrUJDw+v9EvJJcnJyap
Xr55l8fLyusIoAQCA212Nh94lS5bIxcXFsvTt21eJiYkKCwtTixYt1L17d02YMEGLFi2y7LN//
36FhoYqMDBQd999tx588EF17drVqt+JEyeqW7duCggI0JgxY7R27doKc20vmTx5skaPHq1+/
frpnnvuUWpqqoKDgys8UDVixAhFRkbK399f48aN0759+yrMP27UqJEiIiIsc0rnzJlzxcDt5OSkpKQkJScn
q6io6Krj1alTJ7344ot64okn5OHhoYiICE2aNElHjhy56r6VGThwoHJycrRv3z7t27dPubm5GjhwoFWbM2f
OaMaMGZo0aZIiIiIUEBCgmTNnytHRUbNnz5YkzZgxQ76+vkpLS9M999yjAQMGVJgTnJycrAEDBigxMVEtW7
ZUly5dNHXqVL333nuXvTZxcXH64IMPLHf6T506pQ8//
NBqTA8fPqzGjRtb7de4cWMVFxdbTd34pRdeeEFFRUWW5cCBA9c0bgAA4PZS46H30jzMS8vUqVP1+eefq0eP
HmratKnq1q2r6OhoHT9+XGfPnpUkDR06VBMmTFBoaKiSkpK0ZcuWCv0GBQVZfr7rrrskSUePHq3Qrri4WAc
PHlRoaKjV+tDQUBUWFlarz9jYWGVmZmr37t1at26dBgwYcMUxiIuLU4MGDZSamnrFdpdMnDhRhw8fVnp6ut
q0aaP09HS1atVKBQUFVdr/
lxo2bKjIyEhlZmYqIyNDkZGR8vDwsGqza9culZaWWo2RnZ2dOnbsaBmjwsJC3XfffVb7de7c2erz5s2blZm
ZafUlJzw8XOXl5dqzZ0+l9fXv319lZWWWLz0LFy5UrVq19Pjjj1/
zuf6Sg4ODXF1drRYAAGBeNR56nZ2d5efnZ1lKSkr04IMPKigoSFlZWcrLy9Pbb78t6ecn+iUpPj5eu3fvVn
R0tAoKCtShQ4cK0wPs7OwsP196Y0F5efl11VrVPiMiInTu3DnFxcUpKirqqg/
N2draauLEiZoyZYoOHjxYpVoaNGigvn37avLkySosLJSnp6cmT558DWfzX5dC+ty5c694V/p6nT59Wk8//
bTVl5zNmzdr586d8vX1rXQfV1dXPfroo5YpGBkZGXrsscesprM0adKkwp3uI0eOyNXVVY6OjjftfAAAwO2j
xkPvr+Xl5am8vFxpaWnq1KmT/
P39Kw2CXl5eGjx4sLKzszV8+HDNnDmzWsdzdXWVp6encnNzrdbn5uYqICCgWn3a2toqJiZGq1evrnKI7Nu3
r9q0aaNx48Zd8/
Hs7e3l6+t72QfCrqZXr166cOGCSktLFR4eXmG7r6+v7O3trcaotLRUGzZssIxR69at9c0331jtt379eqvP7
du317Zt26y+5Fxa7O3tL1tfXFyccnJytGTJEq1du9byANslnTt3rvAKuhUrVlS40wwAAO5ctjVdwK/
5+fmptLRU06ZNU1RUlHJzc5Wenm7VJjExUREREfL399fJkyf1xRdfqHXr1tU+5siRI5WUlCRfX18FBwcrIy
ND+fn5mjdvXrX7HD9+vEaOHHlNr0ZLSUmpNHT+0pIlS7RgwQL169dP/
v7+MgxDn376qZYuXVrp672qonbt2pZpCrVr166w3dnZWc8884xGjhwpd3d3eXt76/
XXX9fZs2ctAXTw4MFKS0vTyJEjFR8fr7y8vArvyh09erQ6deqkIUOGKD4+Xs7Oztq2bZtWrFih6dOnX7a+r
l27ys/
PTzExMWrVqpW6dOlitX3w4MGaPn26Ro0apdjYWK1atUqLFi3SZ599Vq3xAAAA5nPL3elt166d3njjDaWmpq
pt27aaN2+ekpOTrdqUlZUpISFBrVu3Vq9eveTv76933nmn2sccOnSohg0bpuHDhyswMFDLli3TJ598opYtW
1a7T3t7e3l4eFj9Qoqr6d69u7p3717hjRC/
FBAQICcnJw0fPlzBwcHq1KmTFi1apFmzZik6Orra9V5tXmtKSooeeeQRRUdHq3379vr+++
+1fPlyubm5SZK8vb2VlZWljz76SO3atVN6erpee+01qz6CgoK0Zs0a7dixQw888IBCQkL0t7/9TZ6enlesz
cbGRrGxsTp58mSld859fHz02WefacWKFWrXrp3S0tI0a9asq36BAAAAdw4b40a89wq4zRUXF//
86rLERarl4FTT5QAAcNPtTYms6RKu26W/
v4uKiq76UPotd6cXAAAAuNEIvQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwPQIvQAAADA9Qi8AAABMj9ALAAA
A0yP0AgAAwPQIvQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwPQIvQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwP
QIvQAAADA9Qi8AAABMj9ALAAAA0yP0AgAAwPQIvQAAADA9Qi8AAABMz7amCwBuJVvHhcvV1bWmywAAADcYd
3oBAABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoRe
AAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwA
AAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6AUAAIDpEXoBAABgeoReAAAAmB6hFwAAAKZH6AUAAI
DpEXoBAABgerY1XQBwKzAMQ5JUXFxcw5UAAICquvT39qW/
x6+E0AtIOn78uCTJy8urhisBAADX6tSpU6pXr94V2xB6AUnu7u6SpP3791/1fxrUrOLiYnl5eenAgQNydXW
t6XJwBVyr2wfX6vbBtbJmGIZOnTolT0/
Pq7Yl9AKSatX6eXp7vXr1+EPkNuHq6sq1uk1wrW4fXKvbB9fqv6p6s4oH2QAAAGB6hF4AAACYHqEXkOTg4K
CkpCQ5ODjUdCm4Cq7V7YNrdfvgWt0+uFbVZ2NU5R0PAAAAwG2MO70AAAAwPUIvAAAATI/
QCwAAANMj9AIAAMD0CL2ApLffflstWrRQnTp1dN999+mbb76p6ZLueMnJybr33ntVt25dNWrUSH/
605+0fft2qzbnz59XQkKCGjRoIBcXFz3yyCM6cuRIDVUMSUpJSZGNjY0SExMt67hOt44ff/
xRAwcOVIMGDeTo6KjAwEBt3LjRst0wDP3tb3/TXXfdJUdHR/
Xs2VM7d+6swYrvTGVlZXrllVfk4+MjR0dH+fr6avz48frluwe4VteO0Is73sKFCzVs2DAlJSVp06ZNateun
cLDw3X06NGaLu2OtmbNGiUkJGj9+vVasWKFSktL9Yc//EFnzpyxtPnrX/+qTz/
9VB988IHWrFmjgwcPqk+fPjVY9Z1tw4YN+vvf/66goCCr9VynW8PJkycVGhoqOzs7/fOf/
9S2bduUlpYmNzc3S5vXX39dU6dOVXp6ur7+
+ms5OzsrPDxc58+fr8HK7zypqamaMWOGpk+frsLCQqWmpur111/
XtGnTLG24VtVgAHe4jh07GgkJCZbPZWVlhqenp5GcnFyDVeHXjh49akgy1qxZYxiGYfz000+GnZ2d8cEHH1
jaFBYWGpKMdevW1VSZd6xTp04ZLVu2NFasWGF069bNeP755w3D4DrdSkaPHm3cf//
9l91eXl5uNGnSxJg0aZJl3U8//WQ4ODgY77///
m9RIv5fZGSkERsba7WuT58+xoABAwzD4FpVF3d6cUe7cOGC8vLy1LNnT8u6WrVqqWfPnlq3bl0NVoZfKyoq
kiS5u7tLkvLy8lRaWmp17Vq1aiVvb2+uXQ1ISEhQZGSk1fWQuE63kk8+
+UQdOnRQ37591ahRI4WEhGjmzJmW7Xv27NHhw4etrlW9evV03333ca1+Y126dNHKlSu1Y8cOSdLmzZuVk5O
jiIgISVyr6rKt6QKAmvSf//xHZWVlaty4sdX6xo0b69///
ncNVYVfKy8vV2JiokJDQ9W2bVtJ0uHDh2Vvb6/69etbtW3cuLEOHz5cA1XeuRYsWKBNmzZpw4YNFbZxnW4d
u3fv1owZMzRs2DC9+OKL2rBhg4YOHSp7e3sNGjTIcj0q+/OQa/
XbGjNmjIqLi9WqVSvVrl1bZWVlmjhxogYMGCBJXKtqIvQCuOUlJCRo69atysnJqelS8CsHDhzQ888/
rxUrVqhOnTo1XQ6uoLy8XB06dNBrr70mSQoJCdHWrVuVnp6uQYMG1XB1+KVFixZp3rx5mj9/vtq0aaP8/
HwlJibK09OTa3UdmN6AO5qHh4dq165d4UnyI0eOqEmTJjVUFX5pyJAhWrJkib744gs1a9bMsr5Jkya6cOGC
fvrpJ6v2XLvfVl5eno4ePar27dvL1tZWtra2WrNmjaZOnSpbW1s1btyY63SLuOuuuxQQEGC1rnXr1tq/
f78kWa4Hfx7WvJEjR2rMmDHq16+fAgMDFR0drb/
+9a9KTk6WxLWqLkIv7mj29vb63e9+p5UrV1rWlZeXa+XKlercuXMNVgbDMDRkyBAtXrxYq1atko+Pj9X23/
3ud7Kzs7O6dtu3b9f+/fu5dr+hHj16qKCgQPn5+ZalQ4cOGjBggOVnrtOtITQ0tMJr/
3bs2KHmzZtLknx8fNSkSROra1VcXKyvv/6aa/
UbO3v2rGrVso5otWvXVnl5uSSuVbXV9JN0QE1bsGCB4eDgYGRmZhrbtm0z/
vKXvxj169c3Dh8+XNOl3dGeeeYZo169esbq1auNQ4cOWZazZ89a2gwePNjw9vY2Vq1aZWzcuNHo3Lmz0blz
5xqsGoZhWL29wTC4TreKb775xrC1tTUmTpxo7Ny505g3b57h5ORk/OMf/7C0SUlJMerXr298/
PHHxpYtW4yHHnrI8PHxMc6dO1eDld95Bg0aZDRt2tRYsmSJsWfPHiM7O9vw8PAwRo0aZWnDtbp2hF7AMIxp
06YZ3t7ehr29vdGxY0dj/
fr1NV3SHU9SpUtGRoalzblz54xnn33WcHNzM5ycnIyHH37YOHToUM0VDcMwKoZertOt49NPPzXatm1rODg4
GK1atTLeffddq+3l5eXGK6+8YjRu3NhwcHAwevToYWzfvr2Gqr1zFRcXG88//7zh7e1t1KlTx7j77ruNl15
6ySgpKbG04VpdOxvD+MWv9wAAAABMiDm9AAAAMD1CLwAAAEyP0AsAAADTI/
QCAADA9Ai9AAAAMD1CLwAAAEyP0AsAAADTI/
QCAADA9Ai9AADcAF27dtX8+fOvq49OnTopKyvrBlUE4JcIvQAAXKdPPvlER44cUb9+/
Szrhg0bJnd3d3l5eWnevHlW7T/44ANFRUVV6Ofll1/WmDFjVF5eftNrBu40/
BpiAIAplJaWys7OrkaO3bNnT/Xs2VNjxoyRJH366af685//
rCVLlmjnzp2KjY3VgQMH5OHhoaKiIt177736/
PPP5e3tbdVPWVmZmjZtqtmzZysyMrImTgUwLe70AgCu2bJly3T//ferfv36atCggR588EHt2rXLqs0PP/
yg/v37y93dXc7OzurQoYO+/vpry/ZPP/1U9957r+rUqSMPDw89/PDDlm02Njb66KOPrPqrX7+
+MjMzJUl79+6VjY2NFi5cqG7duqlOnTqaN2+ejh8/rv79+6tp06ZycnJSYGCg3n//
fat+ysvL9frrr8vPz08ODg7y9vbWxIkTJUndu3fXkCFDrNofO3ZM9vb2WrlyZaVjcezYMa1atcrqzm1hYaH
+53/+Rx06dFD//
v3l6uqqPXv2SJJGjRqlZ555pkLglaTatWvrj3/8oxYsWFDpsQBUH6EXAHDNzpw5o2HDhmnjxo1auXKlatWq
pYcfftjyz/KnT59Wt27d9OOPP+qTTz7R5s2bNWrUKMv2zz77TA8//LD++Mc/
6ttvv9XKlSvVsWPHa65jzJgxev7551VYWKjw8HCdP39ev/vd7/TZZ59p69at+stf/
qLo6Gh98803ln1eeOEFpaSk6JVXXtG2bds0f/58NW7cWJIUHx+v+fPnq6SkxNL+H//
4h5o2baru3btXWkNOTo6cnJzUunVry7p27dpp48aNOnnypPLy8nTu3Dn5+fkpJydHmzZt0tChQy97Th07dt
RXX311zWMB4CoMAACu07FjxwxJRkFBgWEYhvH3v//
dqFu3rnH8+PFK23fu3NkYMGDAZfuTZCxevNhqXb169YyMjAzDMAxjz549hiTjrbfeumptkZGRxvDhww3DMI
zi4mLDwcHBmDlzZqVtz507Z7i5uRkLFy60rAsKCjLGjh172f7ffPNN4+67766wPikpyfD19TXatm1rZGdnG
yUlJUbbtm2NjRs3GtOmTTP8/f2NLl26GFu3brXa7+OPPzZq1apllJWVXfXcAFQdd3oBANds586d6t+/v+6+
+265urqqRYsWkqT9+/dLkvLz8xUSEiJ3d/dK98/
Pz1ePHj2uu44OHTpYfS4rK9P48eMVGBgod3d3ubi4aPny5Za6CgsLVVJSctlj16lTR9HR0ZozZ44kadOmTd
q6dauefPLJy9Zw7tw51alTp8L6sWPH6vvvv1dBQYEefvhhJScnq2fPnrKzs9OECROUk5Oj+Ph4xcTEWO3n6
Oio8vJyq7vNAK6fbU0XAAC4/
URFRal58+aaOXOmPD09VV5errZt2+rChQuSfg5uV3K17TY2NjJ+9Zx1aWlphXbOzs5WnydNmqQpU6borbfe
UmBgoJydnZWYmFjluqSfpzgEBwfrhx9+UEZGhrp3767mzZtftr2Hh4dOnjx5xT7//e9/6x//+Ie+/
fZbzZkzR127dlXDhg312GOPKTY2VqdOnVLdunUlSSdOnJCzs3OVagVQddzpBQBck+PHj2v79u16+eWX1aNH
D7Vu3bpC6AsKClJ+fr5OnD
hRaR9BQUGXfTBMkho2bKhDhw5ZPu/
cuVNnz569am25ubl66KGHNHDgQLVr10533323duzYYdnesmVLOTo6XvHYgYGB6tChg2bOnKn58+crNjb2is
cMCQnR4cOHLxt8DcPQ008/rTfeeEMuLi4qKyuzBPhL/
y0rK7O037p1q0JCQq56rgCuDaEXAHBN3Nzc1KBBA7377rv6/vvvtWrVKg0bNsyqTf/+/dWkSRP96U9/
Um5urnbv3q2srCytW7dOkpSUlKT3339fSUlJKiwsVEFBgVJTUy37d+/
eXdOnT9e3336rjRs3avDgwVV6HVnLli21YsUKrV27VoWFhXr66ad15MgRy/
Y6depo9OjRGjVqlN577z3t2rVL69ev1+zZs636iY+PV0pKigzDsHqrRGVCQkLk4eGh3NzcSrfPmjVLDRs2t
LzdITQ0VKtWrdL69ev15ptvKiAgQPXr17e0/+qrr/SHP/
zhqucK4BrV8JxiAMBtaMWKFUbr1q0NBwcHIygoyFi9enWFh8/27t1rPPLII4arq6vh5ORkdOjQwfj6668t2
7Oysozg4GDD3t7e8PDwMPr06WPZ9uOPPxp/+MMfDGdnZ6Nly5bG0qVLK32Q7dtvv7Wq6/
jx48ZDDz1kuLi4GI0aNTJefvllIyYmxnjooYcsbcrKyowJEyYYzZs3N+zs7Axvb2/
jtddes+rn1KlThpOTk/Hss89WaTxGjRpl9OvXr8L6w4cPG82bNzd+/PFHq/Xjxo0z3N3djVatWlmNyQ8//
GDY2dkZBw4cqNJxAVQdv5wCAIBf2bt3r3x9fbVhwwa1b9/+qu0PHz6sNm3aaNOmTVec/3s1o0eP1smTJ/
Xuu+9Wuw8AlWN6AwAA/6+0tFSHDx/Wyy+/rE6dOlUp8EpSkyZNNHv2bMtbIqqrUaNGGj9+/
HX1AaBy3OkFAOD/
rV69WmFhYfL399eHH36owMDAmi4JwA1C6AUAAIDpMb0BAAAApkfoBQAAgOkRegEAAGB6hF4AAACYHqEXAAA
ApkfoBQAAgOkRegEAAGB6hF4AAACY3v8BNb6X91LnJdoAAAAASUVORK5CYII=",
"text/plain": [
"<Figure size 640x480 with 1 Axes>"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# Visualize our model results\n",
"compare_results.set_index(\"model_name\")[\"model_acc\"].plot(kind=\"barh\")\
n",
"plt.xlabel(\"accuracy (%)\")\n",
"plt.ylabel(\"model\");"
]
},
{
"cell_type": "markdown",
"id": "0ba50d51-adb3-4e49-9b9a-85173e747352",
"metadata": {
"id": "0ba50d51-adb3-4e49-9b9a-85173e747352"
},
 "source": [
"## 9. Make and evaluate random predictions with best model\n",
"\n",
"Alright, we've compared our models to each other, let's further evaluate our
best performing model, `model_2`.\n",
"\n",
"To do so, let's create a function `make_predictions()` where we can pass the
model and some data for it to predict on."
]
},
{
"cell_type": "code",
"execution_count": 46,
"id": "d1d5d3e7-9601-4141-8bd7-9abbd016bf6c",
"metadata": {
"id": "d1d5d3e7-9601-4141-8bd7-9abbd016bf6c"
},
"outputs": [],
"source": [
"def make_predictions(model: torch.nn.Module, data: list, device: torch.device
= device):\n",
" pred_probs = []\n",
" model.eval()\n",
" with torch.inference_mode():\n",
" for sample in data:\n",
" # Prepare sample\n",
" sample = torch.unsqueeze(sample, dim=0).to(device) # Add an extra
dimension and send sample to device\n",
"\n",
" # Forward pass (model outputs raw logit)\n",
" pred_logit = model(sample)\n",
"\n",
" # Get prediction probability (logit -> prediction probability)\n",
" pred_prob = torch.softmax(pred_logit.squeeze(), dim=0) # note:
perform softmax on the \"logits\" dimension, not \"batch\" dimension (in this case
we have a batch size of 1, so can perform on dim=0)\n",
"\n",
" # Get pred_prob off GPU for further calculations\n",
" pred_probs.append(pred_prob.cpu())\n",
" \n",
" # Stack the pred_probs to turn list into a tensor\n",
" return torch.stack(pred_probs)"
]
},
{
"cell_type": "code",
"execution_count": 47,
"id": "420c7461-eaa9-4459-9e68-53574c758765",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "420c7461-eaa9-4459-9e68-53574c758765",
"outputId": "f3dd6437-4f0f-4bc2-f9e6-d0969df63a52"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
 "text": [
"Test sample image shape: torch.Size([1, 28, 28])\n",
"Test sample label: 5 (Sandal)\n"
]
}
],
"source": [
"import random\n",
"random.seed(42)\n",
"test_samples = []\n",
"test_labels = []\n",
"for sample, label in random.sample(list(test_data), k=9):\n",
" test_samples.append(sample)\n",
" test_labels.append(label)\n",
"\n",
"# View the first test sample shape and label\n",
"print(f\"Test sample image shape: {test_samples[0].shape}\\nTest sample label:
{test_labels[0]} ({class_names[test_labels[0]]})\")"
]
},
{
"cell_type": "code",
"execution_count": 48,
"id": "1DYqA0r4SkrV",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "1DYqA0r4SkrV",
"outputId": "97bc573d-b39b-4eb2-caad-0379257b555e"
},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[2.4012e-07, 6.5406e-08, 4.8069e-08, 2.1070e-07, 1.4175e-07,
9.9992e-01,\n",
" 2.1711e-07, 1.6177e-05, 3.7849e-05, 2.7548e-05],\n",
" [1.5646e-02, 8.9752e-01, 3.6928e-04, 6.7402e-02, 1.2920e-02,
4.9539e-05,\n",
" 5.6485e-03, 1.9456e-04, 2.0808e-04, 3.7861e-05]])"
]
},
"execution_count": 48,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Make predictions on test samples with model 2\n",
"pred_probs= make_predictions(model=model_2, \n",
" data=test_samples)\n",
"\n",
"# View first two prediction probabilities list\n",
"pred_probs[:2]"
]
},
{
"cell_type": "markdown",
 "id": "e9f40dd9-7987-42a9-84cc-65dc912a6345",
"metadata": {
"id": "e9f40dd9-7987-42a9-84cc-65dc912a6345"
},
"source": [
"And now we can use our `make_predictions()` function to predict on
`test_samples`."
]
},
{
"cell_type": "code",
"execution_count": 49,
"id": "79de2ac1-7d4b-4f81-ae8a-90099bca2a3d",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "79de2ac1-7d4b-4f81-ae8a-90099bca2a3d",
"outputId": "918b07bc-4545-4401-84d5-8796ff5acf4c"
},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[2.4012e-07, 6.5406e-08, 4.8069e-08, 2.1070e-07, 1.4175e-07,
9.9992e-01,\n",
" 2.1711e-07, 1.6177e-05, 3.7849e-05, 2.7548e-05],\n",
" [1.5646e-02, 8.9752e-01, 3.6928e-04, 6.7402e-02, 1.2920e-02,
4.9539e-05,\n",
" 5.6485e-03, 1.9456e-04, 2.0808e-04, 3.7861e-05]])"
]
},
"execution_count": 49,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Make predictions on test samples with model 2\n",
"pred_probs= make_predictions(model=model_2, \n",
" data=test_samples)\n",
"\n",
"# View first two prediction probabilities list\n",
"pred_probs[:2]"
]
},
{
"cell_type": "markdown",
"id": "22d3c080-4eb6-4b5d-a5c4-2319e78228af",
"metadata": {
"id": "22d3c080-4eb6-4b5d-a5c4-2319e78228af"
},
"source": [
"Excellent!\n",
"\n",
"And now we can go from prediction probabilities to prediction labels by taking
the `torch.argmax()` of the output of the `torch.softmax()` activation function."
]
},
 {
"cell_type": "code",
"execution_count": 50,
"id": "f9d97bcc-4310-4851-a1f8-6bcd757e9b26",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "f9d97bcc-4310-4851-a1f8-6bcd757e9b26",
"outputId": "9d0f0bf9-a641-45e7-af77-6621fd1cfcc4"
},
"outputs": [
{
"data": {
"text/plain": [
"tensor([5, 1, 7, 4, 3, 0, 4, 7, 1])"
]
},
"execution_count": 50,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Turn the prediction probabilities into prediction labels by taking the
argmax()\n",
"pred_classes = pred_probs.argmax(dim=1)\n",
"pred_classes"
]
},
{
"cell_type": "code",
"execution_count": 51,
"id": "1141af97-0990-4920-83d4-c13cca3f9abc",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "1141af97-0990-4920-83d4-c13cca3f9abc",
"outputId": "c69cddd4-bbe9-495e-d477-6ea0a6c7d8de"
},
"outputs": [
{
"data": {
"text/plain": [
"([5, 1, 7, 4, 3, 0, 4, 7, 1], tensor([5, 1, 7, 4, 3, 0, 4, 7, 1]))"
]
},
"execution_count": 51,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Are our predictions in the same form as our test labels? \n",
"test_labels, pred_classes"
]
},
{
 "cell_type": "markdown",
"id": "4ea04387-c9ad-424f-8297-defd7b685683",
"metadata": {
"id": "4ea04387-c9ad-424f-8297-defd7b685683"
},
"source": [
"Now our predicted classes are in the same format as our test labels, we can
compare.\n",
"\n",
"Since we're dealing with image data, let's stay true to the data explorer's
motto. \n",
"\n",
"\"Visualize, visualize, visualize!\""
]
},
{
"cell_type": "code",
"execution_count": 52,
"id": "679cb5f7-bb66-42dd-a4d6-400b27b7c019",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 749
},
"id": "679cb5f7-bb66-42dd-a4d6-400b27b7c019",
"outputId": "3aae0abe-9c19-4054-d8db-7e00403666aa"
},
"outputs": [
{
"data": {
"image/png":
"iVBORw0KGgoAAAANSUhEUgAAAt0AAALcCAYAAAA7awxXAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIH
ZlcnNpb24zLjcuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/
bCgiHAAAACXBIWXMAAA9hAAAPYQGoP6dpAACNV0lEQVR4nO3dd3gVZfr/
8U8S0qv0TmhSLHSsCGIBEUXEzoqIFfva9bsqWNHFte2Kq6xgWUV3RVDqIhIEFEGRIgSkBRBDCwQICem/P/
LjaIDnnnDCQALv13V5XZL7zDNz5swzc2fOOZ+EFBcXFwsAAACAb0KP9gYAAAAAxzqabgAAAMBnNN0AAACAz
2i6AQAAAJ/
RdAMAAAA+o+kGAAAAfEbTDQAAAPiMphsAAADwGU03AAAA4DOabgAAAMBnla7pHjhuoC4bc9nR3owyO9TtTc
tMU8jQEC3ctND5mNELR6vb6G7l3rbDYfTC0UoalnS0N6NchqQMUdu32h7SMiFDQzRu+Thftud4Udnmst9S0
lKU/Gry0d4MSSXbEjI0RJl7M4/2pqCCqszztyzX2cOF67W/
Ktu1uMrhGGTguIF6b9F7kqTw0HA1TGyoAW0G6PEuj6tK6GFZRbls3bNVT854UhNXTtTmPZt1QtQJalO7jZ4
850md1fCso715h1VaZpoav9bYfMyoPqM0sO3AQx47+dVk3Xf6fbrv9PuC2zhDdn62npn5jD5d9qk27tqo+M
h4ta7RWveffr/6tOxz2NeHg6vIczlkaIhZf6rrUxrSbciR2ZijwK/
n3210N7Wt3Vav9nw1uA0zxp25bqaz3rVRV6UMTDms6zzeVeT5Kx1f1+Ky4Hp9/
Dlss7Bns54a1WeUcgtyNWnlJN056U6Fh4brsS6PHfDYvMI8RYRFHK5Ve+r3aT/
lFebpvcveU5MTmmjzns2avma6MnIyjtg2HCkNEhoo/YH0wL+HfztcU1ZN0VcDvgr8LDEyMfD/
hUWFCgkJUWjI0X3T4/YJt+v7jd/rjYveUOsarZWRnaFvN3x7TL5GFV1Fnct/
PK4/+fkTPZnypFbctSLws7iIuMD/FxcXq7C4sEI0GvsLdp9Vtuc/
9uqxyivMkyRt2LlBnUd21lfXf6WTap4kSQfsg/zCfIWHhR/
x7fRypK9X5VVR5690bF+Lgzl+uV5XPH7PicP2ykWGRap2XG01SmqkwZ0G6/
wm5+uLX76Q9PvbUM9985zqvlxXLf7eQlLJifiq/1ylpGFJqvpiVfUZ00dpmWmBMQuLCnX/
1PuVNCxJ1V6qpoenPaxiFR/
SdmXuzdSs9bP04vkv6tzG56pRUiN1rtdZj3V5TJe2uDTwuL999zedMuIUxT4fqwavNNAdE+9QVl5WoL7vbZ
mpq6aq1T9aKe75OPX8sKfSd/8+YcqyvVNWTdHZ754deEzvj3pr9fbVh/
ScLGGhYaodVzvwX1xEnKqEVgn8e8qqKarzch19seILtf5Ha0U+G6n1O9er2+huum/
KfaXGumzMZRo4bqCkkrtW63au05+n/
lkhQ0MOuOtm7Zey+GLFF3r87MfVq3kvJSclq0PdDrr7tLs1qN2gwGM+WPSBOr7dUfEvxKv28Nq67rPrtGXP
lkB931vi09dMV8e3OyrmuRid+a8ztWLbilLrGjZ7mGoNr6X4F+J10/ibtLdgb6n6/
I3zdcEHF6j6S9WVOCxRXUd31YL0BYf0fCqzijqX/3hcJ0YlKkQhgX8v37Zc8S/Ea/
LKyerwdgdFPhup2etnK7cgV/dMvkc1/1pTUc9G6ex3z9b8jfMDYx7s7dZxy8eVOr4XbVqkc987V/
EvxCvhhQR1eLuDfvjth0B99vrZ6jKqi6Kfi1aDVxronsn3aE/enkA9+dVkPTPzGQ34fIASXkjQrV/
eekjPuzzP/2AfAbhvyn2Bt7sHjhuometm6rXvXwvM6z++bj/
+9qM5lyxVo6sGtq9GbA1JUrWYaoGfVXupmkbMH6FLP75Usc/
H6rlZz0mSRswfoaavN1XEMxFq8fcW+mDRB4ExD/axgMy9mQoZGqKUtBRJ0o6cHeo/tr9q/
LWGop+LVvM3mmvUT6MCj/c6Vl3HeGVRUedvWa/
FIUNDNHLBSPX9pK9inotR8zea64sVX5Qa6+ctP+uif1+kuOfjVGt4LV3/+fXalr0tUD/
U62xhUaEGjR+kln9vqfU710uSxi8fr/b/bK+oZ6PU5LUmGpoyVAVFBaW282DH76E4lq/Xya8m6/
lZz2vQ+EGKfyFeDV9pqLd/
fLvUOF7HXTDX4qdmPKU6L9fR4s2LJR2583NZ+fbrUnR4dOAuhyRNXztdKzJWaNr10zTh2gnKL8xXjw97KD4
iXrNunKU5g+YoLqLkANi33MvfvazRC0fr3T7vavaNs7U9Z7s+T/
281HpGLxxtvu0aFxGnuIg4jVs+TrkFuc7HhYaE6vWer2vpHUv13mXv6eu1X+vhaQ+Xekx2fraGfzdcH/
T9QN/c+I3W71yvB6c9GKiXZXv35O3R/Wfcrx9u/UHTB0xXaEio+n7SV0XFRd479TDJzs/
Wi3Ne1MhLR2rpHUtVM7am5zJjrx6r+gn19XS3p5X+QHqp38699su+ZviPk2l/
teNqa9KqSdqdu9v5mPyifD1z7jNadPsijbtmnNIy0wInmT/6v6//Ty9f+LJ+uPUHVQmtokFf/
H4i+HTppxqSMkTPd39eP9zyg+rE19Gb898stfzuvN26oc0Nmj1otubeNFfNqzZXr3/3MrftWFZR5nJZPDr9
UQ07b5hS70zVqbVO1cPTHtZnqZ/pvcve04LbFqhZ1Wbq8WEPbc/ZXuYx+4/tr/oJ9TX/
lvn68dYf9ehZjyo8tOSO1urtq9Xzw57q16qfFt++WJ9c8Ylmr5+tuybfVWqM4d8NV5tabfTTbT/
piXOeKNdztOz//L281vM1nVH/DN3S/
pbAvG6Q0CBQt+bSvgZ4X7MbjCEzh6hvy75aMniJBrUbpM9TP9e9U+7VA2c8oJ/v+Fm3dbhNN46/
UTPWzijzmE/
MeELLti7T5P6TlXpnqkZcPELVY6pLUpmOVenAY7wyqyjzt6zXYkkaOnOormp9lRYPXqxezXqp/
9j+gTmbuTdT3d/rrna12+mHW3/QlP5TtDlrs676z1WB5Q/lOptbkKsr/
3OlFm5aqFk3zlLDxIaatW6WBowboHtPu1fL7lymf/b+p0YvGq3nvindWO9//
Pqhsl6vpZLjpmPdjvrptp90R6c7NHji4MAv7mU57g7lWlxcXKy7J92t9xe/
r1k3ztKptU6tcOdn6TB+vGSf4uJiTV87XVNXTdXdne8O/
Dw2PFYjLx0ZuG3/4eIPVVRcpJGXjlRISMlEHdVnlJKGJSklLUUXNr1Qr859VY+d/
Zgub3W5JOmt3m9p6uqppdaXGJmoFtXcdyKqhFbR6D6jdcuXt+itH99S+zrt1bVRV11z8jWlLkp//
NxTclKynu3+rG6fcLvevPj3hiy/KF9vXfyWmlZtKkm6q/
Ndenrm04F6Wba3X+t+pf79bp93VeOvNbRs6zKdXPNk5/
M4nPKL8vVmrzfVpnabMi9TNbqqwkLCFB8Zr9pxtQ8Yz9ovMeExalGtRaBROZi3L3lb/
cf2V7WXqqlN7TY6u8HZuqL1FaU+5/
fHk1qTE5ro9YteV6d3OikrL6vUW+vPdX9OXZO7SpIePftRXfzRxdpbsFdRVaL06txXdVO7m3RT+5skSc92f
1Zfrfmq1N3u7o27H7BtScOSNHPdTPU+sXeZ91llV9Hmclk83e1pXdD0AkklF94RP4zQ6MtG66LmF0mS3rnk
HU1bM03/WvAvPXTWQ2Uac/3O9XrozIfUsnpLSVLzas0DtRdmv6D+p/
QPnD+aV2uu1y96XV1Hd9WIi0coqkqUpJJj6oEzHyjXcyuLPz7/
skiMSlREWIRiwmMOmNeSPZfCQ8PVoloLxYTHBL291518nW5sd2Pg39d+dq0Gth2oOzrdIUm6/4z7NffXuRr
+3XCd2/jcMo25fud6tavdTh3rdpRUcj7f55Oln3geq9KBx3hlVNHmb1mvxZI0sM1AXXvKtZKk5897Xq/
Pe13zNs5Tz2Y99fd5f1e7Ou30/HnPBx7/bp931eCVBvol4xedWO3EMl9ns/
KydPFHFyu3MFczbpihxKiSj3IMnTlUj571qG5oe4OkkuvNM+c+o4enPaynuj0VWH7/49cPlfV6LUm9mvcKz
OVHznpEr8x9RTPSZqhF9RZlmotlvRYXFBXoT5//ST+l/
6TZN85WvYR6kire+Vk6jE33hF8mKO75OOUX5auouEjXnXJdqS/1nFLrlFInsEWbFmnV9lWKfyG+1Dh7C/
Zq9fbV2llvp9Kz0nVa/dN+39jQKupYt6OKi39/
W6tvq77q26qvuW39WvfTxSderFnrZmnur3M1edVkvTTnJY28dGTgCwpfrflKL8x+Qcu3Ldeu3F0qKCrQ3oK
9ys7PDlxUYsJjAgeqJNWJqxP4eMPOvWXb3pUZK/VkypP6/tfvtS17W+A37/
U71x+xpjsiLKJMd8HKytovktS5Xmctv2u5OcY5jc7RmnvWaO6vc/
Xthm81fe10vTbqNQ3tNlRPdC35zfPH337UkJlDtGjTIu3Yu6PUvmtdo3VgrD8+tzpxdSRJW/
ZsUcPEhkrdlqrbO95eat1n1D9DM9J+v5O2OWuz/vL1X5SyLkVb9mxRYVGhsvOzA287Husq8lz2sq/
RkqTVO1YrvyhfZzX4/
UIQHhauzvU6K3VbapnHvP+M+3Xzlzfrg8Uf6Pwm5+vK1lcGjvdFmxdp8ebF+veSfwceX6xiFRUXae2OtWpV
o1XJdtXpeNCxD7c/Pv/DwZpL9RLqec5rL/
tvb+rWVN3avvTbu2c1OEuvff9amccc3HGw+n3aTwvSF+jCphfqspaX6cwGZ0ryPlb1/09j+x/
jlUlFnr9luRZLpY+72IhYJUQmBK4pizYv0oy1MxT3fNz+w2v19tU6sdqJZb7OXvvZtaqfUF9fD/ha0eHRv+
+TzYs0Z8OcUh8ZKSwuPKAnONzz7WAq6/Vakk6t+ft2h4SUfBwu8DqWYS6W9Vr856l/
VmRYpObePDfwrpZU8c7P0mFsus9tfK5GXDxCEWERqhtf94Av8MSGx5b6d1ZeljrU7aB/X/
5v7a9GTI3DtVkBUVWidEHTC3RB0wv0RNcndPMXN+uplKc0sO1ApWWmqfdHvTW442A91/05VY2uqtnrZ+umL
25SXmFeYILt/
5tfSEjIIX+u7ZKPL1GjpEZ655J3VDe+roqKi3TyiJNLvf3nt+gq0YHfLPcJDQktdQKVSn4jLovDsV+kkoao
S6Mu6tKoix45+xE9+82zenrm03rk7EcCb0X1aNZD/77836oRW0Prd65Xjw97HLDv/
vhlln3P81A+vnPDuBuUkZOh13q+pkaJjRRZJVJn/OuMI/oaHU0VfS5bYiNivR/0B6EhoQccq/mFpY/
7Id2G6LpTrtPEXyZq8qrJeirlKY3pN0Z9W/VVVl6Wbutwm+457Z4Dxm6Y2DDo7QrW/uspy/
OzlHcueQnm9ZJU6ly1//O5qPlFWnffOk1aOUnT1kzTee+fpzs73anhFw4v87G6/zFemVT0+Wtdi/fZ/
wuJIQoJHHdZeVm6pMUlevH8Fw8Ye98vhmW9zvZq1ksfLvlQ3/36Xam7qll5WRrabWjgzv7+27/
PkZjXlfF6ve+XOq/X0eu4K+u1+IImF+jjnz/W1FVT1f/U/
oGfV7Tzs3QYm+7Y8Fg1q9qszI9vX6e9Pln6iWrG1lRCZMJBH1Mnro6+//
V7ndPoHEklbyH8+NuPal+nfbm3t3WN1oFsxx9/+1FFxUV6ucfLgZP6p0s/
PaTxEqMSPbc3IztDKzJW6J1L3lGXRl0klXzIvyKoEVtD6VmlvxT685afdW7y72/
pRoRFqLCo8IhtU+sarQPvOKzMWKmMnAwNO2+YGiSWfOb0j19mK6tW1Vvp+1+/14A2AwI/
m7txbqnHzNkwR2/2elO9mveSVPJljz9+SedYV9nmskvTE5oqIixCczbMUaOkRpJKGrT5G+cH3m6sEVNDu3N
3a0/ensCJ92DZvSdWO1EnnnGi/
nzGn3XtZ9dq1MJR6tuqr9rXaa9lW5cd0v46kmrE1NDPW34u9bOFmxeWuvAe6XltaVWjleZsmBN4W18qmY/
73snadzFOz0pXO7WTdPDXq0ZsDd3Q9gbd0PYGdfmhix6a9pCGXzi8TMdqZVfZ5u8fr8Vl2t7a7fVZ6mdKTk
o+aDrPoVxnB3carJNrnqxLP75UE6+bGPgoVfs67bVi24qKO68r+PW6LO8SleW4K+u1+NIWl+qSEy/
RdWOvU1homK45+ZrAOira+fmo5c70P7W/
qsdUV58xfTRr3Syt3bFWKWkpumfyPfp116+SpHtPu1fD5gzTuOXjtHzbct0x8Y4D/
ljD56mfq+XfWzrXk5Gdoe7vddeHiz/U4s2LtXbHWv1n6X/
00pyX1KdFSZ5ks6rNlF+Urze+f0NrdqzRB4s+0Fs/
vHXIz8lre0+IPkHVoqvp7QVva9X2Vfp67de6f+r9h7weP3RP7q6JKydq4i8TtXzbcg2eOPiAfZ2clKxv1n+
jjbs2HlITOm/jPLX8e0tt3LXR+Zhuo7vpnz/8Uz/
+9qPSMtM0aeUkPT79cZ3b+FwlRCaoYWJDRYRF6I15Ja/RFyu+0DPfPHPIz/Pe0+7Vuwvf1aifRumXjF/
01IyntHTL0lKPaV61uT5Y/IFSt6bq+1+/V/+x/
RVdJdoxIo7UXD5UsRGxGtxxsB6a9pCmrJqiZVuX6ZYvb1F2frZualfymf7T6p+mmPAYPT79ca3evlofLflI
oxeNDoyRk5+juybdpZS0FK3LXKc56+do/sb5alW95G3JR856RN9u+FZ3TbpLCzct1MqMlRq/
fLzumnTXwTbpiOveuLt++O0Hvb/
ofa3MWKmnZjx1QBOenJSs7zd+r7TMtFJvxXvZuGujWv69peZtnHfYtvehMx/S6IWjNWL+CK3MWKm/
ffc3jU0dqwfPLPmiV3R4tE6vf7qGzR6m1K2pmpk2U3+Z8ZdSYzw540mNXz5eq7av0tItSzVh5YTA28hlOVa
PNxXpWlwWd3a+U9tztuvaz67V/I3ztXr7ak1dNVU3jr9RhUWFh3ydvfu0u/Vs92fV+
+Pegeb8yXOe1PuL39fQlKFaumWpUremaszPY/SXr//
iHOdIqujX67Ioy3F3KNfivq366oO+H+jG8Tfqv8v+K6linp+PWohrTHiMvrnxGz3y1SO6/
NPLtTt3t+ol1NN5jc8LvGgPnPmA0rPSdcO4GxQaEqpBbQepb6u+2rl3Z2Ccnbk7tSLDHWMVFxGn0+qdplfm
vqLV20s+49kgoYFuaX+LHu/yuCSpTe02+tuFf9OLc17UY9Mf0zmNztEL572gAeMGOMc9GK/
tDQ0J1Zgrxuieyffo5DdPVovqLfR6z9fV7b1uh7bzfDCo3SAt2rxIA8YNUJXQKvrz6X8u9VuzJD197tO6bc
Jtavp6U+UW5qr4qbK9JZWdn60VGSvMt796NO2h9xa9p8e/
flzZ+dmqG19XvZv31pNdn5RU8pv96D6j9fjXj+v1719X+zrtNfyC4bp0zKXOMQ/m6pOv1uodq/
XwVw9rb8Fe9WvVT4M7Di71paB/Xfov3TrhVrV/u70aJDTQ8+c9rwf/96Ax6vHtSM3lYAw7f5iKiot0/
efXa3fubnWs21FT/
zRVJ0SfIKnkC0cfXv6hHpr2kN5Z8I7Oa3KehnQdolsnlHyuOCw0TBk5GRrw+QBt3rNZ1WOq6/
KWl2vouUMllXz2dObAmfq/r/
9PXUZ1UXFxsZpWbaqrT7r6sD6PYPVo1kNPnPOEHp5WcrwPajdIA04doCVblgQe8+CZD+qGcTeo9T9aK6cgR
2vvXVumsfOL8rUiY4Wy87MP2/Ze1vIyvdbzNQ3/brjunXKvGp/
QWKP6jFK35G6Bx7x76bu66Yub1OHtDmpRvYVeOv8lXfjhhYF6RFiEHpv+mNIy0xQdHq0uDbtoTL8xksp2rB
5vKtK1uCzqxtfVnEFz9MhXj+jCDy9UbkGuGiU1Us+mPRUaEqqQkJBDvs7ed/
p9KiouUq9/99KUP01Rj2Y9NOHaCXr6m6f14pwXFR4WrpbVW+rmdjeXeTv9VNGv12VRluPuUK/
FV7S+InC+Dw0J1eWtLq9w5+eQ4v0/GIQKb/TC0Rq9cDR/zQ04hqSkpWjguIFKuy/
taG8KgMOE6zX+6Oj+WSMAAADgOEDTDQAAAPiMprsSalu7bal4JQCVX3JScqk/0gWg8uN6jT/
iM90AAACAz7jTDQAAAPiMphsAAADwWZlzuvf/M6QA3CrLp7aY1yWs/
WC9lnFxcc5ar169zHVu3rzZWSssdP8ludBQ+15JUZH7j9uEhYUFtc5atWqZ6/
zmm2+cta1bt5rLVibMa+DYcyTnNXe6AQAAAJ/
RdAMAAAA+o+kGAAAAfEbTDQAAAPiMphsAAADwGU03AAAA4LMyRwYCwLHKiuGzovRq1qzprHXt2tVc59q1a5
01K8LKigT0Wtaq5eTkOGvW85Ts+MNjKTIQAMqDO90AAACAz2i6AQAAAJ/
RdAMAAAA+o+kGAAAAfEbTDQAAAPiMphsAAADwGU03AAAA4DNyugEc96wsbsuaNWuctWnTppnLbtq0yVmLjI
x01kJCQsxxrXpiYmJQy+Xn55vr9MoOBwBwpxsAAADwHU03AAAA4DOabgAAAMBnNN0AAACAz2i6AQAAAJ/
RdAMAAAA+IzIQwDHv5JNPNusPP/yws7ZhwwZnrU6dOs5a7dq1zXVmZGQ4a1FRUc5aWFiYOa4V/
Tdq1ChzWRevSMDY2NigxgWA4wl3ugEAAACf0XQDAAAAPqPpBgAAAHxG0w0AAAD4jKYbAAAA8BlNNwAAAOAz
IgMBHPMuv/
xys37FFVc4azt37gxqnXFxcWY9OzvbWbNiAXft2mWO27hxY2dt7Nixzlp6erqzFh8fb66zShUuJQDghTvdA
AAAgM9ougEAAACf0XQDAAAAPqPpBgAAAHxG0w0AAAD4jKYbAAAA8Bk5T5VASEhIULWioiI/
Nkehoe7f1SIiIsxl9+7de7g3x9M555zjrH3zzTdHcEtwtFStWtWsZ2RkOGuZmZlBrXPHjh1mvbCw0FkrLi4
OqibZ21urVi1nbcOGDc6ata2Sf+caADiWcKcbAAAA8BlNNwAAAOAzmm4AAADAZzTdAAAAgM9ougEAAACf0X
QDAAAAPqPpBgAAAHxGTncFYGVtS8Fn9lapYr+8VvautWx+fr6z5lcO90UXXWTWb7nlFmetU6dOztpNN93kr
P3vf/8z1xkeHm7WUXHUqVMn6GWtDOqYmBhnLScnxxzXmmPWOcFrXGt74+PjnTVr7nqdo45G/
j6A8omMjDTrubm5QY3rdb7wQ1hYmFkvKCg4Qlti4043AAAA4DOabgAAAMBnNN0AAACAz2i6AQAAAJ/
RdAMAAAA+o+kGAAAAfEZk4CEKDXX/
nmJFdVnxYOWJsklISHDWdu3aFfS4ViygFTvWuHFjc1xrP3zyySdBrVOSMjMznbUdO3Y4ay+//
LKzdsYZZ5jrzMrKMuuoOKx5K9lz14rWtOaJFScoSXl5eWbdxSueLyIiwlnziggL1p49e3wZF0CJYHsPy5df
fmnW3333XWdtzJgxzpoVZeyXihIJ6IU73QAAAIDPaLoBAAAAn9F0AwAAAD6j6QYAAAB8RtMNAAAA+IymGwA
AAPAZkYGH6GjEAr744ovO2iWXXOKsXXvttea4ixYtctZuvPFGZ+3mm2921jp16mSu87XXXnPWrNi/
VatWmePWqVPHWbPiBidNmuSsEQl47PCKsLLmp7WsFfsXFRVlrjMsLMxZs84z0dHR5rhWxGFOTk5Q67S2VZK
ys7PNOgApJCQk6GWDjQUcP368s3bqqaeay953333O2rx585y
1nTt3muPGxsY6ayeddJKzdtlllzlrXvtn8ODBZv1I4U43AAAA4DOabgAAAMBnNN0AAACAz2i6AQAAAJ/
RdAMAAAA+o+kGAAAAfEbTDQAAAPiMnO7DKNgs7nvuucesd+7c2VlLTU111r777jtz3FdeecVZe+ihh4Ia18
q9lqQGDRo4a8uWLXPW2rdvb44bGur+/fHSSy911pYuXWqOi2ODV460lbG/
fft2Z61evXrOWmRkpLlOK8c7IyPDWfN6Lnv27HHWrFxxi1fO+e7du4MaFzieeM0jP8Z98803nbX09HRz3PP
PP99ZmzNnjrOWn59vjrt3715nzeqjrHPquHHjzHVWFNzpBgAAAHxG0w0AAAD4jKYbAAAA8BlNNwAAAOAzmm
4AAADAZzTdAAAAgM+IDDxEVjRdUVGRs9a3b19n7brrrjPXmZaW5qzVqFHDWVu5cqU57oABA5y1Z555xlm75
ZZbnLX169eb67QiA61lP/
roI3PcF154wazj+Pbbb7+ZdSuKyprz8fHxztrUqVPNdZ599tnOWkREhLMWFhZmjhsSEuKs7dq1y1mznqfXO
jMzM806AFuzZs3MuhXVO2bMGGfNmvOtWrUy12nN+23btjlrhYWF5ri5ublB1axxExMTzXVWFNzpBgAAAHxG
0w0AAAD4jKYbAAAA8BlNNwAAAOAzmm4AAADAZzTdAAAAgM+IDNyPFdUlSXl5ec6aFfnzyiuvOGvz588319m
wYUOz7rJixYqgxz3vvPOctU8+
+cRZmzRpkrnOGTNmmHXAD+np6Wbdigy0zglWbc2aNeY6mzdv7qwlJyc7a1ZUl2THamVnZztrVjyYFYfqNS5
QEVnRmsXFxUGPa82jBQsWOGvh4eHmuNa8/+qrr5y1iy+
+2Fm76KKLzHV6RQC7eEWMWvXo6GhnLT8/31k74YQTvDesAuBONwAAAOAzmm4AAADAZzTdAAAAgM9ougEAAA
Cf0XQDAAAAPqPpBgAAAHxWaSMDvSJpQkPdv09YkVpWJKCXVatWOWtvvfWWs3brrbea4+7cudNZq127trO2Y
cMGc9y9e/c6a+3atXPWzjzzTGetevXq5jqDjQxs27atWW/
cuLGzdvbZZztrJ510krPWpEkTc51erxsqDmtuSlKVKu5TYUFBgbNmRYvt2bPHXGdWVpazZp3frPOXJOXm5j
prOTk5zpoVWea1Tq86ECzrWu4V7WfVg40FvOSSS8z6Rx995KxZ56Fdu3aZ444dO9ZZe/
DBB501K0rPq0ew9pEV32f1FpJ9rrHiSa1zZs2aNc11VhTc6QYAAAB8RtMNAAAA+IymGwAAAPAZTTcAAADgM
5puAAAAwGc03QAAAIDPaLoBAAAAn1XonO6QkBBnza/
c2NjYWLPulb3rMmzYMGfNytGUpIcffthZW716tbPWu3dvc1wr/3vlypXOmpVb/Kc//clcZ/
fu3Z01K2czKirKHNfKV1+zZo2zZj0XK0tUsrPBUbFY80Syc7FjYmKcNSvfOzMz01zn5s2bnbWIiAhnzetvF
Fh5v1aGd2RkpLNmZfKicrGuq5J9fFnLWvnKXpnY1jr9Ovasv3Fh/V0Nr/P+ggULnLV169Y5az/+
+KM5buvWrZ01qy+x9r1Xn2RdH61eKTEx0Rx3/
fr1zpp1TrWOk9NOO81c54UXXmjWjxTudAMAAAA+o+kGAAAAfEbTDQAAAPiMphsAAADwGU03AAAA4DOabgAA
AMBnZY4MtKJaQkOD792tSBor6sYrkuaCCy5w1qzIn4EDB5rjzpkzx1m79dZbzWVdHnnkEbOek5PjrD3xxBP
O2k8//WSOa0XzWDFpVrTYDz/
8YK7TGjcjI8NZ84pqDDYi0ooFrFu3rrnshg0bglonjjyvyEDrPGRFoVm1vXv3muvcvXu3s2ZFs3mdb63zhf
U8rXG9It9QeXi9ltYx4hdrHlm8YuCs6/m1117rrH333XfO2m+//Wau07oG1qpVy1m7/
vrrzXGbNm3qrO3YscNZ2759u7NmXcslOxZwyZIlztqyZcvMcfv16+esWfHASUlJzppXxO8111xj1o8U7nQD
AAAAPqPpBgAAAHxG0w0AAAD4jKYbAAAA8BlNNwAAAOAzmm4AAADAZ2WODLQi2YKNayuPYcOGmXUrYm7t2rX
O2tdff22Oe9dddzlrwUYGehkyZIizdsIJJzhrN954ozluamqqs5afn+
+sbdq0yVmzIoYkKTc311mzYtK8oo2Cja3Mzs521iIiIsxlrX2EisXrHGVFbllxqeWJDLSOH2ueWHNekr799
ltnzYopjI+PN8fF8cE6h1txbjt37nTWvGII4+LinLURI0Y4a3/605/
McWfMmOGsvfTSS86atb1ez6VJkybOmhVlvGvXLnPc9PR0Zy0vL89Zs15P65rrVa9WrZqzdvrpp5vjtmjRwl
mzzovW9drrfNupUyezfqRwpxsAAADwGU03AAAA4DOabgAAAMBnNN0AAACAz2i6AQAAAJ/
RdAMAAAA+o+kGAAAAfFbmnO769es7a1YupST98ssvztrmzZudteLiYmetWbNm5jotXbp0cdbKk1XbsGFDZ2
39+vVBj2tlkg8aNMhZW7hwoTmulV1cr149Z61Dhw7OmvV6StLWrVudNStn08pKluznYh1HFq/
ljkY+PfyRmZnprIWHhztrVgaudaxLdraudbxb2yPZf4fAyhi2svC9solRebz33ntmvXbt2s6a9fcvrGO2Ua
NG5jqTkpKcte+++85Zu//++81xrWPa6iFq1KjhrDVu3NhcZ0JCgrOWkZHhrFmZ/
5K9f63naY3rtc6cnBxnrX379s6aVx9lnW+tc411TfbKHK8o12vudAMAAAA+o+kGAAAAfEbTDQAAAPiMphsA
AADwGU03AAAA4DOabgAAAMBnZY4MvPzyy521m266yVw2KirKWbOi9CZMmOCsWZF2ktSqVStn7ddff3XWvKL
prG06/fTTnbW//OUvztq1115rrtPy888/O2txcXHmsm3btnXW1qxZ46z179/
fWZs5c6a5Tiv+KTc311nziizzigvyg1fcEiqPnTt3OmuJiYnOmnVc/
vbbb0FvT5Uq7lNzdna2uawVvWltr7VOa0xUPH379nXWvK6dVpxbTEyMs2bFuaWnp5vr/
PHHH501K1qzZcuW5rhW9J9VsyLvIiIizHXu3r3brAfLj8g7r2uYVbfiI72ux6Gh7vu9Vr9o7YOj9bocKu50
AwAAAD6j6QYAAAB8RtMNAAAA+IymGwAAAPAZTTcAAADgM5puAAAAwGdljgxcuHChs+YVYWXVa9as6azde+
+9ztr27dvNdf7000/OmhVnY0XZSHYU4VtvvRXUuL/
88ou5TiuuKykpyVmzonckqXfv3s7axIkTzWWDVb16dWfNio/
0igOyWFFWwdYk7xhDVB7Wa2nFVFmRWlb0mmTHFHqdhyxW9F+wkYEc65XLjh07nDXrmiHZUXA5OTnOmnX8eB
3PDRo0cNaio6OdNa/jsmrVqs7aCSec4KxZ11yva5FVt/
aRV3yfdT0KNr7W6xpnxTXm5+c7a17xhl7RzC7lOd9afdbZZ58d1PYEgzvdAAAAgM9ougEAAACf0XQDAAAAP
qPpBgAAAHxG0w0AAAD4jKYbAAAA8FmZIwOtSJ+4uDhz2aysLGfNiiAKNp5IknJzc521yMjIoJbz2qZ169Y5
a1Z8X3lij6ZPn+6s3XPPPea4frD2rWRHCVmvqRXzKNmvW7BRVl6vCzFqxw4rlsxiRVhZkYCStHv3bmct2Jh
CyY4ss87F1jnKOu+h4klJSXHWfvjhB3NZKxY3ISHBWQv2Oi/
Z528rHtgritCKtUtPT3fWrGuGV2Sg9Vysa4ZXfJ/FOl9Y8XxefZS1vcHGrHrVrfOb9XqGh4eb6/
TqIY4U7nQDAAAAPqPpBgAAAHxG0w0AAAD4jKYbAAAA8BlNNwAAAOAzmm4AAADAZzTdAAAAgM/
KnNM9depUZ+3RRx81l61Tp46ztnXrVmfNyqe2MkElO0vTypf0yv20sh6t7bWyJ63lJKlWrVrO2tlnn20ua7
FyLYPNE/XKOf/
pp5+cNSvvPS8vzxy3Zs2azpqV82rlfm7YsMFcp9exgspjz549zpqV4W2dD7yyardt2+asWedFr5xga65Yx3
tMTIyztnfvXnOdqDxuv/32oJe9+uqrnbWbbrrJWTv55JPNca05Zl3rd+3aZY5rZVRby1rXRq/
8fSvv3qpZ2yrZmdrWstb5wq/saq9zlLW91nnT6j28/
kbIF198YdaPFLoGAAAAwGc03QAAAIDPaLoBAAAAn9F0AwAAAD6j6QYAAAB8RtMNAAAA+KzMkYFWvNUpp5wS
9AZcc801zlrHjh2dtR49epjj5uTkOGudOnVy1qyoLsmONsrMzAyq9txzz5nrnDhxorNmxRdZsUeSHR9mxRN
ZcYJeUUHnnHOOs/af//zHWbvyyivNcb/
88ktnrVevXs5aamqqs+YVcxUbG2vWUXlY5zcrjrI8kVtWlKUVMeoVLbZ7925nzZq71rhEBkKSPvnkk6BqXk
4//XRnrVWrVs5aw4YNzXETEhKcNetabsVnel3jrNhcK/7QK27XijW15qcVjeh1/
rLi+6zIXGtbJfu5WrGAVr9jnfckafbs2c7aLbfcYi57OHGnGwAAAPAZTTcAAADgM5puAAAAwGc03QAAAIDP
aLoBAAAAn9F0AwAAAD4rc2SgFQ9jxVt5GTNmTFC1Bx98MOh11qlTx1nbsWOHuWzdunWdtTVr1gS9TX6wIgG
9WLE9lry8PLN+7rnnOmspKSnOWocOHcxxf/
zxR2etadOmzlp8fLyzZsU8SlJaWppZR+WxfPlyZ82KLvWK77NkZ2cHvaxl48aNzpoVAWbxijMDymPu3LlB1
YDKhjvdAAAAgM9ougEAAACf0XQDAAAAPqPpBgAAAHxG0w0AAAD4jKYbAAAA8BlNNwAAAOCzMud0W1ncISEh
5rJWvTwZ38FKT08Pelk/
sritDHSvekFBQdDrtV6XYPN8vVhZ3BYrh9vL6tWrg14Wx4fZs2c7a3/605+ctWXLlgW9Tq9Mexev88WKFSu
ctSpV3Kf8Ro0aOWuxsbHeGwYAMHGnGwAAAPAZTTcAAADgM5puAAAAwGc03QAAAIDPaLoBAAAAn9F0AwAAAD
4rc2Sgpbi4uFz1451XbKJfsYq8LkCJ5cuXO2vW/
CvPHJo3b56ztnfvXmctPDzcHHf79u3O2qZNm5y1d99911n75ZdfzHUCALxxpxsAAADwGU03AAAA4DOabgAA
AMBnNN0AAACAz2i6AQAAAJ/RdAMAAAA+OyyRgQBQmf3666/
O2p49e5w1v+I8s7OznbW4uDhz2by8vKDWuXjx4qCWAwCUDXe6AQAAAJ/
RdAMAAAA+o+kGAAAAfEbTDQAAAPiMphsAAADwGU03AAAA4DMiAwHAkJ+f76wlJCT4sk4rijA+Pt5cNicnJ6
h1VqnivhwUFBQENSYA4Hfc6QYAAAB8RtMNAAAA+IymGwAAAPAZTTcAAADgM5puAAAAwGc03QAAAIDPaLoBA
AAAn5HTDQCGSZMmOWuNGzf2ZZ1jxoxx1lq2bGkuG2xOd2FhYVDLAQDKhjvdAAAAgM9ougEAAACf0XQDAAAA
PqPpBgAAAHxG0w0AAAD4jKYbAAAA8FlIcXFx8dHeCAAAAOBYxp1uAAAAwGc03QAAAIDPaLoBAAAAn9F0AwA
AAD6j6QYAAAB8RtMNAAAA+IymGwAAAPAZTTcAAADgM5puAAAAwGc03QAAAIDPaLoBAAAAn9F0AwAAAD6j6Q
YAAAB8RtMNAAAA+IymGwAAAPAZTTcAAADgM5puAAAAwGc03QAAAIDPaLoBAAAAn9F0AwAAAD6j6QYAAAB8d
lw13QPHDdRlYy472ptRbkNShmjguIFHezMklWxL27faHu3NwDHuWJm7luRXk5WSlnK0N0NSyba8OvfVo70Z
OAKO9twqyzWk2+huum/
KfUdkew63lLQUJb+afLQ3Q1LJtoQMDVHm3syjvSkBZdmmY6nPqHK0N2DguIF6b9F7kqTw0HA1TGyoAW0G6P
Euj6tK6FHfPEnSpqxNeu6b5zRx5URt3L1RNWNrqm3ttrrvtPt0XpPzDtt6kl9N1n2n36f7Tr8v6DFS0lJ07
nvnmo+ZccMMdUvudshjhwwN0edXf67LWl4W3MZ5qEz7GRV/
7v5x+6qEVlHV6Ko6tdapuvbkazWw7UCFhlSeew5/
fC4H0yixkdLuSzvkcUcvHK37ptynzEczg984Q8jQkMD/
x4THqG58XZ3V4Czd3fludajbwZd1Hgsq8tz642t6ME91fUpDug05rOsce/
VYhYeGm49Jy0xT49ca66fbflLb2m0PqA9NGaqV21fqw8s/9P1adij82p/
dRndT29pt9WrPV4PbMGPcmetmOutdG3VVysCUw7rOB898UHd3vtvzcda1fWbaTP3p8z9pw583+LZvyuLoXx
kl9WzWU6P6jFJuQa4mrZykOyfdqfDQcD3W5bEDHptXmKeIsIgjtm1pmWk6692zlBSVpL9e8FedUusU5Rfma
+rqqbpz0p1aftfyI7YtZXFmgzOV/kB64N/3TrlXu3J3aVSfUYGfVY2uGvj/I70/
XSrbfkaJijx3/7h9hUWF2rxns6asmqJ7p9yr/y77r7649gtnA5NfmK/wMPsifyS91vM1DTt/
WODfdV6uo1F9Rqlns56SpLCQsFKPryjzWlJgO/cW7NUvGb/o7R/
f1mkjT9O7fd7VgDYDDrpMYVGhQkJCKtUvRodbRZ1bf7y+fPLzJ3oy5UmtuGtF4GdxEXGHfZ1/
vGYdTF5hnucY41eM16NnP3q4NumwOZT9WVxcrMLiwqP6i9fYq8cG9veGnRvUeWRnfXX9Vzqp5kmS5MtxGBc
RZx5XZTn+x68Yr0tOvORwb9ohqxBntMiwSNWOq61GSY00uNNgnd/kfH3xyxeSfn/
r67lvnlPdl+uqxd9bSCp5sa/
6z1VKGpakqi9WVZ8xfZSWmRYYs7CoUPdPvV9Jw5JU7aVqenjawypW8SFv2x0T71CIQjTv5nnq17qfTqx2ok
6qeZLuP+N+zb15buBx63euV58xfRT3fJwSXkjQVf+5SpuzNgfqq7evVp8xfVRreC3FPR+nTu900ldrvgrUu
43upnU71+nPU/+skKEhnr/9ukSERah2XO3Af9FVogP7t3Zcbb31w1vq/
E5njVwwUo1fa6yoZ6MkHfzt5LZvtdWQlCGBuiT1/aSvQoaGHPB22QeLPlDyq8lKHJaoa/
57jXbn7j6k7a5s+xklKvLc/eP21Uuop/Z12uvxLo9r/DXjNXnVZI1eODrwuJChIRoxf4Qu/
fhSxT4fq+dmPSdJGr98vNr/s72ino1Sk9eaaGjKUBUUFUgquQAOSRmihq80VOSzkar7cl3dM/
mewJhvzn9Tzd9orqhno1RreC1d8ekVQT0HSUqMSiw1ryUpKSop8O9O73TSMzOf0YDPByjhhQTd+uWtB33bd
uGmhQoZGqK0zDSlpKXoxvE3amfuzsBc2DffJSk7P1uDxg9S/AvxavhKQ73949tBbfu+7UxOStaFTS/
Uf6/6r/qf2l93TbpLO3J2SCq54540LElfrPhCrf/RWpHPRmr9zvXKLcjVg/
97UPX+Vk+xz8fqtJGnlfoIzrrMdbrk40t0wosnKPb5WJ305kmatHKSJGlHzg71H9tfNf5aQ9HPRav5G8016
qdRB9vECqmizq0/HoeJUYkKUUipnx2sOUpJS1Hndzor9vlYJQ1L0lnvnqV1metKPca6huz/
8ZLkV5MPON4bv9ZYktTun+0UMjRE3UZ3Czx+w84NWrp1qXo262ley0bMH6GmrzdVxDMRavH3Fvpg0QeltnH
feeKif1+k6Oei1eS1Jvrvsv8e0v7bn7U/l29brvgX4jV55WR1eLuDIp+N1Oz1sw/
6kaD7ptwXeM4Dxw3UzHUz9dr3rwXm9h+Pgx9/+1Ed3+6omOdidOa/
ztSKbStUVlWjqwa2r0ZsDUlStZhqgZ8d7Bcka56WZZv2/3jJwY5/
r2v7Fyu+0KUtLjX3zcy0mer8TmdFPhupOi/X0aNfPRo430slx+Fdk+7SXZPuUuKwRFV/
qbqe+PoJFReXfQ5ViKZ7f9Hh0aV+c52+drpWZKzQtOunacK1E5RfmK8eH/
ZQfES8Zt04S3MGzVFcRJx6ftgzsNzL372s0QtH690+72r2jbO1PWe7Pk/
9vNR6Ri8cbTZd23O2a8qqKbqz052KjYg9oJ4UlSRJKiouUp8xfbQ9Z7tmDpypaddP05oda3T1f68OPDYrL0
u9mvXS9AHT9dNtP6ln05665ONLtH7nekklvz3WT6ivp7s9rfQH0kv99nu4rdq+Sp+lfqaxV43VwtsXlmmZ+
bfMl1Ry1yr9gfTAvyVp9Y7VGrdinCZcN0ETrp2gmetmatjs3+/KHa/
7+XhUUeaupXvj7mpTq43Gpo4t9fMhM4eob8u+WjJ4iQa1G6RZ62ZpwLgBuve0e7XszmX6Z+9/
avSi0Xrum5KG/LPUz/TK3Ff0z97/1Mq7V2rcNeN0Ss1TJEk//PaD7pl8j57u9rRW3LVCU/
pP0TmNzglqe8tq+HfD1aZWG/1020964pwnPB9/ZoMz9WqPV5UQmRCYCw+e+WCg/
vJ3L6tj3Y766bafdEenOzR44uBSF8Juo7sF/
d2SP5/+Z+3O261pa6YFfpadn60X57yokZeO1NI7lqpmbE3dNekufffrdxrTb4wW375YV7a+Uj0/7KmVGSsl
SXdOulO5Bbn6ZuA3WjJ4iV48/8VA0/fEjCe0bOsyTe4/
Wal3pmrExSNUPaZ6UNtbEVSGuXUwBUUFumzMZeraqKsW375Y3930nW5tf6tCQn5fh9c15GD2P97n3TxPkvT
V9V8p/YF0jb369/n9xYov1C25mxIiE5zXss9TP9e9U+7VA2c8oJ/v+Fm3dbhNN46/
UTPWzii13idmPKF+rfpp0e2L1P+U/rrmv9codWvqYdlXLo9Of1TDzhum1DtTdWqtUz0f/1rP13RG/TN0S/
tbAnO7QUKDQP3/vv4/vXzhy/rh1h9UJbSKBn0xKFBLy0xTyNCQw/r9EmuelmWbDmb/49+6ti/
dslRb9mxR98bdnftm466N6vVRL3Wq20mLbl+kEReP0L9++pee/
ebZUut9b9F7qhJaRfNunqfXer6mv839m0YuGFnmfVEhPl6yT3Fxsaavna6pq6aW+vxObHisRl46MvD2wYeL
P1RRcZFGXjoyMHFH9RmlpGFJSklL0YVNL9Src1/VY2c/
pstbXS5Jeqv3W5q6emqp9SVGJqpFtRbO7Vm1fZWKVayW1Vua2z19zXQt2bxEa+9dqwaJJQf2+33f10lvnqT
5G+erU71OalO7jdrUbhNY5pnuz+jz5Z/rixVf6K7Od6lqdFWFhYQpPjI+cCfLL3mFeXr/
svcDv6WWxb7H7rtr9UdFxUUa3We04iPjJUnXn3q9pq+drudU0pwcr/
v5eFLR5q6XltVbavHmxaV+dt3J1+nGdjcG/j1o/
CA9etajuqHtDZKkJic00TPnPqOHpz2sp7o9pfU716t2XG2d3+R8hYeVfO62c73OkkrekYmNiFXvE3srPjJe
jZIaqV2ddkFvb1l0b9xdD5z5QODfG3ZtMB8fERZR6s7a/
no176U7Ot0hSXrkrEf0ytxXNCNthlpUL9nvDRMbqk5cnaC2dd9c/+Pdt/yifL3Z683A/F2/
c71GLRyl9X9er7rxdSWVfLZzyqopGrVwlJ4/73mt37le/
Vr10ym1Sn7ZaXJCk8B463euV7va7dSxbkdJUnJSclDberRVtrm1v125u7Qzd6d6n9hbTas2lSS1qtGq1GO8
riEHs//xHpZZ8hGrfXdd/2j8ivHq06KPJPe1bPh3wzWw7cDAMX//Gfdr7q9zNfy74Tq38e/
fk7qy9ZW6uf3NkkquL9PWTNMb897Qmxe/eQh75dA83e1pXd
D0gjI/PjEqURFhEYoJjzno3H6u+3PqmtxVkvTo2Y/
q4o8u1t6CvYqqEqXw0HC1qNZCMeExh237rXlalm06mP2Pf0nOa/
v4FePVo1kPRYRFBP7bf9+8Of9NNUhooL/3+rtCQkLUsnpL/
bb7Nz3y1SN6suuTgY+6NUhooFd6vKKQkBC1qN5CS7Ys0StzX9EtHW4p076oEE33hF8mKO75OOUX5auouEjX
nXJdqS8OnFLrlFI7dtGmRVq1fZXiX4gvNc7egr1avX21dtbbqfSsdJ1W/7RArUpoFXWs27HU2wB9W/
VV31Z9ndtV1rcMUrelqkFig0AjKEmta7RWUlSSUrelqlO9TsrKy9KQlCGauHKi0nenq6CoQDkFOYE7sEdSo
6RGh9Rwe0lOSg6cLCWpTlwdbdmzJfDv43U/
Hw8q6tz1UqziUnfaJAWas8C2bl6kORvmBD5qIkmFxYXaW7BX2fnZurL1lXp17qtq8noT9WzaU72a99IlLS5
RldAquqDJBWqU2Kik1qynejbtqb6t+h7WC9n+Otbp6P2gQ3Bqzd/vqIWElDTmf5zX7/
d9P+ix972WIfr9NYgIiyh1F2/
J5iUqLC7UiW+cWGrZ3MJcVYupJkm657R7NHjiYP1vzf90fuPz1a91v8AYgzsOVr9P+2lB+gJd2PRCXdbyMp
3Z4Mygt/lIq4xza/3O9Wr9j9aBfz/e5XE93uVxDWw7UD0+7KELml6g8xufr6tOukp14n//
hc3rGnIwZT3ed+Xu0sx1M/WvS/9lPi51a6pubX9rqZ+d1eAsvfb9a6V+dkaDM0r/u/
4ZWrh5YZm2JVj7n5vK64/zbN8vzlv2bFHDxIaql1CvXN+hOunNkwIfHerSqIsm959sztOybNPB7H/
8W8avGK+7Ot1lPiZ1W6rOaHBGqevCWQ3OUlZeln7d9WtgO06vf3qpx5xR/wy9/
N3LKiwqVFho2AHj7q9CNN3nNj5XIy4eoYiwCNWNr3vAlwRiw0t/5CArL0sd6nbQvy//
9wFj1Yg5fM1k82rNFaIQLd9W/i/xPfi/BzVtzTQNv2C4mlVtpujwaF3x6RVl+gLI4bb//pSk0JDQA5rf/
KL8Mo23/7fKQ0JCVFRcVObtOVb38/
Ggos5dL6lbU9U4qXGpn+3/0aasvCwN7TY0cFfwj6KqRKlBYgOtuGuFvlrzlaatmaY7Jt2hv377V80cOFPxk
fFacNsCpaSl6H+r/6cnU57UkJlDNP+W+YGPSx1u+2//
vjszf5zX+YVlm9OSDvgiaYgObV5bUreVvB3f+ITfX4PoKtGlLmZZeVkKCwnTj7f+eMDFbN9b0ze3v1k9mvb
QxJUT9b/V/9MLs1/
Qyxe+rLtPu1sXNb9I6+5bp0krJ2nammk67/3zdGenOzX8wuGH5Tn4rTLOrbrxdUt9ZHHf53tH9Rmlezrfoy
mrpuiTpZ/oLzP+omnXT9Pp9U+XFNw15GAfRTyYySsnq3WN1qVu1lQ2B5vb+38WP9i5vW/
OHa65Pem6SYHeIbpKtCR7nga7TQfrYw4mfXe6fkr/SRefePEhPxc/VIjPdMeGx6pZ1WZqmNiwTN/
KbV+nvVZmrFTN2JpqVrVZqf8SoxKVGJWoOnF19P2v3weWKSgq0I+//XhI21U1uqp6NOuhf8z/h/
bk7Tmgvu8LSq2qt9KGnRu0Yefvb+cu27pMmXsz1bpGyW/
9czbM0cA2A9W3VV+dUusU1Y6rXeqtVankTk9hUeEhbePhUiO2htKzfv8M1K7cXVq7Y22px4SHhvuyfcfTfj
7WVNS5a/l67ddasmWJ+rXq57mtK7atOGA7m1VtFmhoo8OjdUmLS/
T6Ra8r5YYUfffrd1qyZYmkkruI5zc5Xy9d8JIW375YaZlp+nrt14fteXjZ12j9cV4v3LSw1GMiwiJUWHzk5
8Krc0s+S35+k/Odj2lXp50Kiwu1Zc+WA/b/
H98WbpDYQLd3vF1jrx6rB854QO8seCdQqxFbQze0vUEfXv6hXu3xatBfBj0aKuPcqhJapdR6//
ilunZ12umxLo/p25u+1ck1T9ZHSz46bOuVfk/N2P/c/sePluxzsGtZqxqtNGfDnFI/m7NhTuDass/
cX+eW/vfGuWpVvfTHZfxWI6aG0neX/j7S/nfbj9Z1rlFSo8DrXy+hXuDn1jw9XA72nL/
85Uud2eDMUsfiwR7Xqnorfbfhu1I3KeZsmKP4iHjVT6gf+Nn3G78vtdzcX+eqedXmZbrLLVWQpvtQ9T+1v6
rHVFefMX00a90srd2xVilpKbpn8j36ddevkqR7T7tXw+YM07jl47R823LdMfGOA8LXP0/9XC3/
bn+O+B+9/qHC4kJ1HtlZny37TCszVip1a6pe//51nfGvkreZzm9yvk6pdYr6j+2vBekLNG/
jPA34fIC6NuoaeFuoedXmGrt8rBZuWqhFmxbpus+uO+C3uOSkZH2z/
htt3LVR27K3Haa9VTbdk7vrg8UfaNa6WVqyeYluGHfDAQdRclKypq+drk1ZmwKpA2XBfsY+R3LuSiUfRdiU
tUkbd23UgvQFen7W8+ozpo96n9jbGVe3z5PnPKn3F7+voSlDtXTLUqVuTdWYn8foL1//
RVLJF87+teBf+nnLz1qzY40+XPyhoqtEq1FiI034ZYJe//
51Ldy0UOsy1+n9Re+rqLjosH5W1kuzqs3UIKGBhqQM0cqMlZr4y0S9/
N3LpR6TnJSsrLwsTV8zXduytyk7P7vM4w/4fIAe++rA+Lr9Ze7N1KasTVqXuU7TVk/
TFZ9eoY+WfKQRF48w7/qfWO1E9T+lvwaMG6CxqWO1dsdazds4Ty/MekETf5koqSSxYeqqqVq7Y60WpC/
QjLQZgc8LPznjSY1fPl6rtq/S0i1LNWHlhAM+S3wsOdJzq6zW7lirx756TN9t+E7rMtfpf6v/
p5UZKw97o1oztqaiq0Rryqop2py1WTv37lRBUYEmr5qsS1tcWuqxB7uWPXTmQxq9cLRGzB+hlRkr9bfv/
qaxqWNLfblYkv6z7D9696d39UvGL3pqxlOat3Ge7upsf3ThcOveuLt++O0Hvb/
ofa3MWKmnZjyln7f8XOoxyUnJ+n7j90rLTNO27G1lvpO9cddGtfx7S83bOO+wba81Tw+ng13b96WW7P+4/
ffNHZ3u0IZdG3T35Lu1fNtyjV8+Xk+lPKX7z7i/VHTp+p3rdf/
U+7Vi2wp9vORjvTHvDd172r1l3sYK8fGSQxUTHqNvbvxGj3z1iC7/9HLtzt2tegn1dF7j85QQmSBJeuDMB5
Sela4bxt2g0JBQDWo7SH1b9dXOvTsD4+zM3akVGXZUTpMTmmjBrQv03Kzn9MD/
SsasEVNDHep20IiLR0gqeStk/DXjdffku3XOqHMUGhKqns166o2L3giM87cef9Og8YN05r/
OVPWY6nrkrEe0K3dXqXU9fe7Tum3CbWr6elPlFuaq+KngYtKC8ViXx7Q2c616f9xbiZGJeubcZw640/3yhS
/r/v/dr3cWvKN68fXK/Mc42M/
Y50jOXUmasmqK6rxcR1VCq+iEqBPUpnYbvd7zdd3Q9gbPDOgezXpowrUT9PQ3T+vFOS8qPCxcLau31M3tSr
5ElRSVpGGzh+n+/92vwqJCnVLrFH157ZeqFlNNSVFJGps6VkNShmhvwV41r9ZcH/
f7OJBleySEh4Xr434fa/DEwTr1rVPVqW4nPdv9WV35nysDjzmzwZm6vcPtuvq/VysjJ+OQ/
hDH+p3ry5SjfeP4ki+nRlWJUr34ejq74dmad8s8ta/T3nPZUX1G6dlvntUD/3tAG3dtVPWY6jq9/
unqfWJvSSV3Ne+cdKd+3fWrEiIT1LNZT73S4xVJJXezHpv+mNIy0xQdHq0uDbtoTL8xZXpuldGRnluHsl3L
M5brvU/
fU0ZOhurE1dGdne7UbR1vO2zrkErusr9+0et6eubTejLlSXVp2EVPnPOE4iLiDjjWDnYtu6zlZXqt52sa/
t1w3TvlXjU+obFG9Rl1wB+SG9ptqMb8PEZ3TLxDdeLr6ON+Hx9wN9xvPZr10BPnPKGHpz2svQV7NajdIA04
dUDgXTap5EvHN4y7Qa3/0Vo5BTlae+9aY8Tf5Rfla0XGikP6BdyLNU8Pp/2v7VmPZWn62ukH/
BGcg+2b5KRkTbpukh6a9pDavNVGVaOr6qZ2N+kv5/
yl1LIDTh2gnPwcdR7ZWWEhYbr3tHt1a4fS3wWwhBQfSsAgKoQhKUOUlpmm0ZeNPtqbAuAwSX41WaMvGx3UX
4sFcKB7Jt+jgqKCw5YsEsxfskxJS9HAcQOD+ouxKJ+xqWP1l6//
omV3Ljss4x2Ov2RZKe90AwAAWE6uebLOqH+G9wNxTIqLiNOL5794tDejFJpuAABwzDmUt/
1x7Lmw6YVHexMOQNNdCXVL7nbAl18AVG73nX5fpf0DLsDxIJjv/yQnJeu+0+87/
BuDIy5lYEq5x+Az3QAAAIDPKmVkIAAAAFCZ0HQDAAAAPqPpBgAAAHxW5i9ShoSE+LkdFUZYmP2nPAsLg/
uzqiNGjHDWTjrJ/oMZ2dnukHqrtmvXLmdNkjIzM521adOmOWsTJ040x7VYx9Gx9PWCyvJcjpd57SU01H3/
oaiobH/J7VANHDjQWcvPz3fW9u7da46bnJzsrL388svOmsXaP1782n9HA/
P6+NCvXz9n7cQTTzSXzcnJCWqdZ5xhRxvOmDHDWXvrrbeCWidKHMl5zZ1uAAAAwGc03QAAAIDPaLoBAAAAn
9F0AwAAAD6j6QYAAAB8Vua/SMm3ob29+eabztptt93mrFkJJJL97f+oqKigx42Li3PWrBSXwYMHO2v//
Oc/zXWSXlKxMK/9c99995n1rl27Omu//
fabs9auXTtz3PDwcGftb3/7m7P28ccfm+OCeV3RnHrqqWb9sssuc9YuuOACZ62goMBZS0hIMNeZlpbmrDVr
1sxZs67lkrR27dqgtslKG5s5c6a5ztmzZ5v1YwXpJQAAAMAxhKYbAAAA8BlNNwAAAOAzmm4AAADAZzTdAAA
AgM9ougEAAACf0XQDAAAAPiOn+zCaMmWKs3b66ac7a6tXrzbHbdiwobNm5Wnn5+eb427evNlZq1GjhrO2cu
VKZ+2cc84x12kpT4Z3Rcv/Js+3crFy9Pv16+esWVm/77//
vrlOK3d34MCBzpqV4e213h49ejhr5557rrP2xhtvmOt8/fXXnbWtW7eay1Ymx/
u8tsa1atbfmpCkpk2bOmtPPfWUsxYREWGOa9X37NnjrFmvs3XNlexrZ/Xq1Z21efPmmeNa4uPjnTUrt9/
rddm9e7ezdsstt3hvWCVBTjcAAABwDKHpBgAAAHxG0w0AAAD4jKYbAAAA8BlNNwAAAOAzmm4AAADAZ8dlZK
AV+VNYWGgue/
LJJztr33zzjbOWl5fnrG3ZssVcpxVBtGPHDmetWrVq5rjWemvXru2sWRFEHTt2NNe5atUqZ61KlSrOWkFBg
TkukYHBOZbmteWdd94x61deeaWztnPnTmfNOi69jtmEhARnLTTUfT8kOzvbHDc2NtZZs55LZGRkUDVJys3N
ddb+9Kc/
OWvWObMiYl7746WXXnLWWrVq5axZ8bWSHaVnXetjYmKctaysLHOd1nXXuj7OmjXLHNfa3sTERGctPT3dHNf
SsmVLZ23SpEnO2iuvvBL0Oo8GIgMBAACAYwhNNwAAAOAzmm4AAADAZzTdAAAAgM9ougEAAACf0XQDAAAAPn
NntB3DvGIBLVb0WFRUlLNmRTiddNJJ5jqtaL+tW7c6a/
Xr1zfHbdasmbMWbLTYI488Yq7zlltucda8ItYslSXKC/
6Jjo521s4//3xzWStWq6ioyFmz4ketmiTl5OQ4a9a83rt3rzlu48aNnTVrnuzZs8dZs6JJJSkuLs5ZGz16t
LPWpEkTc1xULFaUpTVP6tWrZ45rHbPbtm1z1qxoPyn4+Rns3JTsWOEGDRo4a159SUpKirO2fv16Z82KE6xZ
s6a5zqSkJGetXbt2zlqtWrXMcTdv3mzWj2Xc6QYAAAB8RtMNAAAA+IymGwAAAPAZTTcAAADgM5puAAAAwGc
03QAAAIDPKm1koBVdJNnxRdayS5YsMce1InasWK3Y2FhnbdeuXeY6LVbMnlcEkRVjuHv3bmfNivs5/
fTTzXX+8ssvztrQoUOdtX//+9/muMCQIUOcNa9oMStKLy8vL6jtKc/
8syL44uPjzXGtc4I1r61xq1SxLxVZWVnOWkREhLNWtWpVc9zt27ebdRxZ1nXVct1115l1KwbTOn4yMjLMca
24PGvOW3PI2h6vZbOzs50163wgSbVr13bWTjnlFGfNikO14ogl+/
W2zm+DBg0yx33hhRfM+rGMO90AAACAz2i6AQAAAJ/
RdAMAAAA+o+kGAAAAfEbTDQAAAPiMphsAAADwGU03AAAA4LNKm9MdbF6oJI0bN85Za926tblsZmams2Zl2V
oZnDk5OeY6rWWt/GGvfWRlo1qZvdb2WNmnkp0/
bGV3zp071xx39erVZh3HvrPPPttZ8zourTxfK4Payt0NCwsz12nl+VrLeuV/h4eHO2s///
yzs9ayZUtnzZq3kp23be3bv/71r+a4N910k1lHxWFdM0488URzWSs/vl69es5afn6+Oa51Tbbmn/
W3PLyuq1Z2uHWdys3NNcdNSEhw1rp27eqsWeeSn376yVxnZGSks2ZlfDdr1swct3r16s7atm3bzGUrO+50A
wAAAD6j6QYAAAB8RtMNAAAA+IymGwAAAPAZTTcAAADgM5puAAAAwGeVNjLQixVr165dO2dt+/bt5rhW/
I5XRJiLVxyXFRGWl5fnrJUnJs2KL7Kii6wYIcnev1a8U//+/
c1xn376abOOY9+bb77prHXv3t1ctlevXs6aFcFnzTGvaDFrWatmndskO+LQihazziVWdJgkLV261FkbP368
szZr1ixzXFQeV1xxhbMWERFhLmtFBlrXm+joaHNcaw5aNetabsWESlLNmjWdtfbt2ztrXrG4//
nPf5y1xYsXO2vWddXqASTpqquuctasa71XL3T11Vc7a//4xz/
MZSs77nQDAAAAPqPpBgAAAHxG0w0AAAD4jKYbAAAA8BlNNwAAAOAzmm4AAADAZ8dsZGCLFi2ctaSkJGctJy
cn6HVaMTmhoe7fb6yaJBUUFDhrderUcda8YprK81xdvCKIrChCa9kOHToEvU04Pvz73/8OquYlJSXFWTvpp
JOctd9+
+80c14ryslgRopIdKbhjxw5nLSYmJqgxJalnz55mHce+Cy64wFnLyMgwl7ViOa3j3bqeSPY10DqmreW8rtd
WFKF1zfV6LgMGDHDWrNhhK0I0Pz/fXOe2bducNWv/
WRGQktSsWTOzfizjTjcAAADgM5puAAAAwGc03QAAAIDPaLoBAAAAn9F0AwAAAD6j6QYAAAB8RtMNAAAA+Oy
YzemuX7++s2ZlgnrlS1qZ2QkJCc6alVO6d+9ec51WTqmV+
+mVwx0dHR30si5WJqgkxcfHO2tWbnFycnJQ24Nji1detEtxcXHQ61y1apWzZv09AK91Wnm+Vhaw17jW3wuw
lrXmn5XXWx7lyTzGkTdw4EBnzbpmeF0XLNY1OTIy0lw2Ly/
PWbOuq9ZxV6NGDXOdixYtCqq2Z88ec1zrem09F+ua68V63azt9fobBHXr1nXWHn74YWftpZdeMsetDLjTDQ
AAAPiMphsAAADwGU03AAAA4DOabgAAAMBnNN0AAACAz2i6AQAAAJ8ds5GBF1xwgbNmxY55RVhZsUhWhI4V4
xUbG2uuc/v27c7a/
PnznbWOHTua45588snOmhWNaD0Xr4gvK0rI2rf16tUzx8XxIdjoP+uYlezILSvu01KeODNrnnidoyz5+flB
rXPGjBlBr9NCJGDlUrt2bWfNiqbzusZZc8GKyvM6H1jHdEREhLNm9Qhex6w1x6xxk5KSzHF37doV1LjW/
vOKcmzSpImzZr2mXjHIO3fudNasfudYwJ1uAAAAwGc03QAAAIDPaLoBAAAAn9F0AwAAAD6j6QYAAAB8RtMN
AAAA+OyYjQxs1KhRUMtZcT+SHYWzZcsWZ61OnTpBjSlJzZs3d9asODOvmL3s7GxnzdoPVrxaTExM0Ou0Yt2
2bdtmjgtYgo0alKRNmzY5a1bsmBcr+i/Ympfw8PCgxp07d27Q67TmtXUuQcUzbNiwoJY7/
fTTzbp1vb766qudNa9YTisSz4o4tMbdvXu3uU5rjllxg9a10WtcK67YihOsXr26uc6lS5c6a59+
+qmztmbNGnNcKx74WMedbgAAAMBnNN0AAACAz2i6AQAAAJ/
RdAMAAAA+o+kGAAAAfEbTDQAAAPjsmI0MtCL6rAgdrzguKxLPitnLy8sLqibZcUDlWc6KL8rNzXXWrPi1Kl
XsQ8qKPbJER0ebdet18YpiAiyZmZnOmhWH53UuCXYuWOcvyZ6f1rLWcl5RqoDFK3LSqltRvA8//
LA57rJly5w16/
pXUFDgrHld46wIYCti1DqXeImKinLWgo0JlaTJkyc7a1acINy40w0AAAD4jKYbAAAA8BlNNwAAAOAzmm4AA
ADAZzTdAAAAgM9ougEAAACf0XQDAAAAPjtmc7p37drlrFkZnF7Z1oWFhc5a1apVnbXdu3c7a165u9Y6rcxQ
r2zdhIQEZ83aR8Hmm0pSWlqas9a0aVNnLTIy0hy3YcOGztry5cvNZXHsszKovfz666/
OmnW8lyd31yu7P1hWLq81r9PT0/3YHFQyXtcqF69Meut4r127trO2detWc1xre62/7ZCTk2OOa4mLi3PWrO
dpZXhLdv63xToPeZ0X69evH9Q6vbLMvfqEYxl3ugEAAACf0XQDAAAAPqPpBgAAAHxG0w0AAAD4jKYbAAAA8
BlNNwAAAOCzYzYy0IrtseKLFixYYI5rRRTdeeedzpoVO+YVh2dFeVnPxSveyVrWihu0opY2btxorvMf//
iHs/bKK684a16RU7Vq1XLWiAxEeSIDO3To4KxZ88SK+iyPYGPbJHs/
WOO2bt3aHPf777931vzaDzjygp1H5TkGrHg5r+2xrp3Z2dnOWrBxupId7Wddx7z2kRXDZ0UclmffH43X+1j
HnW4AAADAZzTdAAAAgM9ougEAAACf0XQDAAAAPqPpBgAAAHxG0w0AAAD47JiNDIyNjXXWrDibzMxMc9xFix
Y5a8FGeZUn2s96LlZckiSFhrp/5/
Ja1sUrYujbb7911sLCwoJapyQlJSUFvSyODX4cz5L0wAMPOGu7d+921ryOZytu0IoRtaLOvFjnGutccvvtt
5vjjho1KuhtAix5eXnOmjXnpeCPd6tmRQJKdrRfeeIP/Ti/
Wdsq2fHAlvJEtB7ruNMNAAAA+IymGwAAAPAZTTcAAADgM5puAAAAwGc03QAAAIDPaLoBAAAAn9F0AwAAAD4
7ZnO6rZxbKx9306ZN5rhedRcra9QrpzsiIsJZs3I2vcb1yht12bZtm7Nm5RZL0m+//
easxcfHO2s7d+40x01ISDDrOPYFm5MvSa1bt3bWrHlizWuvjFvrPFSeTF5rWeu8mJub66xZ+6c8vHKWy5Ov
jmODdVx6sY4fK0ffmpvW382QpKysLGctLi7OWfM6f1nnGmteW6zeQpISExODGhdu3OkGAAAAfEbTDQAAAPi
MphsAAADwGU03AAAA4DOabgAAAMBnNN0AAACAz47ZyECvKBwXr2g6Ky7PYkX7FRcXm8vm5OQ4a1bMkFcEUU
FBQVDLlieez+u5BrtcsJFJgCR17tzZWbOixY4Gr5g9izVPsrOznTWv816dOnWctfT0dGetPDGPOD5Yx4jXd
SHY48uaY15jWtf68vQBXlGFwfB6LtHR0Yd9ncc77nQDAAAAPqPpBgAAAHxG0w0AAAD4jKYbAAAA8BlNNwAA
AOAzmm4AAADAZ8dsZGBUVJSzVlhY6Kxt2rTJHNeK9cnPz3fWioqKnDUruk+S9u7d66zl5uY6a
17RYlb0nxW5mJWV5axZ8YaSVKtWLWdt+/
btzppXbFtsbKxZx7HPmtdeOnXq5KxZsVrWcWnN+aPFOidY22tFnUlS48aNnTUrMjDYCFFA8o68s453q2ZdV
8tzLbLG9bp2Vq9e3Vmzeg+r5oVIz8OPO90AAACAz2i6AQAAAJ/
RdAMAAAA+o+kGAAAAfEbTDQAAAPiMphsAAADw2TEbGWhFXFnRWF4xXyeffLKz9ttvvzlr0dHRzppXbJYVKW
iN6yU7O9tZCzZy0eu5nHrqqc6atf+Sk5PNcRMTE806YLGOS2v+WecZrwhDK47LqnlFgQY7P60oNK91WufFb
7/9NqjtAbx4HZfh4eHOmnXsxcTEOGuZmZnmOq3zhTWu1/nCihu0nqfFa/
4RxXv4cacbAAAA8BlNNwAAAOAzmm4AAADAZzTdAAAAgM9ougEAAACf0XQDAAAAPqPpBgAAAHx2zOZ0W5mzV
t6llV0tSb169QpqWSuf0ysbPNgMTmsfSHYuqFXbu3evsxYXF2eu84QTTghqXK99VJ68clQeVi6v1zFiOeWU
U5y17du3O2uRkZHOmte5xMriLk9mtrUfrHldnn3bpUsXZ+3tt9921sjpPj749TqXZ85b22TVrDkv2f2FNf/
Kcw2z/
q7Gnj17nDWvc4nVtyA43OkGAAAAfEbTDQAAAPiMphsAAADwGU03AAAA4DOabgAAAMBnNN0AAACAzyptZKBX
lI0Vs5eTk+OsJSUlmeO2atXKWbMif6xoo927d5vrtJ5rfn6+s2bFCEl2XJAVk2Ztj1dkYEJCgrMWbISh1zY
BXvGZGRkZzpo1d63YPy/
lWdZixZ1Z67TOmV7zr3379t4bBgQhIiLCWfOKvLPq1rXTWqcVbStJeXl5zpp1fbSWk4KPG7TmtVfkYmxsrF
nHoeNONwAAAOAzmm4AAADAZzTdAAAAgM9ougEAAACf0XQDAAAAPqPpBgAAAHxWaSMD4+PjzXqwUUFt27Y1x
61evbqzZsXvZGdnO2tecWbWuFWquF9Crzgla9lg4w/
37NljrrNp06ZBjWvFoEnezxXHhmBj9ho3bmzWrWisXbt2BbU9Xtsa7HMpT9RgsOeSgoICc9yGDRsGvU2AxY
rv8+J1bXWx4vu85p+1vdZ11SuK0DpHWXPX6ne8okCtuEEEh04FAAAA8BlNNwAAAOAzmm4AAADAZzTdAAAAg
M9ougEAAACf0XQDAAAAPqPpBgAAAHxWaXO6vbIyg8183rRpkzmulWtp1azt8crKtDKorRxSr2xdK7/
TqlnjeuV6WnnlVjaqV053YmKiWcexwes4cOnatatZt3JuLcFujxT8vPbKHrbOjdZ5yNoer3NURkaGs9asWT
NnbdWqVea4ODaUZ55UrVrVWfPqA6y5Yl3jrHG9not1HbOyuL3Gzc3NddYSEhLMZV28/
r7Fjh07ghrXS3n2b2XHnW4AAADAZzTdAAAAgM9ougEAAACf0XQDAAAAPqPpBgAAAHxG0w0AAAD4rNJGBnpF
5AQbfxUfH2+Oa0UKJiUlOWsRERHOWrBxZVLwEWBerGWtmlfcT2RkpLO2Zs0aZ61NmzbmuI0aNTLrOL6ddNJ
JZj3Y4708cyzYWECvyECr7hWxFuxyUVFRzlqrVq2cNSIDjw/
liYGLi4tz1rzmnzUXrGuRtZzX9To2NjaoZa0ewWubrJ4m2GhSqXy9CQ6OO90AAACAz2i6AQAAAJ/
RdAMAAAA+o+kGAAAAfEbTDQAAAPiMphsAAADwWaXNg0lLSzPru3btcta2bNnirC1YsMAc9y9/+YuzZsUJho
eHm+NarMif7OzsoNcZbPRfXl6es2bFJUlSgwYNnLWRI0c6a3379jXH3bZtm1nHscGKyLRYx51kR25Z9u7d6
6xZ88Rrndbz3LNnjzmuNXcLCgrMZYPZHq91WnGNX375ZVDbg+OHNce85m1ubq6zZh3TWVlZQa/
TmvfWsl6xilZ8nzWuFefpJT8/
P+hlcXDc6QYAAAB8RtMNAAAA+IymGwAAAPAZTTcAAADgM5puAAAAwGc03QAAAIDPKm1kYHnidWrUqOGsTZs
2zRy3ffv2zlqtWrWcNSuez4v1XHbs2BH0OiMiIoLanpycnKBqkh3J+PTTTztriYmJ5rjliUXCsa9du3Zm3Z
oL1rFXvXp1Zy3YeEMvXuc+qx4SEhJUbevWreY64+PjnbVLL73UWRs2bJg5LlC7dm1nrUmTJuay6enpzprVB
1gRfF4xel5RoS5hYWFm3TpHnXDCCc6atQ+8ziU4/
LjTDQAAAPiMphsAAADwGU03AAAA4DOabgAAAMBnNN0AAACAz2i6AQAAAJ/RdAMAAAA+q7Q53V5ZmZ9+
+mlQy+7Zs8ccd+HChWYdwZs9e7azNnbsWHPZjz/+
+HBvDo4hXjnd9erVc9aSk5OdtQYNGjhrVnauJFWrVs1Zi46OdtasPG3Jzt7duXOns5adne2seZ0XrZz877/
/3lwWsEycONFZW7Rokbms9XcsEhISnLXw8HBnrWrVquY6rblrZXh75XuXZ1kX6xwkec/
7YB3P+eDc6QYAAAB8RtMNAAAA+IymGwAAAPAZTTcAAADgM5puAAAAwGc03QAAAIDPQoqP5+wWAAAA4AjgTj
cAAADgM5puAAAAwGc03QAAAIDPaLoBAAAAn9F0AwAAAD6j6QYAAAB8RtMNAAAA+IymGwAAAPAZTTcAAADgM
5puAAAAwGc03QAAAIDPaLoBAAAAn9F0AwAAAD6j6QYAAAB8RtMNAAAA+IymGwAAAPAZTTcAAADgM5puAAAA
wGc03QAAAIDPaLoBAAAAn9F0AwAAAD6j6QYAAAB8dlw13QPHDdRlYy472ptRbkNShmjguIFHezMklWxL27f
aHu3NQAVWmeddWmaaQoaGaOGmhb6va/
TC0eo2upvv6ymL0QtHK2lY0tHejMMqZGiIxi0fd7Q345hXmee7H1LSUpT8avLR3gxJJdsSMjREmXszj/
amHLeqHO0NGDhuoN5b9J4kKTw0XA0TG2pAmwF6vMvjqhJ61DdPkrQpa5Oe+
+Y5TVw5URt3b1TN2JpqW7ut7jvtPp3X5LzDtp7kV5N13+n36b7T7wt6jJS0FJ373rnmY2bcMEPdkrsd8tgh
Q0P0+dWf67KWlwW3cR4q036u7Cr6vNu6Z6uenPGkJq6cqM17NuuEqBPUpnYbPXnOkzqr4VlHe/
OOuLTMNDV+rbH5mFF9Rmlg24GHPLaf8yE7P1vPzHxGny77VBt3bVR8ZLxa12it+0+/
X31a9jns68PBVeT5HjI0xKw/1fUpDek25MhszFHg1/
PvNrqb2tZuq1d7vhrchhnjzlw301nv2qirUgamHNZ1HkuO/
tVVUs9mPTWqzyjlFuRq0spJunPSnQoPDddjXR474LF5hXmKCIs4YtuWlpmms949S0lRSfrrBX/
VKbVOUX5hvqaunqo7J92p5XctP2LbUhZnNjhT6Q+kB/5975R7tSt3l0b1GRX4WdXoqoH/
P9L706Wy7edjQUWed/
0+7ae8wjy9d9l7anJCE23es1nT10xXRk7GEdsGv+QX5is8LPyQlmmQ0KDUvB7+7XBNWTVFXw34KvCzxMjEw
P8XFhUqJCREoSFH983M2yfcru83fq83LnpDrWu0VkZ2hr7d8O0x8TpWlHNnWVXU+f7H4/qTnz/
RkylPasVdKwI/i4uIC/x/cXGxCosLj/ovCgcT7D6rbM9/7NVjlVeYJ0nasHODOo/srK+u/
0on1TxJkg7YB8Gc746EozV/K8THSyLDIlU7rrYaJTXS4E6DdX6T8/
XFL19I+v2tque+eU51X66rFn9vIankxb7qP1cpaViSqr5YVX3G9FFaZlpgzMKiQt0/9X4lDUtStZeq6eFpD
6tYxYe8bXdMvEMhCtG8m+epX+t+OrHaiTqp5km6/4z7NffmuYHHrd+5Xn3G9FHc83FKeCFBV/
3nKm3O2hyor96+Wn3G9FGt4bUU93ycOr3TSV+t+f2C2W10N63buU5/nvpnhQwN8fzt1yUiLEK142oH/
ouuEh3Yv7XjauutH95S53c6a+SCkWr8WmNFPRslqeRu16tzXy01Vtu32mpIypBAXZL6ftJXIUNDDni77INF
Hyj51WQlDkvUNf+9Rrtzdx/Sdle2/
XwsqKjzLnNvpmatn6UXz39R5zY+V42SGqlzvc56rMtjurTFpYHHhQwN0cgFI9X3k76KeS5Gzd9ori9WfFFq
rJ+3/KyL/n2R4p6PU63htXT959drW/
a2QH3Kqik6+92zA9vb+6PeWr19tXPbCosKNWj8ILX8e0ut37lekjR+
+Xi1/2d7RT0bpSavNdHQlKEqKCootZ0j5o/QpR9fqtjnY/XcrOcOaX9IUlhoWKl5HRcRpyqhVQL/
nrJqiuq8XEdfrPhCrf/RWpHPRmr9zvXqNrqb7ptyX6mxLhtzWeDjaV7zYeqqqWr1j1aKez5OPT/
sqfTd6ToUX6z4Qo+f/bh6Ne+l5KRkdajbQXefdrcGtRsUeEzyq8l6ftbzGjR+kOJfiFfDVxrq7R/
fLjWO13E3f+N8XfDBBar+UnUlDktU19FdtSB9gbltT814SnVerqPFmxdLkmavn60uo7oo+rloNXilge6ZfI
/25O0ptZ3PzHxGAz4foIQXEnTrl7ce0r442irqfP/jcZ0YlagQhQT+vXzbcsW/EK/
JKyerw9sdFPlspGavn63cglzdM/
ke1fxrTUU9G6Wz3z1b8zfOD4x5sI9HjVs+rtTxvWjTIp373rmKfyFeCS8kqMPbHfTDbz8E6kfqeAjm+R/
sIzz3Tbkv8PG0geMGaua6mXrt+9cC8/qPr9uPv/
2ojm93VMxzMTrzX2dqxbYVKquq0VUD21cjtoYkqVpMtcDPqr1U7aDnuxHzR6jp600V8UyEWvy9hT5Y9EFgz
IN9jC9zb6ZChoYoJS1FkrQjZ4f6j+2vGn+toejnotX8jeYa9dPvNxO9jlXXMX6kVYime3/
R4dGB36Qkafra6VqRsULTrp+mCddOUH5hvnp82EPxEfGadeMszRk0R3ERJReFfcu9/
N3LGr1wtN7t865m3zhb23O26/PUz0utZ/
TC0WbTtT1nu6asmqI7O92p2IjYA+pJUUmSpKLiIvUZ00fbc7Zr5sCZmnb9NK3ZsUZX//
fqwGOz8rLUq1kvTR8wXT/d9pN6Nu2pSz6+JHDhHnv1WNVPqK+nuz2t9AfSS/
32e7it2r5Kn6V+prFXjdXC2xeWaZn5t5Sc0Eb1GaX0B9ID/
5ak1TtWa9yKcZpw3QRNuHaCZq6bqWGzhwXqx+t+rmwqyryLi4hTXEScxi0fp9yCXHObh84cqqtaX6XFgxer
V7Ne6j+2v7bnbJdUctLu/l53tavdTj/c+oOm9J+izVmbddV/
rgosvydvj+4/4379cOsPmj5gukJDQtX3k74qKi46YF25Bbm68j9XauGmhZp14yw1TGyoWetmacC4Abr3tHu
17M5l+mfvf2r0otF67pvSjfWQmUPUt2VfLRm8pFTDeThl52frxTkvauSlI7X0jqWqGVvTcxlrPmTnZ2v4d8
P1Qd8P9M2N32j9zvV6cNqDgfq+z4f+8cK2v9pxtTVp1STPX8Jf/u5ldazbUT/
d9pPu6HSHBk8cHGgEynLc7c7brRva3KDZg2Zr7k1z1bxqc/
X6d6+Drre4uFh3T7pb7y9+X7NunKVTa52q1dtXq+eHPdWvVT8tvn2xPrniE81eP1t3Tb6r1LLDvxuuNrXa6
KfbftIT5zzhuX8rsooy38vi0emPath5w5R6Z6pOrXWqHp72sD5L/UzvXfaeFty2QM2qNlOPD3sE5n5Z9B/
bX/UT6mv+LfP1460/6tGzHlV4aMkd2Yp2POz//L281vM1nVH/DN3S/pbAvG6Q0CBQ/7+v/08vX/
iyfrj1B1UJraJBX/x+TtrXAO9rdoOx//nu89TPde+Ue/XAGQ/o5zt+1m0dbtON42/
UjLUzyjzmEzOe0LKtyzS5/2Sl3pmqERePUPWY6pLKdo6QDjzGj4YK9R5NcXGxpq+drqmrpuruzncHfh4bHq
uRl44MvBXw4eIPVVRcpJGXjlRISMlkHtVnlJKGJSklLUUXNr1Qr859VY+d/
Zgub3W5JOmt3m9p6uqppdaXGJmoFtXcv+2s2r5KxSpWy+otze2evma6lmxeorX3rlWDxJID+/2+7+ukN0/
S/I3z1aleJ7Wp3UZtarcJLPNM92f0+fLP9cWKL3RX57tUNbqqwkLCFB8Zr9pxtQ9hrx26vMI8vX/
Z+4HfUsti32OTopIO2L6i4iKN7jNa8ZHxkqTrT71e09dO13MqaTyO1/1cWVS0eVcltIpG9xmtW768RW/
9+Jba12mvro266pqTrznggjOwzUBde8q1kqTnz3ter897XfM2zlPPZj3193l/
V7s67fT8ec8HHv9un3fV4JUG+iXjF51Y7UT1a92v1Hjv9nlXNf5aQ8u2LtPJNU8O/
DwrL0sXf3SxcgtzNeOGGUqMKvkox9CZQ/XoWY/
qhrY3SJKanNBEz5z7jB6e9rCe6vZUYPnrTr5ON7a70XoZyi2/
KF9v9nqz1PHvxZoP+UX5euvit9S0alNJ0l2d79LTM58O1GPCY9SiWotAo3Iwb1/
ytvqP7a9qL1VTm9ptdHaDs3VF6ysO+Fx+r+a9dEenOyRJj5z1iF6Z+4pmpM1Qi+ot9MnSTzyPu+6Nux+w3q
RhSZq5bqZ6n9g78POCogL96fM/6af0nzT7xtmql1BPkvTC7BfU/5T+gc+1N6/
WXK9f9Lq6ju6qERePUFSVkncEuzfurgfOfKDM+7ciqmjzvSye7va0Lmh6gaSSX5RH/
DBCoy8brYuaXyRJeueSdzRtzTT9a8G/
9NBZD5VpzPU71+uhMx8KXHeaV2seqFW04+GPz78sEqMSFREWoZjwmINe557r/
py6JneVJD169qO6+KOLtbdgr6KqRCk8NFwtqrVQTHhM0Nu7//nu2s+u1cC2AwNz/
P4z7tfcX+dq+HfDdW5j+zto+6zfuV7tardTx7odJUnJScmBWlnOEdKBx/jRUCGa7gm/
TFDc83HKL8pXUXGRrjvlulJfHDil1imldtKiTYu0avsqxb8QX2qcvQV7tXr7au2st1PpWek6rf5pgVqV0Cr
qWLejiot/f+urb6u+6tuqr3O7/
vhYS+q2VDVIbBBoBCWpdY3WSopKUuq2VHWq10lZeVkakjJEE1dOVPrudBUUFSinICdwB/ZIapTU6JAabi/
JScmBhluS6sTV0ZY9WwL/Pl73c0VXUeedJPVr3U8Xn3ixZq2bpbm/
ztXkVZP10pyXNPLSkaW+LPjHJjw2IlYJkQmBY2/
R5kWasXaG4p6P2394rd6+WidWO1ErM1bqyZQn9f2v32tb9rbAHe71O9eXarqv/
exa1U+or68HfK3o8Ojf98nmRZqzYU6pj4wUFhdqb8FeZednBy5c+y4UfooIiyjTXbCyigmPCTTc0oHzunO9
zp7ftTin0Tlac88azf11rr7d8K2mr52u10a9pqHdhuqJrr/fGTy15u/bHRJS8vZ64HX0OO7UVNqctVl/
+fovSlmXoi17tqiwqFDZ+dkHzPs/T/2zIsMiNffmuYG7ZFLJ67h482L9e8m/
Az8rVrGKiou0dsdatarRSpLUsY7/r6NfKvJ89/LH+bN6x2rlF+XrrAa//
+IWHhauzvU6K3VbapnHvP+M+3Xzlzfrg8Uf6Pwm5+vK1lcGjveKdjwc7vPHH88TdeLqSJK27NmihokNVS+h
Xrm/Q7X/9qZuTdWt7Ut//OasBmfpte9fK/
OYgzsOVr9P+2lB+gJd2PRCXdbyMp3Z4ExJZTtHSAce40dDhWi6z218rkZcPEIRYRGqG1/3gC8JxIaX/
shBVl6WOtTtoH9f/m/tr0bM4Wsmm1drrhCFaPm28n+J78H/
Pahpa6Zp+AXD1axqM0WHR+uKT68o9dbHkbL//pSk0JDQA5rf/
KL8Mo23/52ukJCQg74973Ks7ueKrqLOu32iqkTpgqYX6IKmF+iJrk/o5i9u1lMpT5Vquvf/gk6Ifj/
2svKydEmLS/Ti+S8eMPa+C80lH1+iRkmN9M4l76hufF0VFRfp5BEnH3C89GrWSx8u+VDf/
fpdqbuqWXlZGtptaOBO3/7bv8/BPjZ1uEVXiQ7c5dnncM/rYL4XEx4Wri6NuqhLoy565OxH9Ow3z+rpmU/
rkbMfCVwAvV5Hr+PuhnE3KCMnQ6/1fE2NEhspskqkzvjXGQe8jhc0uUAf//yxpq6aqv6n9g/
8PCsvS7d1uE33nHbPAetomNgw8P9H4nX0S0Wf75ZD3e+hIaEHHKv5haWP+yHdhui6U67TxF8mavKqyXoq5S
mN6TdGfVv1rXDHw/7rKcvzs/
xxvu07ZxzKNdtLMK+XVPoG3P7P56LmF2ndfes0aeUkTVszTee9f57u7HSnhl84vMzH6sF6nyOtQjTdseGxa
la1WZkf375Oe32y9BPVjK2phMiEgz6mTlwdff/r9zqn0TmSSt5W/PG3H9W+Tvsyr6dqdFX1aNZD/5j/
D91z2j0HHEiZezOVFJWkVtVbacPODdqwc0PgLuyyrcuUuTdTrWu0liTN2TBHA9sMDPzGn5WXdcBnISPCIlR
YVFjm7TucasTWUHrW75/
n3JW7S2t3rC31mPDQcF+273jazxVJRZ13Lq1rtD6knOX2tdvrs9TPlJyUfNBv+2dkZ2hFxgq9c8k76tKoi6
SSL08dzOBOg3VyzZN16ceXauJ1EwNvzbav014rtq04pP14JO0/
rwuLCvXzlp91bvLvb+ke6fnQukZrFRQVaG/B3jLddSrLcTdnwxy92etN9WreS1LJl6r+
+IXZfS5tcakuOfESXTf2OoWFhumak68JrGPZ1mUV9nU8HCrbfHdpekJTRYRFaM6GOWqU1EhSSYM2f+P8wMd
BasTU0O7c3dqTtydwPTlY1v6J1U7UiWecqD+f8Wdd+9m1GrVwlPq26lvhj4caMTX085afS/
1s4eaFpX5RrkjXuVY1WmnOhjmBj+FJJXN233V7X2OcnpWudmon6eCvV43YGrqh7Q26oe0N6vJDFz007SENv
3B4mY7ViqJCfpHSS/
9T+6t6THX1GdNHs9bN0toda5WSlqJ7Jt+jX3f9Kkm697R7NWzOMI1bPk7Lty3XHRPvOCAQ/vPUz9Xy7/
bniP/R6x8qLC5U55Gd9dmyz7QyY6VSt6bq9e9f1xn/OkOSdH6T83VKrVPUf2x/
LUhfoHkb52nA5wPUtVHXwNsszas219jlY7Vw00It2rRI13123QG/
WSYnJeub9d9o466NB71g+Kl7cnd9sPgDzVo3S0s2L9EN425QWGjYAds3fe10bcrapB05O8o8Nvv52HCk5l1
Gdoa6v9ddHy7+UIs3L9baHWv1n6X/0UtzXlKfFmXPdr6z853anrNd1352reZvnK/
V21dr6qqpunH8jSosKtQJ0SeoWnQ1vb3gba3avkpfr/1a90+93zne3afdrWe7P6veH/cONOdPnvOk3l/
8voamDNXSLUuVujVVY34eo798/
Zcyb6efuid318SVEzXxl4lavm25Bk8cfMDrEex8mLdxnlr+vaU27trofEy30d30zx/
+qR9/+1FpmWmatHKSHp/+uM5tfG6ZL45lOe6aV22uDxZ/oNStqfr+1+/Vf2x/
RVeJPuh4fVv11Qd9P9CN42/Uf5f9V1LJ58i/
3fCt7pp0lxZuWqiVGSs1fvl43TXproOOcTw4ktfZQxEbEavBHQfroWkPacqqKVq2dZlu+fIWZedn66Z2N0m
STqt/mmLCY/T49Me1evtqfbTkI41eNDowRk5+ju6adJdS0lK0LnOd5qyfo/
kb56tV9ZKPjVT046F74+764bcf9P6i97UyY6WemvHUAU14clKyvt/
4vdIy00p9dM7Lxl0b1fLvLTVv47zDtr0PnfmQRi8crRHzR2hlxkr97bu/aWzqWD14ZskXs6PDo3V6/
dM1bPYw
pW5N1cy0mfrLjNLn0CdnPKnxy8dr1fZVWrplqSasnBD4mE9ZjtWKokLc6T5UMeEx+ubGb/TIV4/
o8k8v1+7c3aqXUE/nNT4vcCJ/
4MwHlJ6VrhvG3aDQkFANajtIfVv11c69OwPj7MzdqRUZdlROkxOaaMGtC/
TcrOf0wP9KxqwRU0Md6nbQiItHSCp5e2b8NeN19+S7dc6ocxQaEqqezXrqjYveCIzztx5/06Dxg3Tmv85U9
ZjqeuSsR7Qrd1epdT197tO6bcJtavp6U+UW5qr4qUN/
KzdYj3V5TGsz16r3x72VGJmoZ8595oA73S9f+LLu/
9/9emfBO6oXX09p96WVaWz287HhSM27uIg4nVbvNL0y9xWt3l7y+c0GCQ10S/
tb9HiXx8u8vXXj62rOoDl65KtHdOGHFyq3IFeNkhqpZ9OeCg0JVUhIiMZcMUb3TL5HJ795slpUb6HXe76ub
u91c4553+n3qai4SL3+3UtT/
jRFPZr10IRrJ+jpb57Wi3NeVHhYuFpWb6mb291c5u3006B2g7Ro8yINGDdAVUKr6M+n/
7nUXW4p+PmQnZ+tFRkrzI+r9GjaQ+8tek+Pf/
24svOzVTe+rno3760nuz5Z5udQluPuX5f+S7dOuFXt326vBgkN9Px5z+vB/
z3oHPOK1leoqLhI139+vUJDQnV5q8s1c+BM/d/X/6cuo7qouLhYTas21dUnXe0c41h3JK+zh2rY+cMCr9/
u3N3qWLejpv5pqk6IPkFSybunH17+oR6a9pDeWfCOzmtynoZ0HaJbJ5R8rjgsNEwZORka8PkAbd6zWdVjqu
vylpdr6LlDJZV85rkiHw89mvXQE+c8oYenPay9BXs1qN0gDTh1gJZsWRJ4zINnPqgbxt2g1v9orZyCHK29d
60x4u/yi/K1ImOFsvOzD9v2XtbyMr3W8zUN/
2647p1yrxqf0Fij+owq9Uf63r30Xd30xU3q8HYHtajeQi+d/5Iu/
PDCQD0iLEKPTX9MaZlpig6PVpeGXTSm3xhJZTtWK4qQ4rJ+iw0VxpCUIUrLTNPoy0Yf7U0BcJiMXjhaoxeO
5q+5AceQlLQUDRw3sMw3qXBsq5QfLwEAAAAqE5puAAAAwGeV8jPdx7tuyd0O+LIKgMqtbe22peIQAVR+yUn
JgVQVgM90AwAAAD7j4yUAAACAz2i6AQAAAJ+V+TPd+/
954SPBWqfXp2LKs6wfkpOTnbUBAwaYyxYVuUPtY2JinLXIyEhz3KFDhzpru3btctYq2r6tiCrLfjga8xqor
JjXx44hQ4Y4a5mZmc5adradX21dry3h4eFmPSwszFlr3Lixs/
bAAw8EtT3HkyM5r7nTDQAAAPiMphsAAADwGU03AAAA4DOabgAAAMBnNN0AAACAz2i6AQAAAJ+V+S9SVrbIw
GDH9XqewcYBLV261Flr1qyZuWxERISzlpeX56x57aOXXnrJWXvyySfNZYN1vMQNVpbnQrQYUHbM62PH0Xgt
rXV6vWYFBQXOWpUq7vTnnj17OmtTp04113m8IDIQAAAAOIbQdAMAAAA+o+kGAAAAfEbTDQAAAPiMphsAAAD
wGU03AAAA4DOabgAAAMBn7nDHCqA8mc5hYWHOmpW1HWwOtyTdc889zlpcXJyztmrVKnNcK4t7x44dzlp8fL
w57jXXXOOsffvtt87alClTzHGDVZ5s2cqSnwsAODK6devmrFnXjF9+
+cWHrSnf3wjJzs521urWreusde7c2Vkjp/vI4043AAAA4DOabgAAAMBnNN0AAACAz2i6AQAAAJ/
RdAMAAAA+o+kGAAAAfBZSXMastfLEuQWrPPE6VvSftexDDz1kjnvzzTc7a1WquBMYc3NznbWIiAhznVYsoL
Xsnj17zHGtZWNiYpw1K7roxRdfNNf56aefmnUXr9e7okUGVrTtcTka8xqorJjXlcsjjzzirL3wwgvO2sqVK
50169roxepLCgoKzGWt/iI6OtpZW7JkibN23nnnmes8XhzJec2dbgAAAMBnNN0AAACAz2i6AQAAAJ/
RdAMAAAA+o+kGAAAAfEbTDQAAAPjMnUFTAVgxLl4RL7Vr13bWfvzxR2ctIyPDHNeK79u1a5ezZkU4JSUlme
usVauWs5aWluaseUUG5ufnO2vW/j3hhBOctf/7v/8z1/
nss886a61atXLWCgsLzXEBAPij2NhYZ826xlnxfVb8r9e45VnOugbm5OQ4a9Y+wJHHnW4AAADAZzTdAAAAg
M9ougEAAACf0XQDAAAAPqPpBgAAAHxG0w0AAAD4rEJHBloxe17xOt98842ztmXLFmfNKzIwLCwsqJoV92NF
DUrShg0bzLpLdHS0WY+KinLWIiMjg1qn1/6rUaOGs/bpp586a/
369TPHLc+xAgA49ljXsaKiImctNNR9P9KqSfZ11apZccSSHWNo4fpXsXCnGwAAAPAZTTcAAADgM5puAAAAw
Gc03QAAAIDPaLoBAAAAn9F0AwAAAD6j6QYAAAB8VqFzuq18ySeffDLoZbdu3eqsWTmakpSfn+
+sWbmf1vZYGd5e4yYmJjprVna113qzs7OdtSpV3IeNta2SnePdrFkzZ61NmzbmuIsWLTLrAADsY/
1dDes6Hx4ebo5r/
d2N3377zVlr3ry5Oa61rHVN3rt3rzkujizudAMAAAA+o+kGAAAAfEbTDQAAAPiMphsAAADwGU03AAAA4DOa
bgAAAMBnFToy0HLzzTebdSuaLjc311nziu+zonm8lg1mTEmKiIhw1qxYQCumsDzblJeX56xlZWWZ49arV89
ZCw11/w542223mePecccdZh0AcHxZvnx5UMtZ1/KkpCRz2VmzZjlr//d//
+esLV261Bx3zZo1zlqtWrWctW3btpnj4sjiTjcAAADgM5puAAAAwGc03QAAAIDPaLoBAAAAn9F0AwAAAD6j
6QYAAAB8VqEjA+Pi4py1yMhIc9mYmBhnrVq1as5adna2OW5+fr6zFh4eHtRyXgoKCpy18sQCWrFIRUVFzpo
V7RcdHW2uMyoqynvDDqJRo0ZBLYeKx4q5tI4tKfhYzrCwsKDH9Vo2WNZ+sOZ8zZo1zXEvueQSZ61nz57O2q
pVq5y1xx57zFyntY+Cfc2A8lq8eLGzZs2/YGuSHbfr1V9YrGuy1StZ+wBHHne6AQAAAJ/
RdAMAAAA+o+kGAAAAfEbTDQAAAPiMphsAAADwGU03AAAA4DOabgAAAMBnFTqnu379+s6aV1amlQ1bntxYK4
/Wysz22l6LlftZnpxu67kEmw3ulWmcmJjorOXm5jprERER5rioPKwsbiuL1k/Bzs+jkUG9e/
dusz5y5EhnzcoJ7t27d9DbdDT2g3VOyMvLc9a8cs5POumkoLcJFcvy5cuDWs76OyBe1zhrjpXnem0ta82Fn
3/+Oeh14vDjTjcAAADgM5puAAAAwGc03QAAAIDPaLoBAAAAn9F0AwAAAD6j6QYAAAB8VqEjAxs3buysecVm
5eTkOGtWZNTGjRvNcatWreqs7dy501mLiopy1qx4K8mOL7LGLU+MlxWhZkUXWXFwkv1crO21nicqF78iO/
2KrQt23Pbt25v1F154wVm7/vrrnTUrSlWSPvroI2ctLS3NWbv//
vudtSuuuMJc55133mnWg2W93nPnznXWXn31VWft22+/NddJZOCxY8+ePc5afn6+s2Zd/
6xrmCTt2rXLWYuOjjaXtVjrtWKFlyxZEvQ6cfhxpxsAAADwGU03AAAA4DOabgAAAMBnNN0AAACAz2i6AQAA
AJ/
RdAMAAAA+q9CRgbVr13bWsrOzzWXj4uKCqnnFg1kRRBEREc6aFaVXUFBgrjM8PNyXca2YISsW0BrXimGS7C
jHoqIiZ82KecTxI9j4vvLECVpRXc8884yz1rFjR3Nca17ffffdztpnn31mjjty5Ehn7emnn3bW/vvf/
zprN9xwg7nO+fPnO2u9e/d21qxIQEkaMWKEs5aamuqsTZgwwVm74447zHVmZmaadRwbNm/
e7KxZ10Zr3krS1q1bgxrXi7Ws1ZesWrUq6HXi8ONONwAAAOAzmm4AAADAZzTdAAAAgM9ougEAAACf0XQDAA
AAPqPpBgAAAHxWoSMDragurzi8pKQkZ82KDLQi+CQpLy/
PrLtYcYLW85SkkJAQZ82KLwp2WyU7ymvv3r3OmtdzsaLbrLjBqKgoc1ygQYMGzlpsbKy57JVXXumsWZF35Z
ljV1xxhbN2zz33OGtXX321Oe7LL7/
srM2YMcNZGzhwoLNmRRhK0tSpU521WbNmOWu5ubnmuK1bt3bW2rRp46w1btzYWWvYsKG5zoyMDLOOY0N6er
qzZsUVW9djSdq9e7ezlpCQ4L1hDta13oopRMXCnW4AAADAZzTdAAAAgM9ougEAAACf0XQDAAAAPqPpBgAAA
HxG0w0AAAD4jKYbAAAA8FmFzum2cimtHGlJiomJcdasbF0rT1uSEhMTnTUrZ9rKFY+OjjbXuWfPHmfNyr2u
Vq2aOa6VkWtljVr73uu5VKniPuSs/G/reUp2vnpRUZG5LCoPa/5de+21zlq/fv3Mca3j1ponc+bMcda+/
PJLc52dO3d21r744gtn7bTTTjPH7dmzp7P25JNPOmtr1qxx1qxMcUk66aSTnLWbbrrJWbvrrrvMcQcPHuys
WXna5513nrM2efJkc50nnniiWcexwTre69SpE/
S41vXR6+8FBDtuWlpa0OPiyOJONwAAAOAzmm4AAADAZzTdAAAAgM9ougEAAACf0XQDAAAAPqPpBgAAAHxWo
SMDrWg/K05QsiPA0tPTnTUrek6yo/
8SEhKcte3btztr1vOU7PhDK86suLjYHDcuLi6oZXft2uWsecXz7dy501krT+xfUlKSs2bte7hFRkaadSv+K
thoLK8o0IYNGzpr1lxYunSpOW5OTo6zNmLECGetSZMmzlpWVpa5TuuYbdasmbP2ww8/
mOPefPPNzlpycrKzNm/
ePGetZcuW5jofeOABZ2369OnOWtWqVc1x69Wr56x16tTJWbPOUV7XDq86jg1bt2511qxrUUhIiDnu+vXrnb
WoqCjvDQtivatWrQp6XBxZ3OkGAAAAfEbTDQAAAPiMphsAAADwGU03AAAA4DOabgAAAMBnNN0AAACAzyp0Z
KAVueUVvVOlivupZWdnO2tekYFWLFliYqKzZsX9eEUQeW2Ty549e8y6FUVYo0YNZ23btm3OWq1atbw3zGHH
jh3OmleMV+3atZ01IgPdTjvtNGfNK77PmgvWMW1FVZZnndZxeeutt5rjWlF6Vnzf3r17g1pOkvLz85016xx
l7QNJ+vzzz521nj17Omtnn322s/bzzz+b62zQoIGzZh1j119/vTnuTz/
95KxZ54SVK1c6a4WFheY6vc7HODZYMaJXXnmls2adZyT7OuY1dy3Wcblhw4agx8WRxZ1uAAAAwGc03QAAAI
DPaLoBAAAAn9F0AwAAAD6j6QYAAAB8RtMNAAAA+IymGwAAAPBZhc7pLioqCnpZKw+zoKDAWfPK/
7Yys61xre2Jjo421xkRERHUuHl5eea4Vk5wcXGxs2bl3Hqts2bNms6atW+9jgUrVxxuderUCXrZ9PR0Z806
Lq38/cjISHOd1rH322+/OWt9+/
Y1xw02m3njxo3OmpWDL0mxsbHO2po1a5y1jIwMc9ycnBxnbcKECc5ajx49nLXu3bub67ReF+t5eu1363jYu
XOnuayL1+tiHZ84dljHj/X3Arz+boaV023Nay/
WevlbFJUHd7oBAAAAn9F0AwAAAD6j6QYAAAB8RtMNAAAA+IymGwAAAPAZTTcAAADgswodGWjFRXlFyFmxWb
Vq1XLWNm/
ebI5rxV9ZEhISnDWv2Cwrfs3iFaNnxSNa+75atWrO2p49e8x1WjGFVjSi1+ttvaZws45n63WWpL179zprVo
ScNTfDw8PNdVrHjxVXuW3bNnPcuLg4Z82KkLO2Nzs721yntU3W8XzyySeb41oRpNac3717t7PmFc9nRU9Wq
eK+zPz000/muNb+rV+/
flA1r3PUrFmzzDqODaeddpqzZkXmVq1a1Rx369atzppX3KfFijFs37590OPiyOJONwAAAOAzmm4AAADAZzT
dAAAAgM9ougEAAACf0XQDAAAAPqPpBgAAAHxWoSMDrZg9rxg9a1krziw1NdUct0OHDs6aFRFWUFBgjmtJSk
py1qwINSuaTbJj+Kz4MCtebd68eeY6GzRo4KxZMU1WHJxkR6HBbdKkSc7aBRdcYC574oknOmtWbNamTZucN
a95bdWtOeYVKbl9+3ZnzYpOjI+Pd9a8okCbN2/urFnzzzp/SdLixYudNWv/tW7dOuh1WtF/
VjRiTEyMOa4Ve2rt3wULFjhrXrFtW7ZsMes4NtStW9dZ85q7lhUrVgS9rMXaJq5/
lQd3ugEAAACf0XQDAAAAPqPpBgAAAHxG0w0AAAD4jKYbAAAA8BlNNwAAAOCzCh0ZaEV1ebHi8ObMmeOseUW
WWXF54eHhzlpYWJiz5hVPZG2Ttc7o6Ghz3CpV3C9/
aKj79zGrtnz5cnOdl112mbO2a9euoNYpSbGxsWYdB7djxw5nbfz48eayderUcdbOP/
98Z61z587OWlZWlrnOzMxMZ806frxERkY6a9axt2fPHmfNK2Zv9uzZztrq1audtc2bN5vjWueoNm3aOGtW3
KdX/
Kh1frPiPq3XU7KPT6/9C1iaNm3qrJUn4tcv1hyzzl+oWLjTDQAAAPiMphsAAADwGU03AAAA4DOabgAAAMBn
NN0AAACAz2i6AQAAAJ/RdAMAAAA+q9A53VbOtFdGq1X/
4YcfnLXhw4eb45500knOWlpamrmsi1fG5vbt24Matzys/O/8/
Pygx33iiSecteLi4qDXWbt27aC3CQfnlVlvHe8jR4501qxM9ebNm5vrtLJ1rdz51NRUc1zr2LPs3r3bWfPK
Dbcyya1M3qioKHNcKxd7zZo1zpo1xzIyMsx1euWrB8v6GwZWzdp/Vk0i//t4YR3T9evXD3pcq2/
xyru3WH8voDzXZBxZ3OkGAAAAfEbTDQAAAPiMphsAAADwGU03AAAA4DOabgAAAMBnNN0AAACAzyp0ZKAVrx
MREWEua8Vqffnll0Fv09KlS4Ne1mXPnj2Hfczy8iuCaOXKlc5ajRo1nDWv+LqioqKgtwlHlnW8L1y40FzWq
36ssGLrvM4XVn3Hjh1Bb9PREGyUY0FBQVA1HD+sa5xXrKTF6k3KExloRWQSGVh5cKcbAAAA8BlNNwAAAOAz
mm4AAADAZzTdAAAAgM9ougEAAACf0XQDAAAAPqvQkYHBxkVJduTPihUrgh63ShX3LrNi60JD3b/
feD3P8uwHixVBZK2zPPF8qampzpoVGei1D4gMBACUlRUrbMnMzDTrVo9QHlZPU54oQhxZ3OkGAAAAfEbTDQ
AAAPiMphsAAADwGU03AAAA4DOabgAAAMBnNN0AAACAz2i6AQAAAJ9V6JzuwsLCoJdNS0s7fBvyB9Y2+ZVtH
Swrh1sKfpuCzfeWpB9//
NFZO+ecc5y1vLw8c9yCggKzDgDAPsH+XQ2v62p+fn7Q22Qpz3UXFQd3ugEAAACf0XQDAAAAPqPpBgAAAHxG
0w0AAAD4jKYbAAAA8BlNNwAAAOCzCh0ZmJWV5ax5RcRZ8Trlid6xooSCjTj0iiCqaHFA5dkHixYtOuzrlKR
du3YFNS4A4PhjxdB6XW8swUbxbtu2zaxHRUU5axWtR4Abd7oBAAAAn9F0AwAAAD6j6QYAAAB8RtMNAAAA+I
ymGwAAAPAZTTcAAADgswodGXjqqac6a+Hh4eayMTExzlp5Iu/
8iAz0K+7Hr3HLsw+2bNnirEVGRjprVlySJHXs2NGsAwCwz7p165y1Ll26OGt79+41x7Wiji1WhKFkXwM3b9
4c1Dpx5HGnGwAAAPAZTTcAAADgM5puAAAAwGc03QAAAIDPaLoBAAAAn9F0AwAAAD6r0JGB8+bNc9YiIiLMZ
a3IwGCj/SSpqKgo6GUrk5CQEGetPPtg48aNztrChQudNa9IpJ07dwa7SQCA48z333/
vrA0aNMhZi4uL82NzzGuuZEf1Llmy5HBvDnzCnW4AAADAZzTdAAAAgM9ougEAAACf0XQDAAAAPqPpBgAAAH
xG0w0AAAD4jKYbAAAA8FlIcXFxcZke6JEheazwep5l3F2VnrUfjpd9UB6VZR8dL/
MaOByY18eOs88+21mbPHmys7Z9+3Zz3EaNGgW1PRs2bDDr1np79+4d9Lg4svOaO90AAACAz2i6AQAAAJ/
RdAMAAAA+o+kGAAAAfEbTDQAAAPiMphsAAADwWZkjAwEAAAAEhzvdAAAAgM9ougEAAACf0XQDAAAAPqPpBg
AAAHxG0w0AAAD4jKYbAAAA8BlNNwAAAOAzmm4AAADAZzTdAAAAgM/+Hwcl/
CxOGhJ0AAAAAElFTkSuQmCC",
"text/plain": [
"<Figure size 900x900 with 9 Axes>"
]
},
 "metadata": {},
"output_type": "display_data"
}
],
"source": [
"# Plot predictions\n",
"plt.figure(figsize=(9, 9))\n",
"nrows = 3\n",
"ncols = 3\n",
"for i, sample in enumerate(test_samples):\n",
" # Create a subplot\n",
" plt.subplot(nrows, ncols, i+1)\n",
"\n",
" # Plot the target image\n",
" plt.imshow(sample.squeeze(), cmap=\"gray\")\n",
"\n",
" # Find the prediction label (in text form, e.g. \"Sandal\")\n",
" pred_label = class_names[pred_classes[i]]\n",
"\n",
" # Get the truth label (in text form, e.g. \"T-shirt\")\n",
" truth_label = class_names[test_labels[i]] \n",
"\n",
" # Create the title text of the plot\n",
" title_text = f\"Pred: {pred_label} | Truth: {truth_label}\"\n",
" \n",
" # Check for equality and change title colour accordingly\n",
" if pred_label == truth_label:\n",
" plt.title(title_text, fontsize=10, c=\"g\") # green text if correct\n",
" else:\n",
" plt.title(title_text, fontsize=10, c=\"r\") # red text if wrong\n",
" plt.axis(False);"
]
},
{
"cell_type": "markdown",
"id": "5ce6dc44-90a5-48c3-91a5-810fa084d98b",
"metadata": {
"id": "5ce6dc44-90a5-48c3-91a5-810fa084d98b"
},
"source": [
"Well, well, well, doesn't that look good!\n",
"\n",
"Not bad for a couple dozen lines of PyTorch code!"
]
},
{
"cell_type": "markdown",
"id": "ab108078-6770-4cb9-ac62-a761ff159aba",
"metadata": {
"id": "ab108078-6770-4cb9-ac62-a761ff159aba"
},
"source": [
"## 10. Making a confusion matrix for further prediction evaluation\n",
"\n",
"There are many [different evaluation
metrics](https://www.learnpytorch.io/02_pytorch_classification/#9-more-
classification-evaluation-metrics) we can use for classification problems. \n",
"\n",
"One of the most visual is a [confusion
matrix](https://www.dataschool.io/simple-guide-to-confusion-matrix-terminology/).\
n",
"\n",
"A confusion matrix shows you where your classification model got confused
between predictions and true labels.\n",
"\n",
"To make a confusion matrix, we'll go through three steps:\n",
"1. Make predictions with our trained model, `model_2` (a confusion matrix
compares predictions to true labels).\n",
"2. Make a confusion matrix using
[`torchmetrics.ConfusionMatrix`](https://torchmetrics.readthedocs.io/en/latest/
references/modules.html?highlight=confusion#confusionmatrix).\n",
"3. Plot the confusion matrix using
[`mlxtend.plotting.plot_confusion_matrix()`](http://rasbt.github.io/mlxtend/
user_guide/plotting/plot_confusion_matrix/).\n",
"\n",
"Let's start by making predictions with our trained model."
]
},
{
"cell_type": "code",
"execution_count": 53,
"id": "065b8090-c9c5-43df-b5c1-b45ba33af1be",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 49,
"referenced_widgets": [
"d3ab200da5f940d5b45396f83bd835e2",
"f35a13b3e55342aeb24b188c1d81a9e5",
"4a282c1974524bd3a7eba45fd3112129",
"44d4196e99a4412f893ba8ac4672915d",
"12d1a54d4107428eae2e64ff0a255c50",
"4d6eb654b2794b0a95f31ac94b52a4ca",
"fe5cff037f714657996f0541baee39f3",
"0670e3e758e6486b9cf4e2797b4b619a",
"3c590fc27b624584ba564e18bc42a2e4",
"629ca5b704b84a958d4ee477907f64a1",
"4d7c25dcdde8414382be4cf63a9cacf9"
]
},
"id": "065b8090-c9c5-43df-b5c1-b45ba33af1be",
"outputId": "92a8bee2-71f5-4504-d534-cc63138c413d"
},
"outputs": [
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "d3ab200da5f940d5b45396f83bd835e2",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Making predictions: 0%| | 0/313 [00:00<?, ?it/s]"
]
},
"metadata": {},
"output_type": "display_data"
}
 ],
"source": [
"# Import tqdm for progress bar\n",
"from tqdm.auto import tqdm\n",
"\n",
"# 1. Make predictions with trained model\n",
"y_preds = []\n",
"model_2.eval()\n",
"with torch.inference_mode():\n",
" for X, y in tqdm(test_dataloader, desc=\"Making predictions\"):\n",
" # Send data and targets to target device\n",
" X, y = X.to(device), y.to(device)\n",
" # Do the forward pass\n",
" y_logit = model_2(X)\n",
" # Turn predictions from logits -> prediction probabilities -> predictions
labels\n",
" y_pred = torch.softmax(y_logit, dim=1).argmax(dim=1) # note: perform
softmax on the \"logits\" dimension, not \"batch\" dimension (in this case we have
a batch size of 32, so can perform on dim=1)\n",
" # Put predictions on CPU for evaluation\n",
" y_preds.append(y_pred.cpu())\n",
"# Concatenate list of predictions into a tensor\n",
"y_pred_tensor = torch.cat(y_preds)"
]
},
{
"cell_type": "markdown",
"id": "362002d9-ec41-4c74-a210-b5d4f53410c4",
"metadata": {
"id": "362002d9-ec41-4c74-a210-b5d4f53410c4"
},
"source": [
"Wonderful!\n",
"\n",
"Now we've got predictions, let's go through steps 2 & 3:\n",
"2. Make a confusion matrix using
[`torchmetrics.ConfusionMatrix`](https://torchmetrics.readthedocs.io/en/latest/
references/modules.html?highlight=confusion#confusionmatrix).\n",
"3. Plot the confusion matrix using
[`mlxtend.plotting.plot_confusion_matrix()`](http://rasbt.github.io/mlxtend/
user_guide/plotting/plot_confusion_matrix/).\n",
"\n",
"First we'll need to make sure we've got `torchmetrics` and `mlxtend` installed
(these two libraries will help us make and visualize a confusion matrix).\n",
"\n",
"> **Note:** If you're using Google Colab, the default version of `mlxtend`
installed is 0.14.0 (as of March 2022), however, for the parameters of the
`plot_confusion_matrix()` function we'd like use, we need 0.19.0 or higher. "
]
},
{
"cell_type": "code",
"execution_count": 54,
"id": "e6c0a05d-d3e0-4b86-9ef7-ee6ea5629b07",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "e6c0a05d-d3e0-4b86-9ef7-ee6ea5629b07",
 "outputId": "b37df16c-c292-4347-807c-91c97bf81f20"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\u001b[2K \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \
u001b[32m519.2/519.2 kB\u001b[0m \u001b[31m10.8 MB/s\u001b[0m eta \
u001b[36m0:00:00\u001b[0m\n",
"\u001b[2K \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \
u001b[32m1.4/1.4 MB\u001b[0m \u001b[31m54.9 MB/s\u001b[0m eta \u001b[36m0:00:00\
u001b[0m\n",
"\u001b[?25hmlxtend version: 0.22.0\n"
]
}
],
"source": [
"# See if torchmetrics exists, if not, install it\n",
"try:\n",
" import torchmetrics, mlxtend\n",
" print(f\"mlxtend version: {mlxtend.__version__}\")\n",
" assert int(mlxtend.__version__.split(\".\")[1]) >= 19, \"mlxtend verison
should be 0.19.0 or higher\"\n",
"except:\n",
" !pip install -q torchmetrics -U mlxtend # <- Note: If you're using Google
Colab, this may require restarting the runtime\n",
" import torchmetrics, mlxtend\n",
" print(f\"mlxtend version: {mlxtend.__version__}\")"
]
},
{
"cell_type": "markdown",
"id": "5245ede6-fd7f-40ad-a0b3-ae678544b84a",
"metadata": {
"id": "5245ede6-fd7f-40ad-a0b3-ae678544b84a"
},
"source": [
"To plot the confusion matrix, we need to make sure we've got and [`mlxtend`]
(http://rasbt.github.io/mlxtend/) version of 0.19.0 or higher."
]
},
{
"cell_type": "code",
"execution_count": 55,
"id": "21383f88-a2dd-4678-94c6-479c592da0ab",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "21383f88-a2dd-4678-94c6-479c592da0ab",
"outputId": "faffbe4c-9c86-4a20-cbd6-c7e8e48e81a5"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"0.22.0\n"
 ]
}
],
"source": [
"# Import mlxtend upgraded version\n",
"import mlxtend \n",
"print(mlxtend.__version__)\n",
"assert int(mlxtend.__version__.split(\".\")[1]) >= 19 # should be version
0.19.0 or higher"
]
},
{
"cell_type": "markdown",
"id": "c91b9346-e25f-48ab-967e-425649331dc6",
"metadata": {
"id": "c91b9346-e25f-48ab-967e-425649331dc6"
},
"source": [
"`torchmetrics` and `mlxtend` installed, let's make a confusion matrix!\n",
"\n",
"First we'll create a `torchmetrics.ConfusionMatrix` instance telling it how
many classes we're dealing with by setting `num_classes=len(class_names)`.\n",
"\n",
"Then we'll create a confusion matrix (in tensor format) by passing our
instance our model's predictions (`preds=y_pred_tensor`) and targets
(`target=test_data.targets`).\n",
"\n",
"Finally we can plot our confusion matrix using the `plot_confusion_matrix()`
function from `mlxtend.plotting`."
]
},
{
"cell_type": "code",
"execution_count": 56,
"id": "7aed6d76-ad1c-429e-b8e0-c80572e3ebf4",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 667
},
"id": "7aed6d76-ad1c-429e-b8e0-c80572e3ebf4",
"outputId": "ae34ae74-2038-4037-f01d-77a807e4de9b"
},
"outputs": [
{
"data": {
"image/png":
"iVBORw0KGgoAAAANSUhEUgAAApYAAAKKCAYAAACH5hvqAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIH
ZlcnNpb24zLjcuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/
bCgiHAAAACXBIWXMAAA9hAAAPYQGoP6dpAADaa0lEQVR4nOzdd1QU198G8GfpvUkT6aAoKhZU7AUb9t5iQy
JGY0OsaGzYY8NujAVNYiyxIXYRCzbsPRbUYJSmSFdAlvcPftlkFZUyMLu8z+ecOcedvTP7XC47fHd25irJz
c3NBRERERFRMamIHYCIiIiIygYWlkREREQkCBaWRERERCQIFpZEREREJAgWlkREREQkCBaWRERERCQIFpZE
REREJAg1sQNQ8UmlUrx69Qr6+vqQSCRixyEiIqIyJjc3F6mpqbCysoKKyufPS7KwLANevXoFGxsbsWMQERF
RGffixQtYW1t/9nkWlmWAvr4+AECz3WJI1LVFTiOsv7YMFDtCiZBKy+Z/ePWhjPZLXbXsfRNQVr/
dSEzNFDtCiTDR1xQ7Qokoq8dCFZWy9/5KTUmBs4ONrOb4HBaWZcA/
fyAk6tplrrA0MDAQO0KJKKsHUxaWyqOsFpbZkrJZWBqwsFQqZbGw/
MfXjh28eYeIiIiIBMHCkoiIiIgEwcKSiIiIiATBwpKIiIiIBMHCkoiIiIgEwcKSiIiIiATBwpKIiIiIBMHC
koiIiIgEwcKSiIiIiATBwpKIiIiIBMHCkoiIiIgEwcKSiIiIiATBwpKIiIiIBMHCkoiIiIgEwcKSiIiIiAT
BwpKIiIiIBMHCkoiIiIgEwcKSiIiIiATBwpKIiIiIBFEmC8tZs2ahZs2aX2zTvHlz+Pn5lUoeRaeiIsGMvr
Vxf00vvPltEO6u7okpPWvItflpZBNk/OEjtxyY1kb2vK2ZHtaNaCy3jx9614K6muL/
iq1fuwYuzvYw0tNCk4YeuBIZKXakYktNTcXE8X6oXNEe5Qx14NmsEa5dvSJ2rELZuGEdGtatCWtzI1ibG6F
Vs0Y4ceyI7Pn3799jvN8o2Fcwg5WpAQb07Yn4uDgRExfN3MBZ0NFQkVtqVqsidqxiW7xoARrVrwszY33YWp
mjV4+uePTwodixCi0tNRUzAybAw60inKyM0KVtc9y8fjXftlP8R8HaRAsb160q5ZTCKUvHw5ycHATOmg7XS
o4oZ6iDapWdsXD+HOTm5oodTRCKOlYK8VdfIpF8cZk1a5bgr7l3717MmTPni22eP38OiUSCmzdv5vv87Nmz
MWDAAAB5fdi/f7/AKUvH+K7VMbRNZfhvuohafnvxw69XMa6LG0a0d5Vrd/
zG33AY+rtsGRx0WvacSwVDqEiA0RvOw33cPkwOvoxv21TG7G/cS7k3hbN7105MnuiPaT/
MxMXI63Bzq4HOHdoiPj5e7GjFMnK4L8LDTmLj5m2IvHYbLVu1Rsd2rfHq5UuxoxVYhQrWmDVnPs5cuILT5y
PRtHkL9OvVDQ/u3wMABEzyx9FDodj6204cOh6O2JgYDOjbU+TURePqWhVPo1/
JlpOnz4kdqdjOnT2D4SNG4kzEJYQeOYEP2dno2L4N0tPTxY5WKBPHjsC502FYsX4zTkZcQ9MWLdGvW3vEvJ
J/Lx0JPYDrVyNhUd5KpKTFV9aOh8uWLMLGDeuxLGgVrt+6jznzF2L50sVYt0Z5C/9/
KPJYKURhGRMTI1uCgoJgYGAgt27ChAmCv6aJiQn09fU/
+3xWVtZX93HgwAF07txZyFiiqO9ijkNXonH0+t+ITkjD/
kvPEXbrJeo4m8q1y8zOQVzSO9mSlP7vz+jEzZf4bm0Ewm69wvP4VBy6+gIrQu6gi4d9KfemcFYGLcOQb30x
yHsIqri6YtXa9dDW0cHW4M1iRyuyd+/eYf+
+PZg7fxEaN2kKJ2dnTJs+C45Ozvh5wzqx4xVYuw6d0MarPZycK8K5YiXMmD0Xunp6uBJ5CcnJyfgleDPmLV
qCZs09Uau2O9Zu2ITLly7gyuVLYkcvNFU1NVhaWsoWU1PTr2+k4EIOHcXAwd5wrVoVbjVqYMOmYLyIjsaN6
9fEjlZg7969w+GD+zBt9nzUb9gEDo5OGD9lOuwdnfDLlg2ydjGvXmL6ZH+s+ikY6mpqIiYunrJ2PLx08SI6
dOoMr/YdYGdvj27de6Jlqza4qmTf3uRHkcdKIQrL/
x5QDQ0NIZFI5Nbp6el9ss3p06dRr1496OrqwsjICI0aNcJff/0l1+aXX36Bvb09DA0N0bdvX6Smpsqe+/
ircHt7e8yZMweDBg2CgYEBhg0bBgcHBwBArVq1IJFI0Lx5c1n7Fy9e4N69e/
Dy8oK9vT0AoFu3bpBIJLLHALBu3To4OTlBQ0MDLi4u+OWXX+QySiQSrFu3Du3atYO2tjYcHR3xxx9/
FPEnWTSXHsajefXycC5vAACobmeCBpUtcPzG33LtmlS1xPNN/
XBzRQ+s8G0AEz3NL+7XUEcDb9MySyx3cWVlZeHG9WvwbNlKtk5FRQWenq0QeemiiMmK58OHD8jJyYGmlpbc
em1tbVy8cF6kVMWTk5ODP3btQEZ6Oup5NMDNG9eQnZ2N5p7/
jl0ll8qwsbFFpBIWllFPHsPRrgJcXZwwZNAAvIiOFjuS4FKSkwEAxsYmIicpuJx/
3kua8sc6LS0tRF66AACQSqUYO8IHw0ePg0sV1/
x2oxTK4vGwfoMGOB1+Co8fPQIA3L59CxcuRKBNWy+RkxWPoo+VQhSWhfXhwwd07doVzZo1w+3bt3Hx4kUMG
zYMEolE1iYqKgr79+9HaGgoQkNDcebMGSxcuPCL+12yZAlq1KiBGzduYPr06Yj83/
UKJ0+eRExMDPbu3StrGxISgubNm8PAwABXruR9+tmyZQtiYmJkj/ft24exY8di/
PjxuHv3Lr777jsMGTIE4eHhcq87ffp09OjRA7du3UL//v3Rt29fPHjw4LM5MzMzkZKSIrcUx5J9t7H7/
DPcXNEDyTu8cXFxF6w5dA87zz2VtTlx82/4rjqLDrOPYvqvV9DY1RL7p7WBiook3306WupjeDtXbDrxZ7Gy
laTXr18jJycH5uYWcuvNLSwQGxsrUqri09fXh0f9Bli0YC5iXr1CTk4Oft/
+Ky5fuojYmBix4xXKvbt3YGVqADNDbfiP+R6/7dyDylVcER8bCw0NDRgZGcm1NzO3QFycco1d3Xoe2LBxCw
4cPIIVq9bi+fNnaOXZVO6DsLKTSqWYON4PDRo2QtVq1cSOU2B6+vpwr1sfQUsWIDYm7720Z9d2XLtyGfH/
+z1bu2IJ1FTV8O13I0VOWzxl8Xg4fuIU9OzVB7XcqsBQVwMN69XGyNFj0bdff7GjFYuij5VSnrNPSUlBcnI
yOnbsCCcnJwBAlSryF7tLpVIEBwfLvu4eOHAgwsLCMG/evM/
u19PTE+PHj5c9VlVVBQCUK1cOlpaWcm0PHDiALl26AADMzMwAAEZGRnLtlixZAm9vb3z//fcAAH9/
f1y6dAlLlixBixYtZO169eqFoUOHAgDmzJmDEydOYNWqVVi7dm2+ORcsWIDZs2d/th+F1aOhA/
o2cYT3itN48CIJbvYm+HGIB2ISM/DbmScAgD/
OP5O1vxf9Fnf+eov7a3uhaVVLnL4jX6xYmejgwLS22HfxGbacfCRYTiq4jZu3YcR338LZwRqqqqqoWas2ev
Xph5tK9DUkAFSs5IJzl68jJTkZB/
btwXDfITh8PPzrGyqRtl7tZP+u7uaGuvU8UNnZHnv+2AXvId+KmEw4fqNH4t69uwg7HSF2lEJbsX4Txo/
+DnWqOkJVVRXVatRClx69cefmDdy+eR2bflqDI+EX5U5skGLY88cu7NyxHVu2/
YYqrlVx+9ZNTJ4wDuXLW2HAwMFixyuzFP6MZXR0NPT09GTL/
PnzYWJiAm9vb7Rt2xadOnXCihUrEPPRmRh7e3u5ayjLly//
1Yta69SpU6BMKSkpOHPmzFevr3zw4AEaNWokt65Ro0afnI1s0KDBJ4+/
dMYyICAAycnJsuXFixcFyv058wfWxdL9d/DH+We4F/
0Wv5+NwurQe5jQ3e2z2zyPT0VC8js4WRrIrS9vrI0js9rh0qN4jPxJsb92NTU1haqqKuLj5e8kjo+L+
+SDhLJxdHLCsZOnEZ+YiodR0Th7/
jI+ZGfD3sFR7GiFoqGhAScnZ9Sq7Y5Zc+ajWvUaWLdmJcwtLZGVlYWkpCS59gnxcbCwUO6xMzIygnPFSnj6
5InYUQThN2YUDh8OxbET4bC2thY7TqHZOzhhT+hJPHrxBpF3nuDQyQh8yP4AW3sHRF48j9cJ8fBwqwg7M13
Ymeni7xfRCJw+GfVrVBI7eqGUxePhtIBJGD9hMnr17otq1arjm/4DMWqMH5b+
+OVvLxWdoo+VwheWVlZWuHnzpmwZPnw4gLyvnS9evIiGDRti586dqFSpEi5d+vfaKnV1dbn9SCQSSKXSL76
Wrq5ugTIdOXIErq6usLGxKWRvhKGpqQkDAwO5pTi0NdUglcpPv5AjzYXKFz6BVzDRQTl9LcS+zZCtszLRwd
HZ7XHj6Rt8t+YcFH1GBw0NDdSq7Y7wU2GydVKpFOHhYahXv8EXtlQeurq6KF+
+PN6+fYuTJ46hYyflvtlMKpUiKzMTNWu5Q11dHWfC/
x27x48e4sWLaNTzqC9iwuJLS0vDs6dRsCxfXuwoxZKbmwu/MaMQcmAfjh4/Bfv/
XbOurHR0dWFhWR5JSW9x5tQJtGnXET36fIMT567i2JlI2WJR3grDR/
vjtz9CxY5cKGXxePguIwMqKvJljoqq6ldrAUWn6GOl8F+Fq6mpwdnZOd/
natWqhVq1aiEgIAANGjTA9u3bUb++cH9UNDQ0AOTdOPBf//0a/B/q6uqftKtSpQrOnz+PwYP/PeV+/
vx5uLrKX+B96dIlDBo0SO5xrVq1BOlDQRy++gKTetTAi9dpuP8iCTUdymF0x6rYFv4YAKCrpYapvWph/
6XniEt6B0dLfcwbUBdRsSk4cTNvyo28orIdohPSMXVbJMwM/
r1xJC7pXan1pbDG+PnD12cw3N3roE7deli9MggZ6ekYNHiI2NGK5cTxY8jNzUWlSi6IinqCaQGTUMmlMgYq
Ub9mTZ+K1m29YG1ji7TUVOze+Tsizp7G3oNHYGhoiIHePpg2eQKMTUygr2+ASf5jUc+jAeoqWWEZMHkC2nf
oBFtbO8TEvMLcwFlQVVVFrz79xI5WLH6jR2Lnju3YvfcA9PT1Zdd+GRoaQltbW+R0BXc67ARyc3PhVLEinj
+NwtyZU+FU0QV9+g+Guro6jE3KybVXV1ODubkFnCoq1xlLoOwdD9t16IQfF82HjY0tqrhWxa1bN7B6xXKlO
g5+jiKPlcIXlvl59uwZNmzYgM6dO8PKygoPHz7E48eP5YozIZibm0NbWxtHjx6FtbU1tLS0oKuriyNHjnwy
BZK9vT3CwsLQqFEjaGpqwtjYGBMnTkTv3r1Rq1YttGrVCgcPHsTevXtx8uRJuW13796NOnXqoHHjxvjtt98
QGRmJTZs2CdqXLxm/
6SJm9HVHkG9DmBloIeZtBjafeIj5f9wEkHf2spqdMfo3d4aRjgZi3mYg7NYrBO64hqwPeZ/
8PN2s4FzeEM7lDfFkQ1+5/ev0FH/6g8/
p1bsPXickIHD2DMTFxsKtRk0cCD0KCwuLr2+swFJSkjHzh6l4+fJvGJuYoGvX7pgZOO+TM/mKLCEhHsO/
9UZsbAwMDA1RtZob9h48As+WrQEAC35cBhUVFQzs1wtZmZnwbNUGy1asETl14b38+yUGD/wGiW/
ewNTMDA0bNsbpcxdl124rqw0/5U1t1aZlc/n1G7dg4GDv0g9URKkpyVg4ZzpiXr2EkbEJ2nXqisk/
zFaq91JBlbXj4dLlKxE4azr8xo5EQnw8ype3gs/QYQiYNkPsaMWmyGMlyVWwKeiDg4Ph5+f3ybVT/
xUXF4fhw4fj8uXLePPmDcqXL4/Bgwdj5syZUFFRwaxZs7B//
365ic2DgoIQFBSE58+fA8ibbqhmzZoICgoCkFcY+vn5ffK/8WzcuBGBgYF4+fIlmjRpgunTp8Pb2/
uT6xoPHjwIf39/PH/+HBUqVJC9zrp167BkyRK8ePECDg4O+OGHHzBw4EDZdhKJBGvWrMH+/
ftx9uxZlC9fHosWLULv3r0L/
DNLSUmBoaEhtDqvhkRdec4EFETiDh+xI5SIjy89KCs+lNF+qauWvRszyurNJm9SFXeKs+Iop//
l6d2UVVk9Fn5uxhRllpKSAotyhkhOTv7iJXgKV1gqujFjxuDDhw+fvWO7sCQSCfbt24euXbsWeR8sLJVPWT
2YsrBUHiwslQsLS+Xy/
7mwVMqvwsVUrVq1T+7iJiIiIiIWloU2bNgwsSMQERERKSQWliLjlQhERERUVij8PJZEREREpBxYWBIRERGR
IFhYEhEREZEgWFgSERERkSBYWBIRERGRIFhYEhEREZEgWFgSERERkSBYWBIRERGRIFhYEhEREZEgWFgSERE
RkSBYWBIRERGRIFhYEhEREZEgWFgSERERkSBYWBIRERGRIFhYEhEREZEgWFgSERERkSBYWBIRERGRINTEDk
DCeb55AAwMDMSOISjjhuPFjlAiEs8vETtCiZCIHaCESCRltWdlTzl9TbEjUCGoqPC9VdbwjCURERERCYKFJ
REREREJgoUlEREREQmChSURERERCYKFJREREREJgoUlEREREQmChSURERERCYKFJREREREJgoUlEREREQmC
hSURERERCYKFJREREREJgoUlEREREQmChSURERERCYKFJREREREJgoUlEREREQmChSURERERCYKFJRERERE
JgoUlEREREQmChSURERERCYKFJREREREJgoUlEREREQmChSUV2pIfF0JHQwUTx/
uJHeWL9HQ0sXhcFzw8MA2JZxcifONouFexkWvjYm+O3Ut8EHtqLl6fmY+I4LGwsTCSPe/
TtT6OrRuBuFPz8C5yKQz1tEq5F8WnLOP1saWLF6JZIw9YmRnC0dYS/Xp1w+NHD/
Ntm5ubi+5d2sNAWxWhIftLN2gxbVi/
DnVrucHcxADmJgZo1rgBjh09InYswaxfuwYuzvYw0tNCk4YeuBIZKXakYlm8aAEa1a8LM2N92FqZo1ePrnj
0MP/fS2XE8VIeijpWZbKwlEgkX1xmzZoldkSldfXqFWzauAHVq7uJHeWr1k3rDU+PSvCZ9TvqfLMYJy8/
xKE138HKzAAA4FChHMJ+HoVHf8Wj7fB1qPvNUizYdBLvsz7I9qGjpY4TFx9icXCYWN0oFmUar49FnDuDYcN
HIOzMBRwIPYbsD9no2tEL6enpn7Rds2oFJBKJCCmLr4K1NebMX4gLl6/h/KWraN7CE726d8H9e/
fEjlZsu3ftxOSJ/
pj2w0xcjLwON7ca6NyhLeLj48WOVmTnzp7B8BEjcSbiEkKPnMCH7Gx0bN8m399LZcPxUh6KPFaS3NzcXLFD
CC02Nlb27507d2LGjBl4+J9PKHp6etDT0wOQd6YjJycHampqpZ7za7KysqChofHVdikpKTA0NETs6yQYGBi
UWJ60tDQ0rOeOoFVrsGjBPLjVqIHFS4NK7PUAwKTRhCJtp6WphoTw+eg1cQuOnn8gW39+qx+OX/
wTs9cfxba5A5D9IQffzvr9q/trUtsJx9d/
D0vPaUhOe1+kTP+VeH5JsffxNWKM14eckjucvE5IgKOtJY6cCEejxk1l62/
fuone3TvjzPlIVHSogO0796Bj566Cvra6Wul+BrcyN8H8hYvh7fNtqb6u0Jo09IB7nboIWrkaACCVSuHsYI
MRI0dj4qQpIqcTRkJCAmytzHHi1Bk0btL06xsoMI6X8hBjrFJSUmBRzhDJyclfrDXK5BlLS0tL2WJoaAiJR
CJ7/Oeff0JfXx9HjhyBu7s7NDU1ERERgczMTIwZMwbm5ubQ0tJC48aNceXKFdk+g4ODYWRkJPc6+/
fvlztLcuvWLbRo0QL6+vowMDCAu7s7rl69Kns+IiICTZo0gba2NmxsbDBmzBi5T0329vaYM2cOBg0aBAMDA
wwbNqzkfkhFMG7MKHi1bw/Plq3EjvJVaqqqUFNTlTv7CADvMz+gYQ0HSCQSeDWqgsfRCQhZOQx/
HZ2Fs5vHoFOzaiIlFp4yjVdBJKckAwCMjU1k6zIyMvCt9wAsDVoFC0tLsaIJJicnB7t27kB6ejo86jcQO06
xZGVl4cb1a3K/
fyoqKvD0bIXISxdFTCaslORPfy+VEcdLeSj6WJXJwrIgpkyZgoULF+LBgwdwc3PDpEmTsGfPHmzduhXXr1+
Hs7Mz2rZti8TExALvs3///
rC2tsaVK1dw7do1TJkyBerq6gCAqKgoeHl5oUePHrh9+zZ27tyJiIgIjBo1Sm4fS5YsQY0aNXDjxg1Mnz49
39fJzMxESkqK3FLSdu/
cgZs3riNw7oISfy0hpGVk4tLt5wjwaYXypgZQUZGgr1dteFS3g6WpAcxN9KCvq4UJgz1x4uKf6DR6A0JO38
WORYPRuJaj2PGLTdnG62ukUimmTByH+g0awbXqv8V/wCR/
eNRvgA6duoiYrvju3rkDUyM9GOpqYszI4dj5xz5UcXUVO1axvH79Gjk5OTA3t5Bbb25hIfetkjKTSqWYON4
PDRo2QtVqyv2hlOOlPBR9rBTv+99SEhgYiNatWwMA0tPTsW7dOgQHB6Ndu3YAgJ9//
hknTpzApk2bMHHixALtMzo6GhMnTkTlypUBABUrVpQ9t2DBAvTv3x9+fn6y51auXIlmzZph3bp10NLKuynE
09MT48eP/
+LrLFiwALNnzy5Uf4vj7xcvMHG8Hw4ePi7LqQx8Zm7HT9P74OnhmfjwIQc3H77EruM3UKuyNVT+d6Y59Ow9
rPr9LADg9uNX8HCzh2/3hoi48VTM6MWirOP1JeP9RuHBvXs4FnZWtu5waAjOnA5HxKVrIiYTRiUXF1y+ehP
JycnYt/
cP+PoMxvGwM0pfXJZ1fqNH4t69uwg7HSF2FCoAjlfp+H97xrJOnTqyf0dFRSE7OxuNGjWSrVNXV0e9evXw4
MGD/DbPl7+/
P4YOHYpWrVph4cKFiIqKkj1369YtBAcHy67v1NPTQ9u2bSGVSvHs2bN8c31OQEAAkpOTZcuLFy8KnLEorl+
/hvj4eDT0cIe+tjr0tdVx7uwZrF29Cvra6sjJySnR1y+qZy/
foM3wtSjXNAAVO81BkyEroK6mimcv3+B1UjqyP+TgwbM4uW0ePo+DjaWROIEFoqzj9Tnj/
Ubj6OFDCD0WhgrW1rL1Z06H49nTKNhYmsBYTwPGennXIw/
o1wvt23iKFbdINDQ04OTsjNru7pgzbwGqu9XAmlUrxI5VLKamplBVVUV8vPx7LD4uDpZl4LIFvzGjcPhwKI
6dCIf1f34vlRXHS3ko+lj9vy0sdXV1C9VeRUUFH9/nlJ2dLfd41qxZuHfvHjp06IBTp07B1dUV+/
btA5B3I8V3332HmzdvypZbt27h8ePHcHJyKlQuTU1NGBgYyC0lqYVnS1y5fhuXrtyQLbXd66Bvv/
64dOUGVFVVS/T1iyvjfRZi36TCSF8breq7IPTsPWR/yMG1+y9QydZMrm1FWzNEx74VKakwlH28/
pGbm4vxfqMRGrIfB4+ehL29g9zz/hMm4+KVmzh/
+bpsAYAFPy7D2g2bxIgsGKlUiszMTLFjFIuGhgZq1XZH+Kl/
Z1SQSqUIDw9DPSW+fjQ3Nxd+Y0Yh5MA+HD1+CvYODl/fSAlwvJSHoo/V/9uvwv/LyckJGhoaOH/
+POzs7ADkFY1XrlyRfXVtZmaG1NRUpKeny4q/
mzdvfrKvSpUqoVKlShg3bhz69euHLVu2oFu3bqhduzbu378PZ2fn0uqWYPT19T+5HkVXVxcm5UwU+jqVVvV
dIAHwKDoBTtammD+mIx49j8e
2g3lzfS3/NRy/
zBuIiBtPcebaE7RpUBntG7ui7Yh1sn1YlNOHhYk+nGxMAQDVnMsjNT0TL+Le4m3KOzG69VXKOl4f8/
cbhT92/o7fd+
+Dvp4+4v537ZCBoSG0tbVhYWmZ7w07NjY2nxShimz6tAC09WoHGxtbpKamYueO7Th75jQOHj4mdrRiG+PnD
1+fwXB3r4M6deth9cogZKSnY9DgIWJHKzK/0SOxc8d27N57AHr6+rJr2gz/
93upzDheykORx4qFJfL+6I4YMQITJ06EiYkJbG1t8eOPP+bdcfpt3nQfHh4e0NHRwdSpUzFmzBhcvnwZwcH
Bsn28e/cOEydORM+ePeHg4IC///4bV65cQY8ePQAAkydPRv369TFq1CgMHToUurq6uH//
Pk6cOIHVq1eL0e0yz1BPC4Hft0cFcyMkpmTgwKnbmLnuCD7kSAEAIafvYvTCPZg42BNLx3fDo+h49JuyFRd
u/XtpwtDuDfCDb1vZ45Mb8m628p29A78eugIqOZs2rAeAT77WXrdhE/oP9BYhUclIiI/
Ht0MGITYmBoaGhqhW3Q0HDx9Dy1atxY5WbL1698HrhAQEzp6BuNhYuNWoiQOhR2FhYfH1jRXUhp/
yPni2adlcfv3GLRg42Lv0AwmI46U8FHmsyuQ8lv8VHBwMPz8/JCUlAQBOnz6NFi1a4O3bt3LTB71//
x6TJk3C77//
jtTUVNSpUwfLly9H3bp1ZW3279+PiRMn4uXLl2jZsiU6d+6MYcOGITc3F1lZWRg8eDDOnz+PuLg4mJqaonv
37li8eLHsBoorV65g2rRpuHjxInJzc+Hk5IQ+ffpg6tSpAPKmG/
Lz85OdJS2o0prHUgxFncdS0ZXGPJZiKMl5LMVU2vNYEhEpmoLOY1nmC8v/D1hYKh8WlsqFhSUR/X/
3/3qCdCIiIiIqfSwsiYiIiEgQLCyJiIiISBAsLImIiIhIECwsiYiIiEgQLCyJiIiISBAsLImIiIhIECwsiY
iIiEgQLCyJiIiISBAsLImIiIhIECwsiYiIiEgQLCyJiIiISBAsLImIiIhIECwsiYiIiEgQLCyJiIiISBAsL
ImIiIhIECwsiYiIiEgQLCyJiIiISBAsLImIiIhIECwsiYiIiEgQamIHIOFIc/
OWsuTthaViRygRVSYeEjtCibi3qL3YEUrEizcZYkcQnE05HbEjlIg3aVliRygR5fQ0xI5AVCA8Y0lEREREg
mBhSURERESCYGFJRERERIJgYUlEREREgmBhSURERESCYGFJRERERIJgYUlEREREgmBhSURERESCYGFJRERE
RIJgYUlEREREgmBhSURERESCYGFJRERERIJgYUlEREREgmBhSURERESCYGFJRERERIJgYUlEREREgmBhSUR
ERESCYGFJRERERIJgYUlEREREgmBhSURERESCYGFJRERERIJgYfk/9vb2CAoKkj2WSCTYv3+/
aHnEFHHuLHp16wxn+wrQ01TBwQP75Z6fN2cWalWvAnNjPVhbmKCjV2tcibwsTthiWLxoARrVrwszY33YWpm
jV4+uePTwodixvurc9BZ4trzDJ0tgj6qftN0yrC6eLe+A1tUs5NZbGWlhk29d3F/
khSuBrRDQqTJUVSSl1YUiS01NxcTxfqhc0R7lDHXg2awRrl29InasL7pyMQLfDeyJxjWcUMlSFyeOHPxs2x
mTxqCSpS6CN6zO9/mszEx0blkflSx1cf/
urZKKLKj1a9fAxdkeRnpaaNLQA1ciI8WOVChpqamYGTAeHtUrwqm8Ibq0aYab16/Kns/
NzcXi+bNRu7IdnMobom9XLzyNeixi4uJR9vH6mLIe5wtCUceqzBSW3t7ekEgkkEgk0NDQgLOzMwIDA/
HhwwexoymdjPR0VHNzw7IV+f9xq1ixEpYFrcLla7dxPPwc7Ozt0KVDWyQkJJRy0uI5d/
YMho8YiTMRlxB65AQ+ZGejY/s2SE9PFzvaF3VZdh51Z5yULQPWXQIAHLoZI9fOp5kDcnM/
3V5FAmzyrQsNVQl6rLiACdtvoUc9a4zzqlQa8Ytl5HBfhIedxMbN2xB57TZatmqNju1a49XLl2JH+6yMjHR
UrlodMxYs/2K744dDcPNaJMwty3+2zY9zpsHc4vPPK5rdu3Zi8kR/
TPthJi5GXoebWw107tAW8fHxYkcrsIljh+Pc6TCsWL8ZJ89fQ1PPVujXtR1iXuX9zq1dsRRbflqDBctW4eC
JCOjo6GJAj454//69yMkLryyM18eU9Tj/NYo8VmWmsAQALy8vxMTE4PHjxxg/
fjxmzZqFxYsXix2ryLKyskR53TZe7TBz9lx07tIt3+d79/0GLVq2goOjI1xdq2LBj8uQkpKCu3dul3LS4gk
5dBQDB3vDtWpVuNWogQ2bgvEiOho3rl8TO9oXJaZn4XVqpmzxdLXA84R0XI5KlLWpYmWAoc0dMGnHp2PSxM
UMFS31Me7Xm3jwKgVn/kzAsiOPMLCxHdRVFfes5bt377B/
3x7Mnb8IjZs0hZOzM6ZNnwVHJ2f8vGGd2PE+q1nLthg3ZSbatO/
82TaxMa8wZ9p4LF2zGepq6vm2ORN2DBFnTmHKzPklFVVwK4OWYci3vhjkPQRVXF2xau16aOvoYGvwZrGjFc
i7d+9wOGQfps2aj/
qNmsDB0Rnjp0yHvaMTftm8Abm5udi0fhXGTJiCtu07w7VadQSt24y42BgcOxQidvxCU/bxyo+yHue/
RpHHqkwVlpqamrC0tISdnR1GjBiBVq1aISQkBM2bN4efn59c265du8Lb27vA+75z5w48PT2hra2NcuXKYdi
wYUhLSwMAHD9+HFpaWkhKSpLbZuzYsfD09JQ9joiIQJMmTaCtrQ0bGxuMGTNG7lOTvb095syZg0GDBsHAwA
DDhg0r9M+gtGVlZWHLxg0wNDREdbcaYscplpTkZACAsbGJyEkKTl1Vgq7uFbA78oVsnZa6ClYMrImZe+7hd
WrmJ9vUtjfGw5gUvE7794PL2T8TYKCtjoqW+qWSuyg+fPiAnJwcaGppya3X1tbGxQvnRUpVfFKpFJNGfYuh
3/uhYmXXfNu8TojDDxNGYfGqjdDS1inlhEWTlZWFG9evwbNlK9k6FRUVeHq2QuSliyImK7icz/zOaWlpI/
LSBUT/9QzxcbFo0ryl7DkDQ0PUdK+Ha1culXbcYikL41UQynic/5iij1WZKiw/pq2tLchZv/
T0dLRt2xbGxsa4cuUKdu/ejZMnT2LUqFEAgJYtW8LIyAh79uyRbZOTk4OdO3eif//
+AICoqCh4eXmhR48euH37Nnbu3ImIiAjZPv6xZMkS1KhRAzdu3MD06dPzzZOZmYmUlBS5pbQdORQKCxN9lD
PQxupVQQg5fBympqalnkMoUqkUE8f7oUHDRqharZrYcQqsTXVLGGir4Y/
Iv2Xrpnd1xfXnb3Hibly+25gZaOJ1qvz74p8C1Exfs+TCFpO+vj486jfAogVzEfPqFXJycvD79l9x+dJFxM
bEfH0HCmrD6qVQVVPDoKHf5/t8bm4uJo/5Dv0GDUX1mrVLOV3RvX79Gjk5OTA3l7+
+19zCArGxsSKlKhw9fX24162PoMULEBuT9zu3Z+d2XLtyCfFxMUiIy3uPmZqZy21nZm6OhPj833+KqiyM19
co63H+Y4o+VmWysMzNzcXJkydx7NgxuTOGRbV9+3a8f/
8e27ZtQ7Vq1eDp6YnVq1fjl19+QVxcHFRVVdG3b19s375dtk1YWBiSkpLQo0cPAMCCBQvQv39/+Pn5oWLFi
mjYsCFWrlyJbdu2yV2L4+npifHjx8PJyQlOTk755lmwYAEMDQ1li42NTbH7WFhNm7fAhcgbCDtzHq3btMWg
b/ooxLUdReU3eiTu3buLbb/tEDtKofT2sMGZPxMQn5JXGLaqao4GFU0RuO+
+yMlKxsbN25CbmwtnB2sY62th3ZpV6NWnH1RUlPNQdvfWDWz7eS0WrtgAiST/
yxB+2bQO6elp+G7MhFJORwCw4qfNyM3NRR1XBzha6GPzhjXo0qOP0v7O/
X+mrMd5ZVOm3hmhoaHQ09ODlpYW2rVrhz59+mDWrFnF3u+DBw9Qo0YN6OrqytY1atQIUqkUD/93d1n//
v1x+vRpvHr1CgDw22+/
oUOHDjAyMgIA3Lp1C8HBwdDT05Mtbdu2hVQqxbNnz2T7rVOnzlfzBAQEIDk5Wba8ePHiq9sITVdXF07Ozqj
nUR9rf9oENTU1bAveVOo5hOA3ZhQOHw7FsRPhsLa2FjtOgVUw1kajSqbYeenf8W9Q0RR25XRwa34bPF7SDo
+XtAMArBvijt9H1gcAJKRkwlRfQ25fpv87U5mQz1fnisTRyQnHTp5GfGIqHkZF4+z5y/
iQnQ17B0exoxXJ1cvn8eZ1Apq7u6BKBQNUqWCAl39HY+GsALSoUwUAcDHiDG5evYxqtsaoUsEArRtUBwD0a
NsEk0b7ihn/i0xNTaGqqor4j87cxcfFwdLSUqRUhWfv4IQ9h07i0d+JiLwbhUNh5/
HhQzZs7RxgZpF3xuh1gvyH6oT4eJh9dDZJ0ZWV8focZT3O50fRx0pN7ABCatGiBdatWwcNDQ1YWVlBTS2ve
yoqKsj96PbY7OxsQV+7bt26cHJywo4dOzBixAjs27cPwcHBsufT0tLw3XffYcyYMZ9sa2trK/
v3f4vXz9HU1ISmpmJ9ZSmVSpGZqdhFycdyc3MxbuxohBzYh+MnT8PewUHsSIXSs5413qRl4tT9f/
+orQuLws5L0XLtjk1uhrn77+PkvbyD0PXnbzGytTPK6Wngzf+us2ziYoqUd9l4EptWeh0oBl1dXejq6uLt2
7c4eeIY5s5fJHakIunSsx8aNmkht86nXxd06dkPPfoOBABMn7sE4ybPkD0fHxcDn75dEPTTNtSoXbdU8xaG
hoYGatV2R/
ipMHTu0hVA3nEiPDwMw78f9eWNFZCOri50dHWRlPQWZ8JOYOrs+bC1c4C5hSUizpxC1ep515inpqTg5rVID
PJR/Gvk/6usjdc/lP04nx9FH6syVVjq6urC2dn5k/
VmZmaI+c81WDk5Obh79y5atGjxSdv8VKlSBcHBwUhPT5cVfufPn4eKigpcXFxk7fr374/
ffvsN1tbWUFFRQYcOHWTP1a5dG/fv3883n6JJS0vD06gnssd/
PX+G27duwtjYBCblymHxwnlo37EzLC3L482b19iwfg1evXqJbj16iZi68PxGj8TOHduxe+8B6Onry65NMTQ
0hLa2tsjpvkwiAXrVs8aeK38jR/rvh6Z/7hT/2Mu37/B34jsAwLmHCXgcm4pl/Wti4cEHMNPXhH87F/
wS8ReycqSl1oeiOHH8GHJzc1Gpkguiop5gWsAkVHKpjIGDh4gd7bPS09Pw17Mo2eO/o5/j/
t1bMDIygZW1DYxNysm1V1dTh5m5BRyd86Z/
srKWv9RFR1cPAGBj7wBLqwolnL54xvj5w9dnMNzd66BO3XpYvTIIGenpGKTA4/
Wx02HHkZubC6eKlfD8aRTmzgiAUyUX9Ok/GBKJBN8OH42VSxbCwdEZNnYOWDJ/Fiwsy6Nth8/
PAqCoysJ4fUyZj/NfoshjVaYKy8/x9PSEv78/Dh06BCcnJyxbtuyTO7i/pH///
pg5cyYGDx6MWbNmISEhAaNHj8bAgQNhYWEh127WrFmYN28eevbsKXdWcfLkyahfvz5GjRqFoUOHQldXF/
fv38eJEyewenX+80WK5fq1q2jf5t9rU6dMGg8A6D9wMFasXoeHDx/it1974s3r1zApVw7u7nVx/
NRZuLp+OkG3ItvwU94UNW1aNpdfv3ELBg72Lv1AhdC4kikqmOhg9+W/
v974I9JcYOjGq5jTsxr2jG2EjKwP2HvlJZYffVQCSYWVkpKMmT9MxcuXf8PYxARdu3bHzMB5UFfPf4oeRXD
35nUM7NFO9njBzCkAgG69+2PRyg1ixSoVvXr3weuEBATOnoG42Fi41aiJA6FH5Y6bii41JQULA39AzKuXMD
I2QbtOXTH5h0DZ79z3Y8cjIyMdk8eNREpyEurWb4hf/zgIrY/uJFcGZWG8PqbMx/kvUeSxkuR+/
B2xkvL29kZSUlK+/
1tOdnY2xo4di507d0JNTQ3jxo3DpUuXYGRkJPu62t7eHn5+frJpiSQSCfbt24euXbsCyJtuaOzYsbh48SJ0
dHTQo0cPLFu2DHp6enKv5eHhgcjISJw6deqTM6JXrlzBtGnTcPHixbxPwE5O6NOnD6ZOnZpvhoJKSUmBoaE
hXiUkwcDAoFDbKjpl+N9giqLKxENiRygR9xa1FztCiXj59p3YEQRnU045pi0qrDdp4sz/
W9LK6Wl8vRFRCUpJSYFFOUMkJyd/sdYoM4Xl/
2csLJUPC0vlwsJSebCwJCoZBS0sy9Rd4UREREQkHhaWRERERCQIFpZEREREJAgWlkREREQkCBaWRERERCQI
FpZEREREJAgWlkREREQkCBaWRERERCQIFpZEREREJAgWlkREREQkCBaWRERERCQIFpZEREREJAgWlkREREQ
kCBaWRERERCQIFpZEREREJAgWlkREREQkCBaWRERERCQIFpZEREREJAgWlkREREQkCBaWRERERCQINbEDkH
Byc3ORm5srdgxBlbHuyNz/sb3YEUqESYelYkcoEbEHxokdgQpIT1NV7AhE/6/
xjCURERERCYKFJREREREJgoUlEREREQmChSURERERCYKFJREREREJgoUlEREREQmChSURERERCYKFJRERER
EJgoUlEREREQmChSURERERCYKFJREREREJgoUlEREREQmChSURERERCYKFJREREREJgoUlEREREQmChSURE
RERCYKFJREREREJgoUlEREREQmChSURERERCYKFJREREREJgoUlEREREQmChSV9YsmPC9GskQfKmxrCwcYS
fXt1w6NHD+XavH//
Hv5jR8HWygyW5QzQv29PxMfFiZS4aOYGzoKOhorcUrNaFbFjFVvlig6f9EtHQwV+Y0aKHe2L9LTVsXh4Czz
8ZRgSD45F+PJ+cK9kKddm+qBGePr7cCQeHItDC3vBycpI9pythQHW+bfFg22+SDw4FveCh+KHgQ2hrqZYh7
lNG9ajYb1asLEwho2FMVo3b4QTx4580i43Nxc9u3SAkY4aQkMOiJBUGOvXroGLsz2M9LTQpKEHrkRGih2pw
L40Vm8TEzHRfyzq1HCFpYkeqlVywKTxfkhOThY5dfEo83h9TlnsE6C4/
VKsI66IvL29IZFIIJFIoK6uDgsLC7Ru3RqbN2+GVCoVO16pOn/uDHy/
G4FTZy8g5NAxZGdno2sHL6Snp8vaTJnojyOHQvHLbztx5EQ4YmJi8E2fniKmLhpX16p4Gv1Ktpw8fU7sSMV
27kKkXJ9CjxwHAHTv0UvkZF+2blxbeNa2g8+Ph1Hnu604ef0vHFrUC1bl9AAA43vXw/
dda2HMyhNoOuY3pL/PxsEFPaGprgoAcLExgYpEglErjqO2bzAmrQ/
H0I41EDikiZjd+oRVhQqYFTgPp89HIjziMpo2a4FvenfHg/
v35NqtXb0CEolEpJTC2L1rJyZP9Me0H2biYuR1uLnVQOcObREfHy92tAL50ljFxLxCbMwrzJm/
CBev3sKaDZsQduIYRo/
wFTt2kSn7eOWnLPYJUOx+SXJzc3PFDqEIvL29ERcXhy1btiAnJwdxcXE4evQoFixYgCZNmiAkJARqamqfbJ
ednQ11dXUREv8rJSUFhoaGeBn/
FgYGBoLvPyEhAY42ljhyIhyNmzRFcnIyHKwtsHnrr+jaPa+YfPjwT9SpURVhZ86jnkd9wV5bVaXk/
rDODZyFgyEHcPnqjRJ7DUUwcbwfjhw+hDv3H5V4oWLSYWmRttPSUEPCgTHoNXM/jkY+la0/
v2YAjl95htnB5/
H09+FYuecqgv64CgAw0NHAX7u+x7AlR7D79MN89zuuV134dqwB18Ebi5TrH7EHxhVr+6+xr2CGwHmLMMjbB
wBw+9ZN9O3RBeERl+HiaI1fd+xBx85dBH3NfwryktSkoQfc69RF0MrVAACpVApnBxuMGDkaEydNKZHXzMzO
KZH9/uPjsfqv/Xv/wDCfQXj1OiXfvxfFUVbHq6SVxT4B4vQrJSUFFuUMkZyc/MVag2cs/
0NTUxOWlpaoUKECateujalTp+LAgQM4cuQIgoODAQASiQTr1q1D586doauri3nz5gEADhw4gNq1a0NLSwuO
jo6YPXs2Pnz4ACDvK61Zs2bB1tYWmpqasLKywpgxY2Svu3btWlSsWBFaWlqwsLBAz56KdeYvJSXvqx0TExM
AwM3r15CdnY3mnq1kbVxcKsPGxhaRly+JkrGoop48hqNdBbi6OGHIoAF4ER0tdiRBZWVlYcf23zBo8BCFPv
ulpiqBmqoK3md9kFv/PvMDGla1hr2lIcqX08Op63/JnkvJyMKVP2PgUcXqs/
s10NVAYur7EstdXDk5Odizeycy0tNlH8gyMjLgO2QgFi9fBQtLy6/
sQXFlZWXhxvVr8Gz573FCRUUFnp6tEHnpoojJiia/
sfpYSnIy9A0MBC8qS0NZGy+gbPYJUPx+Kd9vfynz9PREjRo1sHfvXgwdOhQAMGvWLCxcuBBBQUFQU1PDuXP
nMGjQIKxcuRJNmjRBVFQUhg0bBgCYOXMm9uzZg+XLl2PHjh2oWrUqYmNjcevWLQDA1atXMWbMGPzyyy9o2L
AhEhMTce7cl7+OzczMRGZmpuxxSkpKCfU+71PQ5AnjUL9BI7hWrQYAiIuLhYaGBoyMjOTamltYIC4utsSyC
K1uPQ9s2LgFFSu5IDY2BvPnBqKVZ1NcvXEH+vr6YscTxMED+5GUlIQBg7zFjvJFae+yceneSwT0b4CH0W8Q
l5SB3i0qw6OKFaJeJcHSRBcAEJ+UIbdd/NsMWBjr5rtPRysjjOhSGwEbTpd0/EK7d/
cO2rRojPfv30NXTw+/7vgDlau4AgCmThqPeh4N0KFTZ5FTFs/
r16+Rk5MDc3MLufXmFhZ4+PBPkVIV3pfG6r/evH6NHxfOg/eQoSKkLL6yMl7/
VRb7BCh+v1hYFkDlypVx+/
Zt2eNvvvkGQ4YMkT328fHBlClTMHjwYACAo6Mj5syZg0mTJmHmzJmIjo6GpaUlWrVqBXV1ddja2qJevXoAg
OjoaOjq6qJjx47Q19eHnZ0datWq9cU8CxYswOzZs0ugp5/yHzsKD+7dw/
FTZ0vl9UpTW692sn9Xd3ND3XoeqOxsjz1/7IL3kG9FTCacrcGb0aZtO1hZff6snqLw+fEwfhrvhac7RuBDj
hQ3H8dh1+k/Uauixdc3/
ohVOT2EzOuBvWcfYsuROyWQtngqVnLBuUvXkJKcjAP792DEMB8cOnYKT6OicPZMOM5evCp2RPqfz43Vf4vL
lJQU9O7eCZUrV8GUH2aKmJZIfCwsCyA3N1fua8Q6derIPX/
r1i2cP39e9rU4kPe1yfv375GRkYFevXohKCgIjo6O8PLyQvv27dGpUyeoqamhdevWsLOzkz3n5eWFbt26QU
dH57N5AgIC4O/
vL3uckpICGxsbAXucZ7zfaBw9fAhHT55GBWtr2XoLC0tkZWUhKSlJ7qxlfFwcLCyU96s7IyMjOFeshKdPno
gdRRDRf/
2FU2En8fuuPWJHKZBnMcloM2EndLTUYaCjgdjEdPwytSOexSQjNjHvxjFzIx3ZvwHA3FgHt6PkL1Yvb6KLo
4t749L9VxgZdLxU+1BQGhoacHRyBgDUrO2O69euYv2aVdDS1sazp1GwK19Orv2gb3qhQaPGOHTslBhxi8TU
1BSqqqqIj5efLSI+Lg6WSvQV/+fGKmj1OgBAamoqenZpDz19ffy6c4/o19wXVVkZr/
8qi30CFL9fvMayAB48eAAHBwfZY11d+a/
e0tLSMHv2bNy8eVO23LlzB48fP4aWlhZsbGzw8OFDrF27Ftra2vj+++/RtGlTZGdnQ19fH9
evX8fvv/+O8uXLY8aMGahRowaSkpI+m0dTUxMGBgZyi5Byc3Mx3m80DobsR+ixk7D/
T9+BvIOruro6zoSHydY9evQQL15EC3rjTmlLS0vDs6dRsCxfXuwogti2dQvMzM3Rrn0HsaMUSsb7bMQmpsN
ITxOt6tgj9OITPI9NRsybNLSoZSdrp6+jgbqVy+Pyg1eydVbl9HBsSR/
ceByHYUuPQlluTZRKpcjMysS48ZNwPvIGzl26JlsAYP6PS7Hmp00ipywcDQ0N1KrtjvBT/
x4npFIpwsPDUK9+AxGTFc8/
YwXkfajv3skL6hoa+H33fmhpaYmcrujK4niVxT4Bit8vnrH8ilOnTuHOnTsYN+7zd4XWrl0bDx8+hLOz82f
baGtro1OnTujUqRNGjhyJypUr486dO6hduzbU1NTQqlUrtGrVCjNnzoSRkRFOnTqF7t27l0SXvsp/
7Cjs3vk7duzeB309fcTF5l03aWBoCG1tbRgaGmKQtw8CJk2AsbEJ9A0MMMF/
LOrVb6BUhWXA5Alo36ETbG3tEBPzCnMDZ0FVVRW9+vQTO1qxSaVS/LItGAMGDFKaGwlaudtDIgEe/
f0WTlZGmO/bDI9eJGLbsbsAgDX7rmPyN/
Xx5OVbPI9NxkzvRoh5k4aQ83lnmP8pKqPjUhCw4QzMDLVl+457m5Hva4ph9oypaNXGC9Y2tkhLTcUfu35Hx
Nkz2BtyGBaWlvnesGNtbQt7e4d89qbYxvj5w9dnMNzd66BO3XpYvTIIGenpGDR4yNc3VgBfGqt/isqMd+
+wYfM2pKakIPV/17ubmplBVbXk7+IWmrKPV37KYp8Axe6XcvzFKSWZmZmIjY39ZLqhjh07YtCgQZ/
dbsaMGejYsSNsbW3Rs2dPqKio4NatW7h79y7mzp2L4OBg5OTkwMPDAzo6Ovj111+hra0NOzs7hIaG4unTp2
jatCmMjY1x+PBhSKVSuLi4lGLP5W3csB4A0K6Np9z6dRs2yW4CWbh4GVRUVDCgXy9kZmaiZes2WL5iTWlHL
ZaXf7/
E4IHfIPHNG5iamaFhw8Y4fe4izMzMxI5WbKfCTuJFdHS+U6IoKkNdTQT6NEEFUz0kpr7HgYjHmLnlHD7k5M
0ju3RXJHS01LHarw2M9DRx4e5LdJ66Rza9jGdtOzhXMIZzBWNE/T5cbt/
abZaUen8+JyE+AcOHDkFcbAwMDA1RtVp17A05jBYtW4sdTXC9evfB64QEBM6egbjYWLjVqIkDoUdhYVH462
bF8KWxOnf2NK5eyZuQulY1+eP1rQdPYGdnL0Li4lH28cpPWewToNj94jyW/
+Pt7Y2tW7cCANTU1GBsbIwaNWrgm2+
+weDBg6GiknfVgEQiwb59+9C1a1e57Y8dO4bAwEDcuHED6urqqFy5MoYOHQpfX1/
s378fCxcuxIMHD5CTk4Pq1atj7ty5aNmyJSIiIvDDDz/g9u3beP/+PSpWrIhp06ahd+/
eBc5e0vNYiqkk57Ek4RV1HktFV9LzWIqhNOZFFENJz2MplrI6XqQ8CjqPJQvLMoCFJSkKFpbKo6wWKiwsiU
oGJ0gnIiIiolLFwpKIiIiIBMHCkoiIiIgEwcKSiIiIiATBwpKIiIiIBMHCkoiIiIgEwcKSiIiIiATBwpKIi
IiIBMHCkoiIiIgEwcKSiIiIiATBwpKIiIiIBMHCkoiIiIgEwcKSiIiIiATBwpKIiIiIBMHCkoiIiIgEwcKS
iIiIiATBwpKIiIiIBMHCkoiIiIgEwcKSiIiIiAShJnYAEo6aqgrUVPlZQRlkZueIHaFExIX4ix2hRFj02yR
2BMG93e0rdoQS8S6rbL63NNVVxY5AVCCsQoiIiIhIECwsiYiIiEgQLCyJiIiISBAsLImIiIhIECwsiYiIiE
gQLCyJiIiISBAsLImIiIhIECwsiYiIiEgQLCyJiIiISBAsLImIiIhIECwsiYiIiEgQLCyJiIiISBAsLImIi
IhIECwsiYiIiEgQLCyJiIiISBAsLImIiIhIECwsiYiIiEgQLCyJiIiISBAsLImIiIhIECwsiYiIiEgQLCyJ
iIiISBAsLImIiIhIECwsqcDWr10DF2d7GOlpoUlDD1yJjBQ7UrFFnDuLHl07wcHWCtrqEoQc2C92pELbtGE
9GtarBRsLY9hYGKN180Y4ceyI7PngTT+jQ1tP2FgYw0hHDUlJSeKFLYSlixeieSMPVDAzhJOtJb7p1Q2PHz
2Ua/P0aRT69+4ORxsLWJsbYXD/
PoiPixMp8adUVCSY0c8dD9b3ReKOIbi3rg+m9Kr1STsXayPsDmiD2F8H4/Xv3oj4sStsTHVlz/
u0roxjczog7rfBeLfPF4Y6GqXZjWJR5uNG3eqVUN5I85MlYMIYAEB8XCxGDRsCt0q2cLQyRuumHgg9sE/
k1MWjzOP1OWWxT4Di9qtAhWVISEiBl7IqNjYWo0ePhqOjIzQ1NWFjY4NOnTohLCxMsNewt7dHUFCQYPsT0u
5dOzF5oj+m/
TATFyOvw82tBjp3aIv4+HixoxVLeno6qrvVQNDKNWJHKTKrChUwK3AeTp+PRHjEZTRt1gLf9O6OB/
fvAQAy3mWgVeu28J84ReSkhXP+3Bn4Dh+Bk2cuYH/oMWR/
yEa3jl5IT08HkDd23Tp6ARIJDh45iWOnziE7Kwt9enSBVCoVOX2e8d1qwNfLFeN+Po+ao3fjh22R8O/
mhu87VJW1cbDUR9j8Tnj0Mgltp4ei7rg9WLD7Ot5n58ja6Giq4cSNv7F4z00RelF0yn7cOBJ+Hrce/
iVbdu4/DADo1KUHAGD0cB9EPXmErb/vQfiFa2jfqSu+G/
IN7ty6KWLqolP28cpPWewToNj9kuTm5uZ+rZGKSsFObEokEuTk5Hy9oZJ5/
vw5GjVqBCMjIwQGBqJ69erIzs7GsWPHsGHDBvz555+CvI69vT38/
Pzg5+dXqO1SUlJgaGiIuDfJMDAwECTLx5o09IB7nboIWrkaACCVSuHsYIMRI0dj4iTlKlg+R1tdgp1/7EPn
Ll1L/
LUys0v2fWJfwQyB8xZhkLePbN25s6fRyasVnr96DSMjoxJ5XYlEUiL7BYDXCQlwsrXE4RPhaNS4KcJOHkfP
Lh3wV8wb2e99cnIy7MqXw77Qo2jh2Uqw17bot6lI2+2Z1hbxSe8wYs1Z2brfJ7XCu6wP8Ak6DQDY5u+J7Bw
pvl1x+qv7a1K1PI7P7QjL/luRnJFVpEz/
eLvbt1jbF4QYx42k9OL9XL5k+pTxOHnsMC5cvw+JRAKnCiZYuHQVevXtL2vj6lAe02bPQ/9BPl/
YU+EZ6Zb8WeqyeJwvi30CxOlXSkoKLMoZIjn5y7VGgSpGqVRaoKUsFpUA8P3330MikSAyMhI9evRApUqVUL
VqVfj7+
+PSpUsAgOjoaHTp0gV6enowMDBA7969Efefr+SioqLQpUsXWFhYQE9PD3Xr1sXJkydlzzdv3hx//
fUXxo0bB4lEUqJ/oAsrKysLN65fg2fLf/
9Qq6iowNOzFSIvXRQxGX0sJycHe3bvREZ6Oup51Bc7jqCSU5IBAMbGJgCArMxMSCQSaGpqytpoaWlBRUUFl
y6cFyXjxy79GYcWblZwtjIEAFS3N0GDKhY4fv0FAEAiAbzq2ODxq2SEzGiHv4IH4OyiLuhUz07M2IIoa8eN
rKws7Nn1O/oO8JYdn+vUq4+Qfbvx9m0ipFIp9u/
ZhfeZ79GwcVOR0xZeWRsvoGz2CVD8fhXrGsv3798LlUNhJSYm4ujRoxg5ciR0dXU/
ed7IyAhSqRRdunRBYmIizpw5gxMnTuDp06fo06ePrF1aWhrat2+PsLAw3LhxA15eXujUqROio6MBAHv37oW
1tTUCAwMRExODmJiYz2bKzMxESkqK3FKSXr9+jZycHJibW8itN7ewQGxsbIm+NhXMvbt3UMHMEOZGOhg35n
v8uuMPVK7iKnYswUilUgRMHIf6DRrBtWo1AEDdevWhq6uLmdOmICMjA+np6fhhykTk5OQgNvbz75/
StGTvTeyOiMKtVb2QsvtbXFraHasP3sWOs1EAAHNDbehra2BC9xo4ceMFOs06jJDLz7Fjcms0rmopcvriKW
vHjaOHQpCSnIQ+3wyUrduwZTuys7Ph6lAedub6mDRuJDb/ugsOjs4iJi2asjZeQNnsE6D4/
Sp0YZmTk4M5c+agQoUK0NPTw9OnTwEA06dPx6ZNRfu6SJE9efIEubm5qFy58mfbhIWF4c6dO9i+fTvc3d3h
4eGBbdu24cyZM7hy5QoAoEaNGvjuu+9QrVo1VKxYEXPmzIGTk5PsulQTExOoqqpCX18flpaWsLT8/
B+VBQsWwNDQULbY2NgI22lSOhUrueDcpWsIO3MB3/p+hxHDfPDng/
tixxLMeL9ReHDvHjZv2y5bZ2pmhuDfduLI4VBYmRrAxsIYyclJqFGrdoEv3ylpPRs5om9TZ3gvP4UG4/
di6MrT8Ovqhv4tKgIAVP535is08i+sOngXt58nYsneWzh8NRq+bauIGZ0+sv2XLfBs1RaW5a1k636cNwspy
cnYdeAIjoZfxHffj8V33v3x4N5dEZMSiavQR9958+YhODgYP/74IzQ0/
r3mo1q1ati4caOg4RRBAS5BxYMHD2BjYyNX4Lm6usLIyAgPHjwAkHfGcsKECahSpQqMjIygp6eHBw8eyM5Y
FkZAQACSk5Nly4sXLwq9j8IwNTWFqqoq4uPl77aNj4v7YgFMpUdDQwOOTs6oWdsdMwPno1p1N6xfs0rsWIK
Y4Dcaxw4fwsFjYahgbS33XMtWbXDr/mNERcfi6d/
x2LB5G2JevYS9vaNIaeXNH+yBJXtvYXfEU9yLfovfzzzBqpC7mNi9JgDgdep7ZH+Q4sGLJLntHv6dBBtTvd
IPLKCydNx4Ef0Xzp0+hW8GDZGte/4sCpt/
XoflqzegSTNPVK3uhvFTfkCNWrWxZeM6EdMWTVkar3+UxT4Bit+vQheW27Ztw4YNG9C/
f3+oqqrK1teoUUOwm1gUScWKFSGRSIrdtwkTJmDfvn2YP38+zp07h5s3b6J69erIyir8heaampowMDCQW0q
ShoYGatV2R/ipf+
+Al0qlCA8PQ736DUr0talopFIpMrMyxY5RLLm5uZjgNxqhIftx8OhJ2Ns7fLZtOVNTGBkZ4czpU0iIj0f7j
p1KMennaWuqQSqV/
3CaI5VCRSXvTGX2BymuPUlApQqGcm0qWhkiOiGt1HKWhLJ03Nj52zaYmpmjVdv2snXvMt4BACQq8tfDq6iq
KsysBIVRlsbrH2WxT4Di90utsBu8fPkSzs6fXj8ilUqRnZ0tSChFYmJigrZt22LNmjUYM2bMJ9dZJiUloUq
VKnjx4gVevHghO2t5//59JCUlwdU17zq38+fPw9vbG926dQOQdwbz+fPncvvS0NBQ2Bugxvj5w9dnMNzd66
BO3XpYvTIIGenpGDR4yNc3VmBpaWmIevJE9vj5s2e4dfMmjE1MYGtrK2Kygps9YypatfGCtY0t0lJT8ceu3
xFx9gz2huRNjRIXG4u4uFg8i8q7ru/+vTvQ09OHjY0tjE1MxIz+ReP9RuGPnb9j++590NPTR9z/
rh0yMDSEtrY2AODXbVvg4lIF5czMcOXyRUyeMA4jR/
uhYiUXMaPLHL4Sjck9a+LF6zTcj36Lmo6mGNO5OraFPZK1Wb7/
Nn4Z74mI+zE4cycGbWpZo31dW7SdHiprY2GkDQsjbTiVz/
sQWc3OBKnvsvDidTrepinuB4iycNyQSqXY8ds29O43AGpq//
7JdK7kAgdHJ0zyG4WZcxfC2MQER0NDcDY8DL/sVM65LMvCeH2sLPYJUOx+FbqwdHV1xblz52BnJ3/
X4h9//IFatT6d+LcsWLNmDRo1aoR69eohMDAQbm5u+PDhA06cOIF169bh/
v37qF69Ovr374+goCB8+PAB33//
PZo1a4Y6deoAyDvzuXfvXnTq1AkSiQTTp0//5FOtvb09zp49i759+0JTUxOmpqZidDdfvXr3weuEBATOnoG
42Fi41aiJA6FHYWFh8fWNFdj1a1fRtlUL2ePJE/
0BAAMGDsbPm4NFSlU4CfEJGD50COJiY2BgaIiq1apjb8hhtGjZGgCweeNPWDR/jqx9+9Z5/
V3z0yb0HzhYlMwFsWnDegBAhzaecuvXbtiE/gO9AQCPHz3C7BnT8DYxEbZ29pgwaSpGjvEr5aSf5//
zBcz8xh0rhjWCmaE2Yt5mYNPxPzF/13VZm5DLzzH6pwhM7F4TS79tiEevktHvx5O48ODfr7mGtq2CH/
q6yx6fnJ93RtZ35Wn8Gv649DpUSGXhuHH2dBhe/
h2NvgPk3yvq6ur4dfcBzJv1Awb17Y709DQ4ODhhxbpNaNmmnUhpi6csjNfHymKfAMXuV4HmsfyvAwcOYPDg
wQgICEBgYCBmz56Nhw8fYtu2bQgNDUXr1q1LKquoYmJiMG/
ePISGhiImJgZmZmZwd3fHuHHj0Lx5c0RHR2P06NEICwuDiooKvLy8sGrVKtkgP3/+HD4+Prh06RJMTU0xef
Jk7N69GzVr1pRNin7p0iV89913ePjwITIzMwt0fSdQOvNYkrBKeh5LsSjSNFlCKuo8loqsNOaxFENJzmMpp
tKYx5LoSwo6j2WhC0sAOHfuHAIDA3Hr1i2kpaWhdu3amDFjBtq0aVOs0FQ0LCyVDwtL5cLCUnmwsCQqGQUt
LAv9VTgANGnSBCdOnChyOCIiIiIqe4pUWALA1atXZVPpuLq6wt3d/StbEBEREVFZVujC8u+//0a/fv1w/
vx52f83nJSUhIYNG2LHjh2w/
mieOSIiIiL6/6HQ81gOHToU2dnZePDgARITE5GYmIgHDx5AKpVi6NChJZGRiIiIiJRAoc9YnjlzBhcuXICL
y7/zxLm4uGDVqlVo0qSJoOGIiIiISHkU+oyljY1NvhOh5+TkwMrKKp8tiIiIiOj/
g0IXlosXL8bo0aNx9epV2bqrV69i7NixWLJkiaDhiIiIiEh5FOircGNjY7n56dLT0+Hh4SH7760+fPgANTU
1+Pj4oGvXriUSlIiIiIgUW4EKy3/+ZxgiIiIios8pUGE5eLDi/n/CRERERKQYijxBOgC8f/8eWVny/30W/
0tBIiIiov+fCn3zTnp6OkaNGgVzc3Po6urC2NhYbiEiIiKi/
58KXVhOmjQJp06dwrp166CpqYmNGzdi9uzZsLKywrZt20oiIxEREREpgUJ/
FX7w4EFs27YNzZs3x5AhQ9CkSRM4OzvDzs4Ov/32G/
r3718SOYmIiIhIwRX6jGViYiIcHR0B5F1PmZiYCABo3Lgxzp49K2w6IiIiIlIahS4sHR0d8ezZMwBA5cqVs
WvXLgB5ZzKNjIwEDUdEREREyqPQheWQIUNw69YtAMCUKVOwZs0aaGlpYdy4cZg4caLgAYmIiIhIORT6Gstx
48bJ/
t2qVSv8+eefuHbtGpydneHm5iZoOCIiIiJSHsWaxxIA7OzsYGdnJ0QWIiIiIlJiBSosV65cWeAdjhkzpshh
iP6/0FRXFTsCFcLb3b5iRxCccd1RYkcoEW+vrBY7QonIzc0VO0KJkEgkYkcoEWVxvArapwIVlsuXLy/
QziQSCQtLIiIiov+nClRY/
nMXOBERERHR5xT6rnAiIiIiovywsCQiIiIiQbCwJCIiIiJBsLAkIiIiIkGwsCQiIiIiQRSpsDx37hwGDBiA
Bg0a4OXLlwCAX375BREREYKGIyIiIiLlUejCcs+ePWjbti20tbVx48YNZGZmAgCSk5Mxf/
58wQMSERERkXIodGE5d+5crF+/Hj///
DPU1dVl6xs1aoTr168LGo6IiIiIlEehC8uHDx+iadOmn6w3NDREUlKSEJmIiIiISAkVurC0tLTEkydPPlkf
EREBR0dHQUIRERERkfIpdGHp6+uLsWPH4vLly5BIJHj16hV+++03TJgwASNGjCiJjERERESkBAr0f4X/
15QpUyCVStGyZUtkZGSgadOm0NTUxIQJEzB69OiSyEhERERESqDQhaVEIsG0adMwceJEPHnyBGlpaXB1dYW
enl5J5CMiIiIiJVHowvIfGhoacHV1FTILERERESmxQheWLVq0gEQi+ezzp06dKlYgIiIiIlJOhS4sa9asKf
c4OzsbN2/exN27dzF48GChchERERGRkil0Ybl8+fJ818+aNQtpaWnFDkREREREyqlI/
1d4fgYMGIDNmzcLtTsiIiIiUjKCFZYXL16ElpaWULsjIiIiIiVT6MKye/
fucku3bt1Qv359DBkyBN99911JZCQFsX7tGrg428NITwtNGnrgSmSk2JEEwX4pj4hzZ9Gjayc42FpBW12Ck
AP7xY4kGGUbLz0dTSye0AMPDwci8eIyhAf7w93VVva8rrYGlk/uhSdH5yDx4jJc3zMNQ3s2/
uz+9q8egXc3VqNTc7fSiF9syjZeX1O5ogN0NFQ+WfzGjBQ7WrFsWL8OdWu5wdzEAOYmBmjWuAGOHT0idqxi
mxs465OxqlmtitixABShsDQ0NJRbTExM0Lx5cxw+fBgzZ84siYxKzdvbG127di1w+
+fPn0MikeDmzZsllqkodu/aickT/
THth5m4GHkdbm410LlDW8THx4sdrVjYL+WSnp6O6m41ELRyjdhRBKWM47VuxjfwrF8ZPj9sRZ3e83Hy4p84
tH40rMwMAQCLxvdA64auGDJtG2p2n4vVv53G8sm90KFZ9U/2Nbp/
C+TmlnYPik4Zx+trzl2IxNPoV7Il9MhxAED3Hr1ETlY8FaytMWf+Qly4fA3nL11F8xae6NW9C+7fuyd2tGJ
zda0qN2YnT58TOxKAQhaWOTk5GDJkCJYtW4YtW7Zgy5Yt2LRpExYuXIg2bdqUVEZBJCQkYMSIEbC1tYWmpi
YsLS3Rtm1bnD9/
XuxoSmFl0DIM+dYXg7yHoIqrK1atXQ9tHR1sDVbu62rZL+XS1qsdZgXORZeu3cSOIihlGy8tTXV0bVkT04L
24/
z1KDx98RrzfjqMqBcJ8O3VBABQv4YDfg29jHPXHiM6JhGb957H7UcvUaeqndy+3CpVwNiBnhg+61cxulIky
jZeBWFmZgZLS0vZcuRwKBydnNCkaTOxoxVLh46d4NWuPZwrVkTFSpUwe8486OnpIfLyJbGjFZuqmprcmJma
moodCUAhC0tVVVW0adMGSUlJJRSn5PTo0QM3btzA1q1b8ejRI4SEhKB58+Z48+aN2NEUXlZWFm5cvwbPlq1
k61RUVODp2QqRly6KmKx42C9SBMo4XmqqKlBTU8X7rGy59e8zs9GwlhMA4NKtZ+jYrLrsDGbTOhVR0c4cJy
89kLXX1lJH8AJv+C3chbg3qaXXgWJQxvEqrKysLOzY/
hsGDR7yxXmrlU1OTg527dyB9PR0eNRvIHacYot68hiOdhXg6uKEIYMG4EV0tNiRABThq/
Bq1arh6dOnJZGlxCQlJeHcuXNYtGgRWrRoATs7O9SrVw8BAQHo3LkzAGDZsmWoXr06dHV1YWNjg++//
15u+qTg4GAYGRnh2LFjqFKlCvT09ODl5YWYmBhZm5ycHPj7+8PIyAjlypXDpEmTkPvR9ztHjx5F48aNZW06
duyIqKio0vlBFNHr16+Rk5MDc3MLufXmFhaIjY0VKVXxsV+kCJRxvNIyMnHp1lME+LZDeTNDqKhI0Ld9XXi
4OcDS1AAA4L9oNx48jUXU8XlIiVyBkDXfw2/hLpy//u/
x7sfxPXDp1jOEnr4jVlcKTRnHq7AOHtiPpKQkDBjkLXYUQdy9cwemRnow1NXEmJHDsfOPfaii5P9zYN16Ht
iwcQsOHDyCFavW4vnzZ2jl2RSpqeJ/QCt0YTl37lxMmDABoaGhiImJQUpKityiiPT09KCnp4f9+/
cjMzMz3zYqKipYuXIl7t27h61bt+LUqVOYNGmSXJuMjAwsWbIEv/
zyC86ePYvo6GhMmDBB9vzSpUsRHByMzZs3IyIiAomJidi3b5/cP
tLT0+Hv74+rV68iLCwMKioq6NatG6RSaYH7k5mZqRQ/dyIqu3x+2AaJBHh6fB6SLwdhZL9m2HX0KqTSvA/
T3/dthnrV7dFj7Ho07L8IU5btQ9CU3mjh4QIA6NCsOprXq4SJi/
8QsxuUj63Bm9GmbTtYWVmJHUUQlVxccPnqTZw9fxm+342Ar89gPLh/
X+xYxdLWqx269+yF6m5uaN2mLfaFHEJyUhL2/
LFL7GiFnyC9ffv2AIDOnTvLnSLPzc2FRCJBTk6OcOkEoqamhuDgYPj6+mL9+vWoXbs2mjVrhr59+8LNLe8O
RD8/P1l7e3t7zJ07F8OHD8fatWtl67Ozs7F+/Xo4OeV91TNq1CgEBgbKng8KCkJAQAC6d+8OAFi/
fj2OHTsml6VHjx5yjzdv3gwzMzPcv38f1apVK1B/
FixYgNmzZxf8B1BMpqamUFVVRXx8nNz6+Lg4WFpalloOobFfpAiUdbye/f0abYaugI6WBgz0tBD7OgW/
LByCZy9fQ0tTHbNHd0If/59xNCLvJom7j1/BzcUafgNbIvzyQzSvWwmO1qaIPbtYbr+/
LxmK8zei0NZ3hRjd+iplHa+Civ7rL5wKO4nfd+0RO4pgNDQ04OTsDACo7e6Oa1evYM2qFVi97ieRkwnHyMg
IzhUr4emTJ2JHKfwZy/
DwcNly6tQp2fLPY0XVo0cPvHr1CiEhIfDy8sLp06dRu3ZtBAcHAwBOnjyJli1bokKFCtDX18fAgQPx5s0bZ
GRkyPaho6MjKyoBoHz58rK7AJOTkxETEwMPDw/
Z82pqaqhTp45cjsePH6Nfv35wdHSEgYEB7O3tAQDRhbg2IiAgAMnJybLlxYsXhf1xFIqGhgZq1XZH+Kkw2T
qpVIrw8DDUU+LrVNgvUgTKPl4Z77MQ+zoFRvraaNWwCkJP34G6mio01NUg/
ehSoJwcKVRU8k5ILNlyHHV7L4BH34WyBQAmLd2DYTMV90YeZR+vr9m2dQvMzM3Rrn0HsaOUGKlU+tlvL5VV
Wloanj2NgmX58mJHKfwZSwcHB9jY2HxyQW9ubm6JFzjFpaWlhdatW6N169aYPn06hg4dipkzZ6J58+bo2LE
jRowYgXnz5sHExAQRERH49ttvkZWVBR0dHQCAurq63P4kEskn11B+TadOnWBnZ4eff/
4ZVlZWkEqlqFatGrKysgq8D01NTWhqahbqdYtrjJ8/
fH0Gw929DurUrYfVK4OQkZ6OQYOHlGoOobFfyiUtLQ1R//lE/
vzZM9y6eRPGJiawtbX9wpaKTRnHq1WDKpBIgEfP4+FkY4b547ri0bM4bAu5iA8fpDh79THm+3XFu/
fZiI5JRBN3Z/TvWA+Tl+0FAMS9Sc33hp0XMW/
x1yvFvqlSGcerIKRSKX7ZFowBAwZBTa3Q5YFCmj4tAG292sHGxhapqanYuWM7zp45jYOHj319YwUWMHkC2n
foBFtbO8TEvMLcwFlQVVVFrz79xI5WtMIyJiYG5ubmcusTExPh4OCgkF+Ff46rqyv279+Pa9euQSqVYunSp
VBRyTuJu2tX4a5TMDQ0RPny5XH58mU0bdoUAPDhwwdcu3YNtWvXBgC8efMGDx8+xM8//4wmTfKm5IiIiBCw
RyWnV+8+eJ2QgMDZMxAXGwu3GjVxIPQoLCwsvr6xAmO/lMv1a1fRtlUL2ePJE/
0BAAMGDsbPm4NFSlV8yjhehnpaCBzdGRUsjJCYnIEDYTcxc81BfPiQd734oCmbETi6C4LnD4axgQ6iYxIxa
00oft6tHMe8L1HG8SqIU2En8SI6GoO8fcSOIpiE+Hh8O2QQYmNiYGhoiGrV3XDw8DG0bNVa7GjF8vLvlxg8
8BskvnkDUzMzNGzYGKfPXYSZmZnY0SDJLeQpNxUVFcTFxX0S/q+//oKrqyvS09MFDSiEN2/
eoFevXvDx8YGbmxv09fVx9epVjB49Gh06dMCYMWNQs2ZNBAUFoVOnTjh//
jwCAgLw8uVLvH37FkZGRggODoafn5/cVEv79+9Ht27dZGctFy1ahB9//
BGbNm1C5cqVsWzZMuzYsQOenp7Yv38/pFIpzM3N0a5dO8ycORPR0dGYMmUKrly5gn379qFr1654/
vw5HBwccOPGDdSsWbNA/UtJSYGhoSHi3iTDwMCgBH6CRFTWGNcdJXaEEvH2ymqxI5SIwn47pizK0nRG/
1UWxyslJQWWpkZITv5yrVHgM5b+/
nlnBiQSCaZPny77ehjIm2bn8uXLBS6ESpuenh48PDywfPlyREVFITs7GzY2NvD19cXUqVOhra2NZcuWYdGi
RQgICEDTpk2xYMECDBo0qFCvM378eMTExGDw4MFQUVGBj48PunXrhuTkZAB5RfmOHTswZswYVKtWDS4uLli
5ciWaN29eAr0mIiIiKl0FPmPZokXe109nzpxBgwYNoKGhIXtOQ0MD9vb2mDBhAipWrFgySemzeMaSiAqLZy
yVS1k8AwbwjKUyEfyMZXh4OABgyJAhWLFiBQsYIiIiIpJT6Jt3tmzZUhI5iIiIiEjJFXoeSyIiIiKi/
LCwJCIiIiJBsLAkIiIiIkGwsCQiIiIiQbCwJCIiIiJBsLAkIiIiIkGwsCQiIiIiQbCwJCIiIiJBsLAkIiIi
IkGwsCQiIiIiQbCwJCIiIiJBsLAkIiIiIkGwsCQiIiIiQbCwJCIiIiJBsLAkIiIiIkGwsCQiIiIiQbCwJCI
iIiJBqIkdgITzMjEDydlla0htyumIHaFEpGd+EDtCiVBXLZufVcvieL29slrsCCWi4fxTYkcoERemeoodoU
R8yJGKHaFEqJXBY6FEIilQu7LXcyIiIiISBQtLIiIiIhIEC0siIiIiEgQLSyIiIiISBAtLIiIiIhIEC0siI
iIiEgQLSyIiIiISBAtLIiIiIhIEC0siIiIiEgQLSyIiIiISBAtLIiIiIhIEC0siIiIiEgQLSyIiIiISBAtL
IiIiIhIEC0siIiIiEgQLSyIiIiISBAtLIiIiIhIEC0siIiIiEgQLSyIiIiISBAtLIiIiIhIEC0siIiIiEgQ
LS8KVixH4bmBPNK7hhEqWujhx5OBn286YNAaVLHURvGG13Prhg3qhmbsLqtmZoJGbIyaM+hZxsTElHV0Q69
eugYuzPYz0tNCkoQeuREaKHalQFs0LhKmeutxSv1Y1AMDbxERMGT8WHrWqwtpUHzUqOyJggh9SkpNFTv11S
xcvRPNGHqhgZggnW0t806sbHj96KNfm6dMo9O/dHY42FrA2N8Lg/
n0QHxcnUuKCycnJwY9zZ8HDrRIcLQ3RoGZlLP9xPnJzc/
NtP3ncSFgZaeLntStLOakwlO39ZaavgbldXXFqQhNcCGiGnd/VQ5Xy+rLnr8/
wzHcZ1MBW1mZ5n+o4NLYhLk5thmPjGmFOV1eY6mmI0Z1CU7bx+ljEubPo1b0zKjpYQ19LFQdD9ss9Hx8Xh+
+GDkFFB2uYG+uhW6d2ePLksThhi0lRx4qFpQAkEgn279//2edPnz4NiUSCpKSkUstUGBkZ6ahctTpmLFj+x
XbHD4fg5rVImFuW/+Q5j0ZNsWLDLzgWcROrNm3Hi+fPMGZo/5KKLJjdu3Zi8kR/
TPthJi5GXoebWw107tAW8fHxYkcrlMpVquJe1AvZcujEaQBAbMwrxMbEYPa8RTgXeROr1m9C2MnjGPv9MHE
DF8D5c2fgO3wETp65gP2hx5D9IRvdOnohPT0dAJCeno5uHb0AiQQHj5zEsVPnkJ2VhT49ukAqlYqc/
vPWBC3B1s0bMG9xEM5cvoVps+dj7cql2PTTmk/aHjl4ANeuRMKyvJUISYtP2d5f+lpq2DLEHR+kuRi9/
SZ6rruM5SeeIPX9B1mb1ksj5JZZBx5AmpuLsAf/
9unq8yRM+eMuuq+5jIm778DaWBuLe1UXo0uFomzjlZ+MjHRUr14DS4NWffJcbm4u+vbujufPnmHH7n2IuHw
NNrZ26Nyujey4oiwUeawkuZ/7mEwyCQkJmDFjBg4dOoS4uDgYGxujRo0amDFjBho1agSJRIJ9+/
aha9eu+W6flZWFxMREWFhYQCKRfPZ1vL29kZSU9MUiNT8pKSkwNDTE9ccx0NM3KNS2H6tkqYs1W3agdbtOc
utjY16hV/
tm2Pz7AQwb0AODh42E97BRn91P2LFD+N67D+5Gv4W6unqR89iU0ynytgXRpKEH3OvURdDKvDOwUqkUzg42G
DFyNCZOmlJir5ue+eHrjQpo0bxAHAk9gNMXrxWo/
YG9f2DE0MGIjk+GmpqaYDkAQF215D6rvk5IgJOtJQ6fCEejxk0RdvI4enbpgL9i3sDAIO/
3Pjk5GXbly2Ff6FG08Gwl2GsLOV6D+nSFqZkFlq3+SbZu6MA+0NLWxuoNwbJ1Ma9eomOrJti+JxQDe3eF74
hR8P1+jGA5jHVL/gyaGO+vhvNPFXnb0S2dUNPGEN8GXy/wNkt7V4eupiqG/3Lzs22aVjLFsj7VUX/
eaXyQFu1P7oWpnkXarjDEGK8POSX3IVBfSxXbd+1Bp85dAQCPHz9C7epVEHn9Nqq4VgWQ10cnOyvMnD0X3j
5DBXtttRI8FgLijFVKSgosyhkiOTlZdszND89YFkCPHj1w48YNbN26FY8ePUJISAiaN2+ON2/
eFGh7DQ0NWFpafraozMnJUegzLFKpFJNGfYuh3/uhYmXXr7ZPepuIkD07Uatu/
WIVlSUtKysLN65fg2fLfwsQFRUVeHq2QuSliyImK7ynUU9Q1dkW7tUq4Tufgfj7RfRn26akJENf30DworKk
JafkfX1vbGwCAMjKzIREIoGmpqasjZaWFlRUVHDpwnlRMhZEnXoNEHEmHFFPHgEA7t25jchLF+DZqq2sjVQ
qxZjvfDBi9Di4VPn6e04RKeP7q1klU9x/
lYJFPavh5PjG2O5bF91qff5ssYmuOhpXLIf9Nz5/2Y+BlhraV7fArRfJRS4qS4MyjldhZWVmAgA0NbVk61R
UVKCpoYmLCnzM+JiijxULy69ISkrCuXPnsGjRIrRo0QJ2dnaoV68eAgIC0LlzZ1m7169fo1u3btDR0UHFih
UREhIie+7jr8KDg4NhZGSEkJAQuLq6QlNTEz4+Pti6dSsOHDgAiUQCiUSC06dP55spMzMTKSkpcktJ2rB6K
VTV1DBo6PdfbLd4zg+o4WCGelVsEPPyBdYF7yzRXMX1+vVr5OTkwNzcQm69uYUFYmNjRUpVeO5162HV+k3Y
tT8Ui4NWI/qv5+jYpgVSU1M/
afvm9WssXTQfg4YI98m8NEilUgRMHIf6DRrBtWre9aN169WHrq4uZk6bgoyMDKSnp+OHKRORk5ODWAW+vnf
UuIno0qMXmtZ1g62pLto0rQffEaPRvXc/WZs1QUugqqaKb4d//lsBRaeM768KxlroWacCXiRmYORvN/
HHtZeY6FURHd0s823fqUZ5ZGTl4NSDhE+eG9PSCeenNMPpSU1haagF/
523Szp+sSjjeBVWJZfKsLGxxawZU/H27VtkZWVh2ZIf8fLl30pzTwCg+GPFwvIr9PT0oKenh/
379yPzf5928jN79mz07t0bt2/fRvv27dG/
f38kJiZ+tn1GRgYWLVqEjRs34t69e1i5ciV69+4NLy8vxMTEICYmBg0bNsx32wULFsDQ0FC22NjYFLufn3P
31g1s+3ktFq7Y8MWv8QHg2+/9sP/kBWzZGQIVVVVMGu372RsSSDit2nihS/
eeqFrNDZ6t2mDHnoNITk7Cgb275dqlpqSgX8/OcKlcBZOmzRApbdGM9xuFB/
fuYfO27bJ1pmZmCP5tJ44cDoWVqQFsLIyRnJyEGrVqQ0VFcQ9tIfv+wN7dO7Bm4zYcO3MZK9ZtwvpVy7Fr+
y8AgNs3r2Pj+tUIWrvxq+85EpaKRII/Y9Kw+tRTPIxNw97rr7Dv+iv0rFMh3/
ada5bHkTuxyMrn69xtF6LRb0MkRvx6AznSXAR2Vc4zz2WJuro6ftv5B548fgzb8qYwN9bDuTPhaNPWS6GPG
cqGP8mvUFNTQ3BwMLZu3QojIyM0atQIU6dOxe3b8p8+vb290a9fPzg7O2P+/
PlIS0tD5Bfu0MrOzsbatWvRsGFDuLi4wMDAANra2tDU1ISlpSUsLS2hoZH/NVABAQFITk6WLS9evBC0z/
919fJ5vHmdgObuLqhSwQBVKhjg5d/
RWDgrAC3qVJFra1LOFA5OFdGoWUsErd+KM2HHcPOaYtyllh9TU1OoqqoiPl7+LuL4uDhYWuZ/
hkIZGBoZwcm5Ip49jZKtS01NRe9uHaCnp4+tv/
+h0JcofGyC32gcO3wIB4+FoYK1tdxzLVu1wa37jxEVHYunf8djw+ZtiHn1Evb2jiKl/
bo5MwIwym8CuvbojSpVq6Fn3/7w/X4MVi3/EQBw+UIEXifEo241Z9iU04FNOR38/
eIvzP5hMupVryRy+oJTxvfX69QsPE2Qv4nj2esMWBpofdK2lq0hHEx1se8zX4MnvctGdOI7XH76FgF77qFJ
RVO4WRfvGviSpIzjVRS1arvjQuR1/
B2XiMfPX2LfwSNITEyEvYPiHjM+puhjxcKyAHr06IFXr14hJCQEXl5eOH36NGrXro3g4GBZGzc3N9m/
dXV1YWBg8MW7szQ0NOS2KQxNTU0YGBjILSWlS89+OHjqMg6cvChbzC3L49vv/
bBpx4HPbvfPNaNZXzjLKzYNDQ3Uqu2O8FNhsnVSqRTh4WGoV7+BiMmKJy0tDc+fPYWFRd4BJjUlBb26tIO6
ugZ+3bUPWlqf/
pFURLm5uZjgNxqhIftx8OhJ2Ns7fLZtOVNTGBkZ4czpU0iIj0f7jp0+21Zs7zMyPjk7oqqqitz/
vWd69O2PsPPXcOLcFdliWd4KI8b4Y/
vez08FpmiU8f1180US7E3lbxi0K6eNmOT3n7TtUtMK91+l4HFc2lf3q/K/
E88leXNbcSnjeBWHoaEhzMzM8OTJY1y/dhUdOnb+
+kYKQtHHSrmu3heRlpYWWrdujdatW2P69OkYOnQoZs6cCW9vbwD45AyQRCL54g052traCvM1V3p6Gv569u/
Zrb+jn+P+3VswMjKBlbUNjE3KybVXV1OHmbkFHJ3zzp7cun4Ft29cg7tHAxgaGiP6r6dYsWgObO0dUauOR6
n2pbDG+PnD12cw3N3roE7deli9MggZ6ekYNHiI2NEKbMbUSWjbriNsbG0RG/
MKi+YFQlVFFd179UVqSgp6dmmHdxkZWLdxK1JTU5CamndNrqmpGVRVVUVO/3nj/Ubhj52/Y/
vufdDT00fc/
64dMjA0hLa2NgDg121b4OJSBeXMzHDl8kVMnjAOI0f7oWIlFzGjf1Frrw5YuXQRKljbwKWyK+7evoWf1qxA
3wGDAQAmJuVg8tF7Tk1NHebmFnCuqLj9yo+yvb9+u/
wCW4a4w6exHU7ci0fVCgboXrsC5ob+KddOV0MVrV3NsezEp/
MfVqtggKpW+rgRnYzU9x9gbayNES0c8CIxA7f/Vuz5Y5VtvPKTlpaGp1FPZI//ev4ct2/
dhLGxCWxsbbFvz26YmprB2sYW9+7dweTx49Cxcxe0bN1GxNSFp8hjxcKyiFxdXQs9LdDXaGhoICcnR9B9Fs
Tdm9cxsEc72eMFM/
OmKujWuz8Wrdzw1e21tLVx4vABrFoyDxkZ6TA3t0STFq0RtGEyNP5zx64i6tW7D14nJCBw9gzExcbCrUZNH
Ag9CgsLi69vrCBevXyJYUMG4G3iG5QzNYNHg0Y4Gh4BUzMzRJw9g2tX8i5HqOtWWW676/
cew9bOXoTEBbNpw3oAQIc28tOsrN2wCf0HegMAHj96hNkzpuFtYiJs7ewxYdJUjBzjV8pJC2fuj8vx47xZC
Bg/Fm9ex8PCsjwGDhmKcZOmiR1NcMr2/rr/KhUTdt3BKE8n+Da1x6u377Hk2GMcuSv/
lWPbahaABDh299PJ+N9n58Czsjm+a+YIbQ0VvE7NwoWoN5h87jmycxT7mnNlG6/83Lh2Fe3btpQ9Dpg0HgD
wzYBB+GnjFsTGxiJg0gTEx8fB0rI8+vUfiMlTfxArbpEp8lhxHsuvePPmDXr16gUfHx+4ublBX18fV69exe
jRo9GhQwds2rQp33ksjYyMEBQUBG9vb5w+fRotWrTA27dvYWRkhODgYPj5+X0yYfr8+fPx008/4fjx4yhXr
hwMDQ0LdC2ckPNYKpqSnsdSLELOi6hIFPmrvuIoi+NVGvNYiqE481gqstKYx1IMJTmPpZhKeh5LMRR0Hkue
sfwKPT09eHh4YPny5YiKikJ2djZsbGzg6+uLqVOnCvpavr6+OH36NOrUqYO0tDSEh4ejefPmgr4GERERUUn
hGcsygGcslU9ZPAMG8IylMuEZS+XCM5bK5f/
zGcuy13MiIiIiEgULSyIiIiISBAtLIiIiIhIEC0siIiIiEgQLSyIiIiISBAtLIiIiIhIEC0siIiIiEgQLSy
IiIiISBAtLIiIiIhIEC0siIiIiEgQLSyIiIiISBAtLIiIiIhIEC0siIiIiEgQLSyIiIiISBAtLIiIiIhIEC
0siIiIiEgQLSyIiIiISBAtLIiIiIhIEC0siIiIiEoSa2AFIOBVMdGBgoCN2DCoAXU2+9ZSJhpqG2BEEl5md
I3aEEnFhqqfYEUqEcbNpYkcoEW/PzBM7QonI/
iAVO4LgCtonnrEkIiIiIkGwsCQiIiIiQbCwJCIiIiJBsLAkIiIiIkGwsCQiIiIiQbCwJCIiIiJBsLAkIiIi
IkGwsCQiIiIiQbCwJCIiIiJBsLAkIiIiIkGwsCQiIiIiQbCwJCIiIiJBsLAkIiIiIkGwsCQiIiIiQbCwJCI
iIiJBsLAkIiIiIkGwsCQiIiIiQbCwJCIiIiJBsLAkIiIiIkGwsCQiIiIiQbCwJCIiIiJBsLCkAlu/
dg1cnO1hpKeFJg09cCUyUuxIgmC/
lAv7pZg2bViPhvVqwcbCGDYWxmjdvBFOHDvySbvc3Fz07NIBRjpqCA05IEJSYSjbeOnpaGDx2PZ4uGcCEk/
NQvj6YXCvXCHftisndsG78/MwqndD2TpbSyOsm9IND3aPR+KpWbi3yx8/
fNsS6mqqpdWFIlO2sfrY0sUL0ayRB6zMDOFoa4l+vbrh8aOH+bbNzc1F9y7tYaCtitCQ/
aUb9H9YWArk+fPnkEgkuHnzpthRSsTuXTsxeaI/pv0wExcjr8PNrQY6d2iL+Ph4saMVC/
ulXNgvxWVVoQJmBc7D6fORCI+4jKbNWuCb3t3x4P49uXZrV6+ARCIRKaUwlHG81k3pBs+6zvAJ/
AN1Bq7EycgnOLTCB1amBnLtOjd1Rb2qNniVkCK33sXODCoqEoxafAC1B6zApJWHMbRrPQR+17o0u1FoyjhW
H4s4dwbDho9A2JkLOBB6DNkfstG1oxfS09M/
abtmlfjvL0lubm6uqAkEkJCQgBkzZuDQoUOIi4uDsbExatSogRkzZqBRo0alkuH58+dwcHDAjRs3ULNmzVJ
5zX+kpKTA0NAQcW+SYWBg8PUNiqBJQw+416mLoJWrAQBSqRTODjYYMXI0Jk6aUiKvWRrYL+XCfgknMzunRP
b7X/
YVzBA4bxEGefsAAG7fuom+PbogPOIyXByt8euOPejYuYugr6mpXvJn0MQYL+Nm04q8rZaGGhJOzECvKb/
h6MV/z3Sd3/Q9jl96hNk/nwQAWJka4OzPw9HJPxj7Fg/C6l0XsHrXhc/ud9w3jeHb1QOuvZcWOdvbM/
OKvG1BiHXMyP4gLbF9v05IgKOtJY6cCEejxk1l62/
fuone3TvjzPlIVHSogO0796Bj566CvW5KSgqsLYyRnPzlWqNMnLHs0aMHbty4ga1bt+LRo0cICQlB8+bN8e
bNG7GjFUt2drbYEQAAWVlZuHH9GjxbtpKtU1FRgadnK0ReuihisuJhv5QL+6U8cnJysGf3TmSkp6OeR30AQ
EZGBnyHDMTi5atgYWkpcsKiU8bxUlNTgZqaKt5nyf9NeZ+ZjYZudgAAiUSC
TTN6Yvn2c3jwrGBn8wx0tZCY+k7wvEJRxrEqiOSUZACAsbGJbF1GRga+9R6ApUHiv7+UvrBMSkrCuXPnsGj
RIrRo0QJ2dnaoV68eAgIC0LlzZwB5b5iNGzeiW7du0NHRQcWKFRESEiK3n7t376Jdu3bQ09ODhYUFBg4ciN
evX8ueP3r0KBo3bgwjIyOUK1cOHTt2RFRU1Gdz5eTkwMfHB5UrV0Z0dDQA4MCBA6hduza0tLTg6OiI2bNn4
8OHD7JtJBIJ1q1bh86dO0NXVxfz5uX/SS4zMxMpKSlyS0l6/
fo1cnJyYG5uIbfe3MICsbGxJfraJYn9Ui7sl+K7d/cOKpgZwtxIB+PGfI9fd/
yBylVcAQBTJ41HPY8G6NCps8gpi0cZxystIwuX7vyFAO8WKG+qDxUVCfq2qQGParawNNUHAIwf0AQfcqRYs
7tgBZdjBROM6NkAm/Yr7vWKyjhWXyOVSjFl4jjUb9AIrlWrydYHTPKHR/
0G6NBJ2G8AikLpC0s9PT3o6elh//79yMzM/Gy72bNno3fv3rh9+zbat2+P/
v37IzExEUBecerp6YlatWrh6tWrOHr0KOLi4tC7d2/
Z9unp6fD398fVq1cRFhYGFRUVdOvWDVLpp6e7MzMz0atXL9y8eRPnzp2Dra0tzp07h0GDBmHs2LG4f/
8+fvrpJwQHB39SPM6aNQvdunXDnTt34OPjk29fFixYAENDQ9liY2NTlB8dEZGgKlZywblL1xB25gK+9f0OI
4b54M8H93E49CDOngnHgsXLxI74/5bPnD8gkUjw9MAUJIfPxsheDbHr5G1Ipbmo5WKFkb0aYti8PQXal5Wp
AUKWeWNv+F1sOXi1hJPTf433G4UH9+5hy7btsnWHQ0Nw5nQ4Fi5eLmKyf6mJHaC41NTUEBwcDF9fX6xfvx6
1a9dGs2bN0LdvX7i5ucnaeXt7o1+/
fgCA+fPnY+XKlYiMjISXlxdWr16NWrVqYf78+bL2mzdvho2NDR49eoRKlSqhR48ecq+7efNmmJmZ4f79+6h
W7d9PDWlpaejQoQMyMzMRHh4OQ0NDAHmF7ZQpUzB48GAAgKOjI+bMmYNJkyZh5syZsu2/+eYbDBky5It9Dg
gIgL+/
v+xxSkpKiRaXpqamUFVVRXx8nNz6+Lg4WCrxV1rsl3JhvxSfhoYGHJ2cAQA1a7vj+rWrWL9mFbS0tfHsaRT
sypeTaz/om15o0KgxDh07JUbcIlHW8Xr2MhFtRm2EjpY6DHS1EPsmFb8E9sGzV2/
RqIY9zI118WjPRFl7NTVVLBzVDqN6N0Tlnktk68ub6uPoqm9x6U40Ri7aL0JPCk5Zx+pzxvuNxtHDh3Dk5G
lUsLaWrT9zOhzPnkbBxtJErv2Afr3QsFETHD5euu8vpT9jCeRdY/
nq1SuEhITAy8sLp0+fRu3atREcHCxr898iU1dXFwYGBrK7wm7duoXw8HDZ2U89PT1UrlwZAGRfdz9+/
Bj9+vWDo6MjDAwMYG9vDwCyr7n/0a9fP6Snp+P48eOyovKf1wgMDJR7DV9fX8TExCAjI0PWrk6dOl/
tr6amJgwMDOSWkqShoYFatd0RfipMtk4qlSI8PAz16jco0dcuSeyXcmG/
lI9UKkVmVibGjZ+E85E3cO7SNdkCAPN/XIo1P20SOWXhKPt4ZbzPRuybVBjpa6FVvYoIPfcA24/
eQN1Bq+DhvVq2vEpIwfLt59DJP1i2rZWpAY6tGoobD19i2Pw9UPR7f5V9rP6Rm5uL8X6jERqyHwePnoS9vY
Pc8/4TJuPilZs4f/m6bAGABT8uw9oNpf/+Uvozlv/
Q0tJC69at0bp1a0yfPh1Dhw7FzJkz4e3tDQBQV1eXay+RSGRfY6elpaFTp05YtGjRJ/
stX748AKBTp06ws7PDzz//DCsrK0ilUlSrVg1ZWVly7du3b49ff/
0VFy9ehKenp2x9WloaZs+eje7du+eb/R+6urpF+wGUsDF+/vD1GQx39zqoU7ceVq8MQkZ6OgYN/
vLZVUXHfikX9ktxzZ4xFa3aeMHaxhZpqan4Y9fviDh7BntDDsPC0jLfGwqsrW0/+SOpDJRxvFrVc4ZEIsGj
6NdwsjbB/JHt8Cg6AdsOXcOHHCkSU+Rvwsn+kIO4xDQ8js6718DK1ADHVn+L6NgkBKw+CjOjf/
9WxSWmlWpfCkMZx+pj/n6j8MfO3/H77n3Q19NH3P+uDzUwNIS2tvZn3182NjaivL/
KTGH5MVdXV+zfv79AbWvXro09e/bA3t4eamqf/
kjevHmDhw8f4ueff0aTJk0AABEREfnua8SIEahWrRo6d+6MQ4cOoVmzZrLXePjwIZydnYvWIZH16t0HrxMS
EDh7BuJiY+FWoyYOhB6FhYXF1zdWYOyXcmG/
FFdCfAKGDx2CuNgYGBgaomq16tgbchgtWir2PIdFoYzjZainhcDhbVDBzBCJKe9w4Mw9zPzpOD7kFGxaHM9
6TnC2MYWzjSmiDkyWe067UdGnQippyjhWH9u0YT0AoH0bT7n16zZsQv+B3iIk+jKln8fyzZs36NWrF3x8fO
Dm5gZ9fX1cvXoVo0ePRocOHbBp0yZIJBLs27cPXbt2lW1nZGSEoKAgeHt749WrV6hZsyaaNWuGSZMmwcTEB
E+ePMGOHTuwceNGSCQSmJubo127dpg5cyaio6MxZcoUXLlyRbbfj+exDAoKwvTp03HkyBE0btwYx44dQ8eO
HfHDDz+gZ8+eUFFRwa1bt3D37l3MnTsXAPLNWRClMY8lEZUtpTGPpRhKYx5LMRRnHktFVtLzWIqlJOexFEt
B57FU+jOWenp68PDwwPLlyxEVFYXs7GzY2NjA19cXU6dOLdA+rKyscP78eUyePBlt2rRBZmYm7Ozs4OXlBR
UVFUgkEuzYsQNjxoxBtWrV4OLigpUrV6J58+af3aefnx+kUinat2+Po0ePom3btggNDUVgYCAWLVoEdXV1V
K5cGUOHDhXoJ0FEREQkLqU/Y0k8Y0lEhcczlsqFZyyVy//
nM5Zl4q5wIiIiIhIfC0siIiIiEgQLSyIiIiISBAtLIiIiIhIEC0siIiIiEgQLSyIiIiISBAtLIiIiIhIEC0
siIiIiEgQLSyIiIiISBAtLIiIiIhIEC0siIiIiEgQLSyIiIiISBAtLIiIiIhIEC0siIiIiEgQLSyIiIiISB
AtLIiIiIhIEC0siIiIiEgQLSyIiIiISBAtLIiIiIhKEmtgBSDi5ubnIzc0VO4agJBKJ2BGoEKTSsvX794+y
+GuooVY2zyuUtWPgP96emSd2hBJh3MBf7Agl4u3FZWJHEJx6AY8ZZfPIQkRERESljoUlEREREQmChSURERE
RCYKFJREREREJgoUlEREREQmChSURERERCYKFJREREREJgoUlEREREQmChSURERERCYKFJREREREJgoUlER
EREQmChSURERERCYKFJREREREJgoUlEREREQmChSURERERCYKFJREREREJgoUlEREREQmChSURERERCYKFJ
REREREJgoUlEREREQmChSURERERCYKFJRXakh8XQkdDBRPH+4kdpdgizp1Fj66d4GBrBW11CUIO7Bc7kmDW
r10DF2d7GOlpoUlDD1yJjBQ7UrHk5OQgcNZ0uFZyRDlDHVSr7IyF8+cgNzdX7GjFMjdwFnQ0VOSWmtWqiB2
r2MpqvwDg5cuX8Bk8ENaWpjAx0EHdWm64du2q2LGKZcP6dahbyw3mJgYwNzFAs8YNcOzoEbFjfZGejiYW+3
fFw5AfkHhuEcI3jYa7q41cGxd7c+xe6oPY8Hl4fXYBIrb6wcbCSPa8RTl9bJr9DZ4dnYXXZxfgwi/
+6NrCrZR7UjSKeoxXEzsAKZerV69g08YNqF5dOd54X5Oeno7qbjUwyNsHfXt1FzuOYHbv2onJE/
2xas161K3ngdUrg9C5Q1vcuvcQ5ubmYscrkmVLFmHjhvXYsDEYVVyr4vr1qxju6wMDA0N8P2qM2PGKxdW1K
kKPnpA9VlMrG4fmstivt2/
fomXzxmjarAX2HTwMM1MzPHnyGMZGxmJHK5YK1taYM38hnJ0rIjc3F7/+shW9unfBpSs34Fq1qtjx8rXuh9
5wdSoPn5nbEZOQgn7t3HFozXDU7v0jXiUkw6FCOYT9PBpbQy5j7k/
HkJL+Hq5Olnif9UG2j42zvoGRvjZ6+W/
G6+Q09GlbG78uGIRGg5bj1qOXIvbuyxT5GM8zlkXk7e0NiUQiW8qVKwcvLy/
cvn1b7GglJi0tDT6DBmDNug0wMlbug+g/
2nq1w6zAuejStZvYUQS1MmgZhnzri0HeQ1DF1RWr1q6Hto4OtgZvFjtakV26eBEdOnWGV/
sOsLO3R7fuPdGyVRtcvXpF7GjFpqqmBktLS9liamoqdiRBlMV+LVu8CNbWNtiwcTPq1q0HewcHtGrdBo5OT
mJHK5YOHTvBq117OFesiIqVKmH2nHnQ09ND5OVLYkfLl5amOrq2cMO0lQdx/
sZTPP37Neb9fAxRL17Dt0dDAMDs79vj2IUHmLYqFLcevcSzl29w6Ow9JLxNk+2nvps91u48h6v3o/
H8ZSIWbT6JpNR3qFXFWqyuFYgiH+NZWBaDl5cXYmJiEBMTg7CwMKipqaFjx45ixyox48aMglf79vBs2UrsK
PQFWVlZuHH9mtw4qaiowNOzFSIvXRQxWfHUb9AAp8NP4fGjRwCA27dv4cKFCLRp6yVysuKLevIYjnYV4Ori
hCGDBuBFdLTYkQRRFvt1KPQgaru7o3/
f3rCrYIH6dWtj86afxY4lqJycHOzauQPp6enwqN9A7Dj5UlNVgZqaqtzZRwB4n5mNhjUdIJFI4NWoCh5HJy
Bk5TD8dWw2zm4Zi07Nqsm1v3T7+f+1d99hTWRfH8C/
Q0cpogiCgjSliIII9g4KirooYm9YsCtWRFHBXl7X3hZde8Gu6+rasRcsYENRxIYFKwpKS877B7/
MErGuYBI9n+fh0cxMknMzycyZO7egdSNXGBkUgSAICGjkCh1tDRy7kPgji/
NNlP0Yz4nld9DW1havxF1dXTFq1Cg8ePAAz549AwCEhISgfPnyKFKkCGxsbDB27FhkZ2fLvcakSZNgYmICf
X199OzZE6NGjYKrq+tn3zczMxNv3ryR+ytsm6M2IvbSRUyYNLXQ34t9n+fPn0MikcDExFRuuYmpKZ48eaKg
qL7fsBGj0DqgLSpXcoRhUS3UrOqG/gMHo137jooO7bt4VK2GP5atwM6/9mLu/
EW4ezcJXg3r4u3bt4oO7bv8rOVKSrqDyKVLYGtnh527/0Gv3n0wfMhgrF29StGhfberV67AuJgeDItqY1D/
Pojash2OTk6KDuuj0t5l4szlJIT2aAQzYwOoqQlo16QKqlW0QiljA5gU14N+UR0M79oQB07fQPOBS7Er+go
2zuiG2m7/1i53Cl0FTQ11PDo0CamnZmD+6AC0HbECdx4+V2DpPk/Zj/
Gq3+BFSaSlpWHt2rWws7NDiRIlAAD6+vpYuXIlzM3NceXKFfTq1Qv6+voYOXIkAGDdunWYPHkyFi1ahFq1a
mHjxo2YNWsWrK2tP/teU6dORURERKGXSebhgwcYMSwYf+3ZDx0dnR/
2vozltXXLJkRtXI8Vq9fB0akCLsfFImT4EJiZmaNT566KDu8/8/ZpIv6/
YqVK8KhaDQ52Vti6ZRO6BfZQYGTf52ctl1QqhVsVd0yYNAUA4Fq5Mq5fu4plkUvRqYvqfg8BoLy9Pc6ej0V
qaiq2b9uCXt27Yv+ho0qbXHYftx5Lx7XDnb3hyMmRIPZmMjbtv4TKDmWgJggAgN1Hr2H+hmMAgMsJj1Ctkh
V6taqBExdzayTH92mCYvq6aNJvMV68Tkfzes5YO7UrvHotwLXExwormyrjxPI77N69G3p6egByO4GYmZlh9
+7dUFPLrQgOCwsTt7WyssLw4cOxceNGMbGcP38+evTogcDAQADAuHHjsH//
fqSlpeFzQkNDMXToUPHxmzdvYGFh8ZlnfJ+LFy8gJSUFNatVEZdJJBKcOH4MSxYtxOu0DKirqxfa+7NvY2x
sDHV1daSkPJVbnvL0KUqVKqWgqL7fmNCRGDY8BAFt2gEAnJ0r4sH9e5g1Y5pKJ5YfKlasGOzKlced27cVHU
qB+lnKVcrMDA6O8r3b7R0csWP7NgVFVHC0tLRga2cHAHCrUgUXzsdg4fy5WLB4qYIj+7ik5Bdo3Hshiuhow
aCoNp68eIs1UzojKfkFnr9OR3aOBPFJ8jV4N5NSUNM1t/
LGunQJ9G1bB25tpyP+Tu7x8sqtR6hV2Qa9A2ph0LQtP7xMX0PZj/F8K/
w7NGjQALGxsYiNjcW5c+fg7e2NJk2a4N69ewCAqKgo1KpVC6VKlYKenh7CwsJwP08bo5s3b6Jq1apyr/
nh44/
R1taGgYGB3F9hatDQEzEXL+NMzCXxz62KO9q174gzMZc4qVQyWlpaqOxWBUcOHxKXSaVSHDlyCFWVtL3U13
j/
7p140Sajpq4OqVSqoIgKR1paGpLuJKKUmZmiQylQP0u5atSoJbbzlbl9KwGWlmUVFFHhkUqlyMzMVHQYX/
QuIwtPXrxFMX1deFV3wO5jV5GdI8GF6/dRvqx8D+lyliVx//ErAEARHS0AgFQqP2SZRCKFmprwY4L/
D5T9GM81lt+haNGisPvf1R0ALFu2DIaGhoiMjISvry86duyIiIgIeHt7w9DQULzVrWr09fVRwVm+wXPRokV
RvETxfMtVTVpaGhLz1KDcTUpCXGwsjIoXh6WlpQIj+z6DgoeiV/
euqFLFHe4eVbFg3hy8S09Hl66Big7tP2vi2xwzpk+BhYUlHJ0qIC7uEhbMnY3OKlwmAAgNGY6mvs1haVkWj
x8/wqQJ4VBXV0dA2/aKDu27/KzlGjA4GA3r1sKMaVPg37oNzsecw5/
LIrFgkXLW6n2tsWNC4e3TBBYWlnj79i2iNq7HsaPR+GvPPkWH9kle1e0hCAIS7qXAtowxpgxujoS7KVi9K3
c8x9lrorFmSmecuHQHR8/
fRuMaDmhaxwnefRYBAG7efYrb959hQWgAQuf+hRep6WhR3xme1cqj1ZDlCizZlynzMZ4TywIkCALU1NTw/
v17nDp1CmXLlsWYMWPE9bKaTBl7e3vExMSgS5cu4rKYGNUfOkWVXLxwHt5eDcTHISNymxh06twVkX+uVFBU
3y+gTVs8f/YMEyLG4emTJ6jk4oqdu/+Bqanpl5+spGbNnocJ4WMRPLg/nqWkwMzMHN17BiF0zDhFh/
Zdkh8mo2vnDnj54gWMS5ZEzZq1EX38NEqWLKno0L7Lz1oud3cPbNy8DePDRmPq5ImwsrLGjFmz0a6Dancie
5aSgh6BXfDk8WMYGhrCuWIl/LVnHzy9Gik6tE8y1NPBhP6+KG1SDC/
fvMPOw5cxftEe5Ehy72Lsir6CgVO3YEQ3T8wa1hIJ91PQPmQlTsUlAQByJFL4BUdi0oBm2PJ7D+gV0ULigx
foGb4B+07FK7JoX6TMx3iBVH3aCgXp1q0bnj59ihUrVgDIHTR3wYIFWLx4MQ4fPow3b97A398fa9asgYeHB
/7++29ERERAIpHg9evXAHI77/Tq1QuLFy9GzZo1ERUVhZkzZ8LGxgaXLl366ljevHkDQ0NDPHn+utBvi/
9ogqC8tyNYfh/eUvpZ8NeQKdrPeiw0qjH0yxupoFenf1d0CAXuzZs3MC1hiNTU1M/mGlxj+R3+
+ecfmP2vvZC+vj4cHBywefNm1K9fHwAwZMgQDBgwAJmZmfD19cXYsWMRHh4uPr9jx464c+cOhg8fjoyMDLR
p0wbdunXDOSWZlokxxhhj7FtwjaWSadSoEUqVKoU1a9Z89XO4xpIpC66xZKxw/
KzHQq6xVB1cY6kC3r17hyVLlsDb2xvq6urYsGEDDh48iAMHDnz5yYwxxhhjSoYTSwUSBAF79uzB5MmTkZGR
AXt7e2zduhVeXjxlImOMMcZUDyeWCqSrq4uDBw8qOgzGGGOMsQLBA6QzxhhjjLECwYklY4wxxhgrEJxYMsY
YY4yxAsGJJWOMMcYYKxCcWDLGGGOMsQLBiSVjjDHGGCsQnFgyxhhjjLECwYklY4wxxhgrEJxYMsYYY4yxAs
GJJWOMMcYYKxCcWDLGGGOMsQLBiSVjjDHGGCsQnFgyxhhjjLECwYklY4wxxhgrEJxYMsYYY4yxAsGJJWOMM
cYYKxAaig6AFRxBECAIgqLDYL8wNTX+/qkKqZQUHUKh+Fm/
gxnZEkWHUChenf5d0SEUCqM6oxQdQoGjnMyv2o5rLBljjDHGWIHgxJIxxhhjjBUITiwZY4wxxliB4MSSMcY
YY4wVCE4sGWOMMcZYgeDEkjHGGGOMFQhOLBljjDHGWIHgxJIxxhhjjBUITiwZY4wxxliB4MSSMcYYY4wVCE
4sGWOMMcZYgeDEkjHGGGOMFQhOLBljjDHGWIHgxJIxxhhjjBUITiwZY4wxxliB4MSSMcYYY4wVCE4sGWOMM
cZYgeDEkjHGGGOMFQhOLBljjDHGWIHgxJIxxhhjjBUITiwZY4wxxliB4MSSfbUlixbC3s4KxfR0UKdmNcSc
O6fokAoEl0u1cLlUx9u3bzFiWDAcylmhhGERNKxXCxfOxyg6rO9y4vgx+Ps1h7WlOXQ1BezauUPRIX2z5X8
sQa2qlWFpagRLUyM0rl8LB/btldvm3NnTaNHEC6WNDWBpaoSmjerj/fv3Cor4v1PV/
aVXRAszg5vh5rYQvIyeiCN/9EUVxzLi+j/CAvD+9DS5v52zA+VeY/
OMLkjYPgqvoifizl+jsXxcG5gZ6xd67D9FYhkeHg5XV9dPrl+5ciWKFSv2Xe/
RrVs3+Pn5fddrqLLNm6IQMmIoxoSNx+lzF1Gpkgta+HojJSVF0aF9Fy6XauFyqZb+fXrhyKGDWPbnapy7cB
meXo3QrEkjPEpOVnRo/
1l6ejoqVnLBnHkLFR3Kf2ZeujTGT5iMIyfP4fCJs6hTrwE6tmmF+OvXAOQmla1/80UDz0Y4eOw0Dh0/
g159+kNNTfVSBlXdX4tD/
dHQoxy6T9gE905zcPDsLfw9ryfMSxqI2+w7fRNWvpPEv67jNsq9xrGLd9ApbB1c2s1Ch9FrYVOmBNZP6VTo
sQtERIX+Ll9w+vRp1K5dGz4+Pvj777+/+fnh4eHYsWMHYmNjP7p+5cqVCA4OxuvXr/
9zjN26dcPr16+xY8eO//waX6N+/fpwdXXFnDlzvvo5b968gaGhIZ6+SIWBgcGXn/
Af1KlZDVXcPTBn3gIAgFQqhZ21Bfr2H4gRI0cVynv+CFwu1cLlKjhSaeEe+t+/fw/
TEgbYtGUHfJr6istrVXdHY28fjI+YVCjvq6YmFMrrfoyupoCoLdvR4je/
Qn+vjGxJob6+demSmDB5Ojp3645G9WqifkMvjBk/oVDfEwB0NNUL/T1kfuT+Mqrz33+3OtoaeHYwAgEhq/
HPqZvi8pMrBmD/6QRE/LEff4QFoJieDtqMWvPVr+tb2xGbpneGYd0w5Eik3xwX5WQi8/
wcpKZ+PtdQisuP5cuXY+DAgTh27BgePXqk6HDYB7KysnDp4gU09PQSl6mpqaFhQy+cO3NagZF9Hy6XauFyq
ZacnBxIJBJo6+jILdfV1cXpUycVFBX7kEQiwdbNUXiXng6PatXxLCUF52POoaSJCRo3qI3yVubwbdwAp0+d
UHSovwwNdTVoaKgjIytHbnlGZg5quliJj+u42eDe32GI2zgMc0f4obhBkU++ppGBLtp5u+LMlfv/
Kan8FgpPLNPS0hAVFYW+ffvC19cXK1eulFsfHR0NQRBw6NAhuLu7o0iRIqhZsyZu3rz58RcEkJiYCBsbGww
YMACfqpDduXMn3NzcoKOjAxsbG0RERCAnJ+ej2+YVERGBkiVLwsDAAH369EFWVpa4LjMzE4MGDYKJiQl0dH
RQu3ZtxMTItyc6evQoqlatCm1tbZiZmWHUqFHi+3br1g1
Hjx7F3LlzIQgCBEHA3bt388WQmZmJN2/
eyP0VpufPn0MikcDExFRuuYmpKZ48eVKo712YuFyqhculWvT19VGteg1MnzoJjx89gkQiwYb1a3H2zGk8ef
xY0eH98q5dvYIyJQ1hWqwIhg7qhzUbt8DB0Ql3794BAEybPAFdA3tiy46/4eJaGX5NGyPx9i0FR/
1rSHuXhTNX7iE00BNmxvpQUxPQztsV1ZwtUapEbhvJA2duoueETWg6KBJhi/aiTmVr7JwdmK/GflI/Hzw/
PAGP9o2HhWkxBIxcXejxKzyx3LRpExwcHGBvb49OnTrhzz///GgyOGbMGMyaNQvnz5+HhoYGunfv/
tHXu3z5MmrXro0OHTpgwYIFEIT8t0WOHz+OLl26YPDgwbh+/
TqWLl2KlStXYvLkyZ+N9dChQ4iPj0d0dDQ2bNiAbdu2ISIiQlw/
cuRIbN26FatWrcLFixdhZ2cHb29vvHz5EgCQnJyMpk2bwsPDA3FxcVi8eDGWL1+OSZNybwnNnTsXNWrUQK9
evfD48WM8fvwYFhYW+eKYOnUqDA0Nxb+PbcMYY4q27M/
VICLYWZeBkb4OFi+cj4C27VWyrd7Pplx5exw7cwEHj55C91690S+oO27EX4dUmlub1a17L3Ts0g2VXCtjyo
zfYVfeHmtXr1Bw1L+O7hFREATgzl9jkHp0Evq3qYVNB+Ig/V9+tPngZfx9Ih7XEp/
ir2PX0Wr4Krg7WaCum43c68xedwzVu86D76BlkEgJy8a1KfTYFf7rXr58OTp1ym1M6uPjg9TUVBw9ejTfdp
MnT0a9evXg5OSEUaNG4dSpU8jIyJDb5tSpU6hfvz6GDx8uJmsfExERgVGjRqFr166wsbFBo0aNMHHiRCxdu
vSzsWppaeHPP/9EhQoV4OvriwkTJmDevHmQSqVIT0/
H4sWLMXPmTDRp0gROTk6IjIyErq4uli9fDgBYtGgRLCwssGDBAjg4OMDPzw8RERGYNWsWpFIpDA0NoaWlhS
JFiqBUqVIoVaoU1NXztz8JDQ1Famqq+PfgwYMvfs7fw9jYGOrq6khJeSq3POXpU5QqVapQ37swcblUC5dL9
djY2mLfwWikvHyLm4n3cezkWeRkZ8PK2ubLT2aFSktLCza2dnB1q4LxE6bAuWIlLFk4H6VKmQEA7B2d5La3
t3fAw0I+17B/JSW/RON+f6BEg7Eo5zcNdXoshKaGGpKSX350+7uPXuLZqzTYlikht/
xF6jvcfvAch2Nuo8vY9WhSywHVnC0LNXaFJpY3b97EuXPn0L59ewCAhoYG2rZtKyZieVWqVEn8v5lZ7hc/
b4/J+/fvo1GjRhg3bhyGDRv22feNi4vDhAkToKenJ/7JagnfvXv3yee5uLigSJF/
2zDUqFEDaWlpePDgARITE5GdnY1atWqJ6zU1NVG1alXEx8cDAOLj41GjRg25WtRatWohLS0NDx8+/
GzMeWlra8PAwEDurzBpaWmhslsVHDl8SFwmlUpx5MghVK1eo1DfuzBxuVQLl0t1FS1aFGZmZnj16hUOHtiH
Zs1bKDok9gGpVIqsrExYlrWCmZk5bifINze7fesWLCwKNyFh+b3LyMaTF29RTF8XXtXKY/fx6x/
drnRJA5QwLIInz99+8rVkt8m1NDUKJVaZwn31L1i+fDlycnJgbm4uLiMiaGtrY8GCBTA0NBSXa2pqiv+XJW
ayKnsAKFmyJMzNzbFhwwZ07979s8lWWloaIiIi0KpVq3zrdD5oaM5yDQoeil7du6JKFXe4e1TFgnlz8C49H
V26Bn75yUqMy6VauFyq5cD+fSAilC9vj8TE2xgTOhLl7R3QWYXLlZaWhsTbt8XHd5OSEBcbC6PixWFpqRqJ
V8S40fBq7AMLC0u8ffsWWzZtwIljR7F11x4IgoCBQ4Zh6qQIOFdyQcVKLtiwdjVuJdzAqvVRig79m6nq/
vKqVg6CICDh3jPYlimBKQOaIuHeM6zefR5FdbUwpocndhy5iicv0mBTpjgm92+CxIcvcOBsAgDAw8kCVZzK
4FTcXbx++x7WpUtgfFAjJD58jrNX7xVq7ApLLHNycrB69WrMmjULjRs3llvn5+eHDRs2oE+fPl/
9erq6uti9ezeaNm0Kb29v7N+/H/
r6Hx8I1M3NDTdv3oSdnd03xRwXF4f3799DV1cXAHDmzBno6enBwsICxsbG0NLSwsmTJ1G2bFkAQHZ2NmJiY
hAcHAwAcHR0xNatW0FEYnJ88uRJ6Ovro0yZ3IFPtbS0IJEU7rAS/0VAm7Z4/uwZJkSMw9MnT1DJxRU7d/
8DU1PTLz9ZiXG5VAuXS7W8eZOK8WGjkZz8EEbFi8PPrxXGT5gsV1Ggai5eOA9vrwbi45ARQwEAnTp3ReSfK
xUU1bd5nvIMfXsG4umTxzAwNEQF54rYumsPGng2AgD0HTAYGRkZGD1yGF6/eokKFSth2+5/
YG1jq+DIv52q7i9DPR1M6OOD0iaGePnmHXZGX8X4JfuQI5FCQyqFs60ZOjapgmL6Onj8/
C0Onk3AhD8OIOt/w1K9y8zCb/
WcEdbTC0V1tPDkxVvsP5OA6SsPi9sUFoWNY7ljxw60bdsWKSkpcjWTABASEoLDhw8jJiYG0dHRaNCgAV69e
iUOch4bG4vKlSsjKSkJVlZWcuNYpqWloUmTJiAi/
PPPP9DT08s3juW+ffvQrFkzhIWFoXXr1lBTU0NcXByuXr36ybaZ3bp1w9atW9G8eXOEhYXh7t276N69OwID
AzF16lQAQHBwMDZv3ozly5fD0tISM2bMwK5du5CYmAgjIyMkJyejfPnyCAwMxIABA3Dz5k307NkT/
fv3R3h4OAAgKCgIsbGx2LRpE/T09FC8ePEvNnT/EeNYMsZ+LoU9jqWi/
MhxLH+kwh7HUlF+5DiWP9L3jGOprJR+HMvly5fDy8srX1IJAP7+/jh//
jwuX778za+rp6eHvXv3gojg6+uL9PT0fNt4e3tj9+7d2L9/
Pzw8PFC9enXMnj1brGn8FE9PT5QrVw5169ZF27Zt0aJFCzEhBIBp06bB398fnTt3hpubG27fvo19+/
bByMgIAFC6dGns2bMH586dg4uLC/r06YMePXogLCxMfI3hw4dDXV0dTk5OKFmyJO7fv//
NnwFjjDHGmCIoxcw77PtwjSVj7FtxjaVq4RpL1cI1lowxxhhjjH0nTiwZY4wxxliB4MSSMcYYY4wVCE4sGW
OMMcZYgeDEkjHGGGOMFQhOLBljjDHGWIHgxJIxxhhjjBUITiwZY4wxxliB4MSSMcYYY4wVCE4sGWOMMcZYg
eDEkjHGGGOMFQhOLBljjDHGWIHgxJIxxhhjjBUITiwZY4wxxliB4MSSMcYYY4wVCE4sGWOMMcZYgeDEkjHG
GGOMFQhOLBljjDHGWIHQUHQA7PsREQDg7Zs3Co6EMaYqpFJSdAiFQk1NUHQIhSIjW6LoEApFlqa6okMoFJS
TqegQChxJcsskyzk+hRPLn8Dbt28BAHbWFgqOhDHGGGM/
s7dv38LQ0PCT6wX6UurJlJ5UKsWjR4+gr68PQSjcq/U3b97AwsICDx48gIGBQaG+14/
E5VIdP2OZAC6XquFyqRYu1/cjIrx9+xbm5uZQU/t0S0qusfwJqKmpoUyZMj/0PQ0MDH6qH6cMl0t1/
IxlArhcqobLpVq4XN/
nczWVMtx5hzHGGGOMFQhOLBljjDHGWIHgxJJ9E21tbYwfPx7a2tqKDqVAcblUx89YJoDLpWq4XKqFy/
XjcOcdxhhjjDFWILjGkjHGGGOMFQhOLBljjDHGWIHgxJIxxhhjjBUITiwZY4wxxliB4MSSMcYYY4wVCE4sG
UPuVFUAcP/
+fQVHwn5lRASpVKroMArMxYsXFR0CYz+dvIP5KOPxghNLxgAIgoAdO3YgICAA165dU3Q4hYJHFlNemZmZAH
K/
hw8ePFBwNAXj9OnTcHd3x8KFCxUdSqGRSCSKDqHAyY4TR48exYEDBxQcDfsYQRDw5MkTxMfHQ01NDVu2bMG
2bdsUHZaIE0v2VZTxqqggyA6iDx48wNy5c9GzZ09UqFBBwVEVDFnZrl+/
DolEAkEQFBwR+5jExESMGTMGr169wubNm2FtbY3ExERFh/
XdatSogUmTJmHo0KFYvHixosMpELLj4Nu3bwEA6urqiI2NxZMnTxQZVoGQHS8EQcCRI0fQtGlTpKenIycnR
8GRfb8Pz1+qfpGdmpqKDh06YPbs2Zg7dy7atGmD9PR0RYcl0lB0AEz5SaVSqKnlXoP89ddfSE5OhpWVFcqV
KwdbW1sFR/
d9BEHA8ePHsXPnThgaGuK3335TdEgFRhAE7Nq1C0OHDsXq1atRs2ZNRYf03YgIgiDgzJkzSE9Ph6enp6JD+
m5XrlzB0qVLce3aNURHR2PFihWwtbUVy6rKRo8eDXV1dQwYMAAA0LdvXwVH9H3U1NTw6NEjBAUFoX///
sjKykLLli1x9uxZlCpVStHhfRfZd+3Ro0c4f/48Ro8eDT8/P5VPwgCI56/Y2Fi4urqq/O/
K0NAQPXr0QHh4OJYtW4YpU6agc+fOynPMIMa+0siRI0lPT48qVapExYoVo7p169KqVasUHdZ3+/
3330kQBDI0NKQLFy4oOpzvJpVKiYjo0aNH1LJlS1q0aJGCIyoYsnJt3bqVzM3NqU+fPvTw4UMFR1UwQkNDS
RAEatiwoVyZZGVWddOmTSM1NbWf4rt46dIl8vf3pwoVKpC2tjatX7+eiIgkEomCI/
s+UqmUkpKSSBAEKl68OM2cOVPRIX23vPvk+PHjZGJiIu4vVSU7Jty9e5esrKzIwsKC+vfvT1evXs23jaLwr
XD2VWJiYnDw4EHs27cPcXFxOHDgAMqXL4+5c+ciKipK0eF9lyFDhiAyMhJqamr4888/
cffuXUWH9F0EQcCxY8cwfPhwvH79Gg0aNACg+rd/
BEHAgQMH0KlTJ0ycOBGzZ89G6dKlFR3Wd5G10dPR0cGQIUNw69YtTJ48GTdu3ACQW2ZV328AEBISgsmTJ2P
AgAEqe1uc/
texytXVFc2aNcP169dhaWkJfX19ALm1YqraZIj+V9NlZWWF2bNn49WrV7h06RKeP3+u6ND+s7x32tauXYv1
69cjPT0dISEhWLdunYKj++9kNZImJiY4fPgwJk6ciFOnTmHu3Lli/
wCF11oqNK1lKmHatGnUvXt36tixo9wV4NWrV8nf35/
atWtH2dnZCr9K+hqyGBMSEujcuXN08OBBcd28efPI3NycRo8eTffu3VNUiAXi8OHDZGxsTGpqarR161ZxuS
rso0/JzMykvn370rBhw4iI6PXr1xQTE0PBwcE0btw4unHjhoIj/
H4bNmygMmXKUJ8+feTKExcXp8Covp7s+3Xt2jU6fvw47d27V279lClTVL7mcuPGjdS8eXNatmwZdezYkWrX
rk1RUVHielWquZTtrw9jnjVrFgmCQNOmTaPU1FRFhFZgQkJCqFSpUrR48WKaNm0a1alTh8qVK0crVqxQdGj
fRLav7t27R9evX6fExERxXWRkJFWuXJl69+4t1lxOnDiRtm/frohQiRNL9kVjx44lQRDI2tqa7t+/L7du/
fr1pKGhQXfu3FFQdF8v761UBwcHcnBwICcnJ3Jzc6MHDx4QEdHcuXOpdOnSNHbsWEpKSlJgtN/
v5MmTZGVlRc2aNaPz58+Ly1U5uWzfvj1VrlyZkpKSqHPnztSwYUOqWbMmlSxZklq1aqXo8L6K7POPiYmhtW
vX0oIFC+jevXviyX3Dhg1kYWFBffv2pWPHjtGECRNIEAR6+fKlUu87WWzbtm0jCwsLqlChAunr61PLli0pP
j5e3G7KlCmkra1Ns2bNUlSo30xWttu3b5Oenh7Nnz+fiHL3YZs2bah27dq0efNmcft9+/
bRkydPFBLr15KV6fDhwzR48GDq3r07hYWFietnzpxJgiDQ9OnTVTa5vH37Njk4OMglWLGxsdSrVy+ysbGhD
Rs2KC64b5D33OXo6EhmZmZkZ2dHLVq0oMzMTCLKTS6rVq1K9evXp7Zt25IgCApr2sWJJZPzqavtuXPnkiAI
NGHCBHr+/Lm4/
OzZs+Tg4EA3b978USF+l6NHj5Kenh5FRkZSRkYGHT16lARBoCVLlojbzJs3j3R0dGjixImUnZ2twGi/
juygc/nyZdqxYwetW7eOUlJSiCi3vNbW1tShQwe6ePGiIsP8ZrJynT9/
XqxZPnXqFFWuXJm0tbUpICCAtm3bRkS5yYyrqyu9fPlSYfF+jbwniOLFi1PDhg3J1NSUvLy8aMWKFZSTk0N
ERJs2bSJHR0dydnYmCwsLOnfunCLD/qy8ye7+/
fupWLFiFBkZSUS5FzeCIJCvry9duXJF3C4sLIxKlChBr169+tHh/
mfHjh2jlStXUmhoqNzy8+fPU9u2bal27do0e/
ZsCg8PJ0EQVKL977Zt20hPT4/69+9PI0aMIDs7O3J1daWsrCwiyq251NLSovDwcHrz5o2Co/
12Dx48ICMjI1q9erXc8kuXLpGVlRWZmZnRmjVrFBTdtzly5Ajp6urS4sWL6dChQ7RlyxaysbGh6tWri8eNq
KgoGjx4MLVq1Uru9/ajcWLJRHmTyjt37tDVq1flkkhZzcmwYcMoOjqarl69Sj4+PuTh4aEyt39mzZpF/
fr1I6LcMpYtW5b69u2bb7tFixZRQkLCjw7vP9uyZQuVLVuW3NzcqEaNGqSnp0eHDh0iIqLo6Giytramzp07
K3WCklfeBMzCwoKGDx9OycnJlJ2dTenp6fnKMWjQIGrSpAmlp6crItxvEh0dTaamprRs2TIiIrpy5QppaGh
Q1apVacmSJeJv6cqVK3TmzBmxNl3ZbNu2ja5fv05EufvrzZs3NGjQIAoPDyei3N+XjY0NdezYkczNzalBgw
YUFxcn7tu8xxZlExwcTDNmzBAfp6amko+PDwmCQC1btiQikrvovHTpEgUFBZGDgwNVqFBB7g6BskpOTiZnZ
2eaN28eERElJSVRqVKlqGfPnnLbRUREkJGRkVLvL6J/jxl5/33+/Dk1atSIhg0bli/
+gIAAqlu3Lnl4eNCBAwd+eLzfKiIiIt9dmcTERLKysqKAgAC55bILA0XhxJIRkXytQ2hoKFWsWJF0dHSoVq
1aYiJGRDRp0iQSBIEEQaCuXbtSq1atxC+xKiSXnTp1osDAQHrx4gVZWFhQUFCQWPaVK1eqZE/
Is2fPkpGRkVhLdO3aNRIEgaZMmSLuk+joaDIwMKBevXpRRkaGIsP9av/88w/p6urS0qVL6f379x/
d5vz58zRs2DAqVqyYSrRDzM7OpmnTplFwcDAR5Z4YZMmXj48P2djY0LJly8QaCGV1+fJlcnFxoZYtW4oXYJ
mZmbR9+3ZKSEigly9fUpUqVahHjx5ERLR7924SBIFq165N165dU2ToX5STk0PLli3LV8N/
4sQJ8vf3JwMDA7HMeU/gr1+/pqdPn4p3C5Td9evXqVy5cpSVlUUPHz6kMmXKUO/evcX1u3fvFv//
4sULRYT41fKeez5Mqn7//XcqVqwYzZ07V9w3b968odatW9OiRYuoVq1aNHr06B8a73/RtWtXcnd3Fx/
LLmxWrFhBFSpUUKoLUE4smZzp06dT8eLFaffu3XTkyBGaOHEiOTs7y10pLViwgARBoHnz5tHr16+JiJTyRJ
i3ZuTdu3dElFvL4u3tTSVLlqRevXoRUe5BSSKRUP/+/
alfv37itqpi3bp11KFDByLKrSWStc+Tefv2LRHl3sq7deuWQmL8VhkZGdS5c2caOXIkEeXWGF28eJFCQ0Mp
IiKCXrx4QZcvX6aBAwdS5cqVVSKplImPj6fr169TWloa1apVi7p3705EucOHFCtWjCpUqCDWZiqzP//
8k+rXr0+tW7cWay5lFy2bN2+mqlWrim2vd+zYQU2aNCFHR0eVaru8Z88eGj9+vPg4JiaGGjRoQBYWFnT79m
0iIpVoLpPX1atXSSKR0OPHj6levXq0detWsrS0pN69e4tlSUhIoE6dOtHx48eJSLnbZedNKhctWkRt2rShd
u3a0dSpU8Xl48aNIxMTE2revDkFBQVRjRo1yM3NjYhyKxs8PT2VuoxERH///
TfZ2trSxo0b5Zbv2LGDrK2tKTk5WUGR5ceJ5S8u748pNTWVfH196ffffxeXpaen08aNG8nJyYn+7//+T1w+
efJkEgSB5syZo5TtpGTl2rVrFzVu3JgOHDhAEomEbt68SbVr1yZbW1vat28fEeXWNIwZM4ZKlSol18lAWX1
4AIyIiKCGDRvSvXv3yNLSkoKCgsSD7bZt2yg4OFjlkmUiog4dOlCdOnXo9u3bFBgYSA0bNiR3d3cqWbKkmE
hfvXpVqTtJfOxkJTt5Hzt2jJydncUavJiYGPL09KTOnTsr9agEeROpP/
74g5o2bUoBAQFyvVRnzJhBtra24r4JDQ2lqVOnqlQSJpVKadGiRSQIAk2cOFFcHhMTQ97e3mRlZSUmzsp4Y
U2UP+m9cuUKlSlThu7fv0+vXr2i+vXrk5qaGnXs2FFuu+HDh1P16tWV+rf1IVnv79GjR9Pw4cPJ2tqaAgMD
xfVr166lIUOGUOPGjalv377iRVCrVq0oODhYae64yY4ZycnJlJiYKLYbf/
r0KbVs2ZJ8fX3FTkdZWVk0atQoqlq1qlK1L+fE8hf24Q9JKpWSm5ub3K1votwvr7+/
P7Vv315u+fTp00kQBFq0aJFSXu3JGqZPmDBBrF0gIrpw4QJVrlyZnJ2dycHBgby8vMjc3FylOrecOHFCvCI
/
deoU1a9fn4oXLy4eSGX7Njg4mDp06KD0De8/9v3ZuXMneXh4kLq6OgUEBNCWLVuIKLfJgoeHh9K3p5SV6cS
JEzR9+nQaNWoUHTx4UDyhHThwgKytrWnnzp0kkUho/
Pjx1KNHD7GGWVl92JvY2dmZ1NXVqU2bNuIt4oSEBDIwMKBKlSpR3bp1ydDQkGJjYxUZ9n/
y7t07Wrp0KampqYltR4mIzp07R02bNiUDAwOlrYGdOXMmtW7dWu53cvr0aXJwcBBvF1+7do2MjY3J19eXVq
9eTfv376cBAwaQoaGhSt0FWL9+PZUvX57OnDlDRLk15kWKFBFHJZDJe85LSUmhMWPGUPHixcUad0XLO7JCu
XLlyNramgwNDWngwIF0584dSkpKotatW1PZsmXJ3t6eGjRoQEZGRkp37uLE8hcVExNDz549I6LcGXVWrlxJ
REQDBw4kb2/
vfD+08ePHk5eXF2VkZMj9OGfPnq00P8q8kpKSyNbWlhYsWEBEuQeUrKwsOnv2LL1//55evHhBu3btopEjR9
KGDRtUYrgkmczMTAoODqaGDRsSUW57IVkHiWXLllF2djY9fvyYQkNDydjYWOnbtOVNwMLDw2nUqFFiL863b
9/SqVOn5Lbv27cvNW/e/JPtLpXJli1bSE9Pj+rVq0fVqlUjQRBo+PDh9ODBA3rx4oU4pp6Tk5NSniA+Zf/
+/SQIAs2ePZv++usvCgkJoYoVK1Lr1q3FWv8rV65Qr169aMSIEUr/
HST6t9bx4cOHcncupFIpLVy4MF9yeerUKfL391fa5iX//
PMPaWtry12s7N27l1xcXIjo3yTrwoUL1LBhQypbtiw5OjqKnayUWWZmplzCvHjxYho3bhwR5d6lMjIyojlz
5lBkZCSpq6uLbX1lXrx4QYGBgWRjY0OXLl36kaF/UXR0NOnq6tLs2bPpwoULNH/
+fKpRowa1bNmSkpKS6Pnz53Ts2DEaPnw4zZ8/
Xyk7mXJi+QtKSUkhQRBo4MCB1Lt3b9LX1xeHJoiNjSVTU1Pq3LmzOAbW27dvqX79+mKbRCLl7agjS1Ju3Lh
BVapUoQsXLtDz589p5syZVK9ePTI0NKS6devSyZMnFRzp9zl//
jxpa2uLt0RevXpFvr6+VLFiRSpWrBjVrl2brK2tVSZR2bp1KxkaGlKHDh2oe/fuZGRklK+G/OrVqzR06F
AqVqwYXb58WUGRfr1bt26RpaUlRUZGit/
LDRs2kLGxsTjI+71792jp0qU0d+5cpTxBfEgqlZJEIqGePXtSu3bt5Nb98ccf5OjoSG3atBGTrZycHKW8my
GzaNEiOnz4sHjLePPmzWRhYSGOw3n48GGxdk+WXOa9La7sFzdHjhwhPT09CgwMJIlEQjt27CBXV1cikr9Lk
JGRQU+ePKGUlBSlrzHfsmULtWrViipXrkwTJkwQl9+5c4eeP39Obm5uNG3aNCLK/
Q2WLl2aBEGgkJAQude5d+9evnGZFUm2P4YNGyZXy0r0790bWZtzZceJ5S/
m6NGjlJSURBcuXCBtbW3S1dWlw4cPE9G/
V+xnzpwha2trcnNzI2dnZ6pWrRo5OzuLB1hlPlHIbvneu3ePihcvTt7e3mRqakp+fn40depU2rdvHzk6Oqr
UzB95P2+JRCI+Hjp0KHl6eooHx/T0dDp//
jwtWbKEjhw5olS9BD9H1jNaVrt869YtKl68OAUFBYnbnD17lvr06UMuLi5Ke0s1JSWFYmJixAuyK1eukI2N
DcXGxsrtw3Xr1pGamhodO3ZMUaF+t/79+5OXl1e+HrjBwcGko6ND3t7eSj22rWx/
2Nvbk6WlJZ06dYouX75M1tbWNHPmTDpy5Ah5e3uTpaUlbd68WRyEesmSJeKg4ari0KFDpKenR4MGDaJNmzZ
RjRo1aP/+/
RQdHU3Xrl2jCxcu0K5du+jx48eKDvWLlixZQgYGBjRkyBAKDg4mdXV1Wrhwobj+7NmzZGlpKV6k3bp1izp0
6EAHDhyQawerzOewoUOHkpeXF+Xk5MhV4MyYMYOMjY2VPvEn4sTyl/
LmzRvq0aMHjRw5kk6fPk06OjqkpqZGQ4YMoUePHhHRvz+4W7du0YYNGygkJIQWLFggXtErc+P72NhY0tbWp
tOnTxNR7jR4o0aNolmzZsk1Qvfy8qK5c+cqKsz/
5MCBA7R9+3a5jlI7d+4kOzs7seemqrp48SJVqlSJiHIvCGRTGsrExMQQUW4trex7qmyuXbtGtWrVIh8fH2r
VqhXl5ORQTEwMaWpqiu2+8g7z5OzsLNcZTtXMnDnzozXiq1evpooVK1L79u2V9sLmw7st9erVIwcHB1q1ah
WNGDFCbp2/v3++5HLZsmVK2fwnrw8Tp4MHD1LRokWpSJEiZGtrS9bW1mRmZkb29vZUpkwZMjc3V/
rmQJGRkaSpqSk3i0779u1p3rx54vE9MTGRbG1taeDAgXT9+nXy9vamli1bip+Hsnayymv27Nmkp6cnTs0oi
33//v3k5ORET58+VWR4X4UTy1/MunXrqGzZsmL7ygMHDpCamhr179//
i1esyv6jvHv3LjVr1owMDAzo7NmzRCR/Ms/
JyaHQ0FAyNTWV68yj7N69e0cDBw4kQRDIz89PvM1DRNS5c2e5sc1UgexAeeTIETp48CBdv36datasSQcOHM
g37ElcXBx16tRJqecBv3r1KhUrVkycYz5v4hIQEEBOTk5yPaYzMzOpSpUq9Mcffygi3G8i21fx8fEUFxcn1
wTBw8ODKlSoQDExMWJ7t5EjR1JoaKjSjnso2zdJSUk0f/
588ThQtWpVEgSBvL2989XC+vv7k62tLa1du1bhA09/iWx/vXnzhtLS0uTWHT16lEqWLElNmzal+/
fv04sXLyg1NZWeP38uDhunrI4cOUKCIFBERITcchcXF6pUqRLp6+tTrVq1aN68eTRr1iwqU6YMlS1blqpVq
6a0d9pkx7jExESKj4+XO8bJRh2Ii4sTf1vBwcHk5uam9PuKiBPLX0beH1WnTp2odevW4vyvf/
31F6mpqdGgQYPEachat25NUVFRCon1a+Utk+z/9+7dozZt2pCurq44+4VEIqE///yT/
Pz8qHTp0irT7vBDp06dotGjR5OpqSlVrVqV5s2bR9u2bSMvLy/666+/
FB3eF+XdX0eOHKEiRYrQtm3bKDExkdzd3UlHR4e6du0q95yhQ4dSgwYNxAshZfPixQuqXbs2DRo0SG65LIE
5ceIE+fj4kL29PR06dIiOHj1KY8aMIWNjY7lkU5lt3ryZTExMyMLCgmxtbcXRCN6/
f09Vq1Yla2tr8vDwoMaNG5OWlpbSdtSR7ZPLly9T+fLlqWXLlnK1X40aNSIjIyM6dOhQvovoRo0aUcWKFZV
6dAXZ7+vvv/+m+vXrk5ubG9WtW5euXr0q1rYePnyYihQpQn369FGpIcgSEhKoTp061KJFC/
EORqtWrcjOzo6ioqJo7969VKFCBXJ3d6e4uDhKTk6m06dPi/
tcWe60rVq1SpzIgoho48aNZGFhQSYmJmRnZ0dt2rShrKwsSklJIR8fH9LX1ycPDw9q2LAhGRoaKl1Ho0/
hxPIn97FONkeOHKGWLVuKt4yJcmdZ0NbWpsaNG5ObmxuVL19e6a/
OiXKvwmW3cGQH1rt371KbNm2oSJEi4g/xypUrNGTIEKVu9yUjK0dcXBxt376dNm/
eLDebR0pKCvXq1Ys8PT1JV1dXnGZT2a7IP+Xhw4c0c+ZMmjRpkrhsz549pKGhQUFBQbRv3z46f/
48BQcHK31HnWvXrpGtrS0dPXr0kx3azp07Rx07diRtbW2ys7OjChUqKP3Fjey79OLFC3JwcKAVK1bQ4cOHa
erUqaSpqUlhYWHitosWLaIxY8bQiBEjlP4WcXx8PBkZGdGoUaM+OqB0rVq1yMrKio4fP55vfyrrrf28du7c
Sfr6+jRmzBg6dOgQ1axZk1xcXGjPnj1icnno0CESBIH69+
+vMscMotzk0sfHh3x9falWrVrk5uYmN9TThQsXSBAE2rlzp9zzlKWjaUpKCjVr1oyqVatGGzdupEePHpG1t
TUtXryYDh8+TBs3bqQyZcpQw4YNxf0SGRlJkyZNokmTJqlE5z4ZTix/YnnbzPz+++/
isC3Z2dnUvHlz8vf3l9s+OjqahgwZQiNGjFCJNpWpqank5eVFxsbG4gFG9oNMSEggV1dXKlmypHiFq8xl+Z
CslqhcuXJkaWlJJUqUoL/+
+kvshSqVSik5OZlmzJhBLi4uYq9+ZXfnzh0SBIEMDQ3zdYCIiooiNzc3KlGiBDk7O5OHh4fSdtSRWbduHWl
oaIjfu7wnMVmtV3p6OsXHx9OzZ8/o3r17Slv7+qGDBw/
SqFGjaMCAAWJS8vbtW1qwYAGpq6vnmwZP2ZOU9+/
fU0BAAPXv319ueVZWFt25c0e8ePPx8SFLS0s6efKk0iQlX+POnTvk7u5Os2fPJiKiZ8+ekbW1NZmYmJCJiQ
nt2bNHbBp09OhRlZgM4kMJCQnk5eVFhoaGtGnTJiL6t0PjhQsXyMnJiU6cOKHgKD8tNjaWOnXqRA0aNKAhQ
4ZQx44d5Spw4uPjydzcnDp16qTAKL8fJ5Y/qdjYWBIEgXbs2EGDBw+m4sWLy4239vjxY3JwcBDHr/
xY42ZVSMROnz5NTZo0IWtr63yNz7t27UpqampkZmZG79+/
V5mTxMWLF8nIyIhWrFhBT548oSdPnlDPnj1JT0+P9u7dS0TyJ3FlHig8PT2dnj17RkeOHBGbWaxfv54EQaA
2bdrkm1f5yZMnFB8fT3fu3FHKGZ0+dPLkSdLR0REHb/
+YefPmUaNGjVRmjnai3HagY8aMIXV1dapSpYrcOllyqaOjIw6bRKT8iWV2djbVqVOH5s+fLy77559/
KDg4mAwMDKhMmTLUunVrIspNLg0NDcWOV6rg5s2bNH36dEpLS6NHjx6RnZ2dOLWrh4cHubi40I4dO8SLBFV
1+/Zt8vb2piZNmsiNrNCsWTOqX7++0h/nY2NjqWPHjmRtbU3Vq1cXl8vOt8uXLycnJye6d++e+JtS9t/
Whzix/
IlFRESQrq4u6enpyd1OzMnJoezsbIqIiKABAwbQu3fvlP7HSPTvjysrK0uuYfqVK1fI09OTrK2t6e7du+Ly
4OBg2rRpk1L3otu/f3+
+TlPbt28nNzc3evXqldwBJTAwkMzMzMSES9kPOjdv3qQuXbqQg4MD6ejokL6+PrVv356Sk5Np27Zt4lR5qt
AY/
VMePnxIJiYm1KJFC7nvXt59MmzYMBo1apTS7qe88sZ49+5dioiIEGfXyistLY1mzpxJJUqUoGfPnqlE2VJT
U8nBwYF69epFN27coClTppC9vT35+/vT3Llzafny5VS2bFlxnEpPT0+lHfz8U2Tx9uvXj/z9/
cWhaTp37kyCIFD58uXzdepRRbLb4k2bNqXjx49Tq1at5JpvKfv57MqVK9SuXTsqUqQILVmyRG7drl27qEyZ
Mko9reuXcGL5k8n7g5o0aRIJgkAaGhq0devWfNuePHmSTE1Naffu3USkvAkKkXzD9JYtW5KLiwv17NmT9uz
ZQ0RE169fJy8vLzIyMqKxY8dS586dyczMTGmH0JDNWy5r65S35i4yMpKKFCkiXsHKarpu3bpFZcqUoYMHDy
ok5m8RFxdHZmZm1KdPH1q5ciXFx8dTSEgIWVtbk729Pd2/
f1+suZwyZYrYkUwVbd26lbS1talz585yHVfS09MpNDSUypYtq/Rte2W/rw/vUty/
f59Gjx5Nenp6+U6A6enpSjU/8dc4dOgQaWhoUNmyZUlfX5+WLFkiJmNZWVnUuHHjfAPzKyPZ/
kpMTKSbN2/
mq1lt2rSp3IDgQ4YMoUuXLol3DX4GCQkJ5OvrS5qammRvby8mlapwp40o95zVvn17qlatGi1evJiIci/
YRowYQQ4ODirTZOZjOLH8SUVERFBQUBBdvXqVIiIiSFNTk9auXUtE8snnkiVLyNXVValmIPiUv/
76i7S0tGjw4ME0YcIEcnd3pxo1atC8efOIiOjRo0c0ePBgcnd3p0aNGil1DzpZk4MtW7aQpqYmDRo0SByL7
dmzZ1ShQgXq1auX3O1T2TSV0dHRCon5a8XFxVGRIkUoNDQ030E+KiqKKlWqRFWrVqWMjAxasmQJaWpq0tix
Y1U2uZRIJLRkyRLS0NAgBwcHCgwMpL59+1KLFi3IxMREZTrqHDp0iLp160YdOnSQS0oePHhAY8aMIX19fbk
erarq/
v37dP78+XwnbolEQgEBARQWFkYSiURpa73yzift6OhIzs7OZGpqSh06dBA7ePj5+ZGjoyP9+eef1LdvXzI0
NFTpGrBPiY+Pp4EDB6pEn4CPuXz5MrVv3560tbWpcuXK1L59e3JwcBBHNFFVnFj+JPK2jdy/
fz+VK1dOnAGEiCg0NJQ0NTXFKQCJcq9iV61aRf7+/rR///
4fGu+3kEqllJqaSg0aNJCbwislJYX69+9P1atXl6vFS01NVeqp1v78809au3atGOP27dvFKTafPn1KEomE5
syZQzVq1KDAwEBKTU2lhw8f0rhx48jKykqpax3u379PxsbGFBAQIC6TSqVyB/w//
viDihYtKo7jOHnyZDIyMqLnz5//8HgL0tmzZ6l169bk6upKderUoZCQEKXvyZk3STEwMKBevXpRSEgIWVlZ
UYsWLcTjyoMHD2jcuHEkCAKtWLFCgREXjszMTAoLCyNzc3Ol32dEucMG6enpUWRkJKWlpdHevXtJEARav34
9EeXWJtepU4ecnJzIxcVFqS+yC4oyJpVf01zp+vXr1LFjRzI1NaXw8HCVrqmU4cRSxX04Ft769etp8ODBNG
TIECKS/
7GNHj2aBEGgQYMGUc2aNcnZ2ZmIcnsgy3pOKwvZnMREuQOE5+TkkIeHh9j+Sbbu+fPnVLFiRQoODlZYrN9C
Vg5XV1faunWrWCMpSy779etHaWlp9P79e5o/
fz5VrFiRNDU1ydnZmUqXLi13saCMkpKSyMPDg1q0aJFvRqC8B9e6deuSn5+f+FjVbql+irJPIiD73eStjYu
NjaXy5cuL7SiTkpLIzMyMBEGg2rVri8eQu3fv0qRJk5R6sPr/
Ys2aNTRo0CAyNTVV+tplmfDwcHF2qtu3b5OdnZ3cFKgyjx8/
Vtk7AapMdqxLTU2lrKwscfzTTyWYly5doqCgIJUY0uprcGKpwrp160bh4eFE9O+QC7Vq1SJBEMjT0/
OjQ6DMmTOHvL29qXPnzkpbq5d3+IUNGzZQly5dKCkpierWrUuBgYFERHK3qoKDg8nT01PpT+qy/
fHu3Tvy8fGhKlWq0ObNmz+ZXEqlUnr37h1t2bKFoqOjVeagI2tY7+3tLZdc5j2o1q9fnzp06PDRdarsY4P2
K4u8s84sXbqUzp07R0S5Y4jKLkTv379PNjY21KtXL3GOaT8/P5Vrv/a1bty4QfXr16eWLVsq/
RicMlKplHx9fWn06NGUkZFBpUuXpqCgIPH7Nm/ePLHmkv14sv2we/
duat68Obm7u1Pz5s1p165dn32eKo0a8SWcWKqwnTt3igd8Wfu87OxsateuHZmbm9OKFSvE5DFvcpl39ghlG
wT9ypUrFB4eThKJhJ49e0Y2NjbivN779u0jQRDyzbHcpk0bCgwMVNo2UXnJTszv3r0jT09Pcnd3p82bN+e7
Ld6/
f3+l7s3+JXmTy7zjykkkEnrw4AE1adIk31BXrPB8bNYZWac9otxaS6lUSn5+ftSxY0eSSqWUlpZG7u7uJAg
CNW7cWFGhF7qnT5+q3MgEq1evptq1a5OxsTH17dtXrhKhR48e1L9//58qUVE1u3btIh0dHZo+fTpt3ryZAg
MDSRAEpZ2VqqBxYqmCPjwRR0ZGUps2bcTbONnZ2eTr60uurq4UFRUljlv2YY2esp3QZWNvLly4kA4fPkwTJ
06kPn36yA2PsXDhQhIEgdq3b09Dhw6l3r17k56enkoMEC77vGW3fdPT08nT0/
OjNZdaWloUGBiYb5xHVfKpmsuQkBBycXFRmRrYn8WXZp15/fo1ubi4iNMcZmRkUM+ePenvv/
9W2tEVfnayY8bDhw/
pxo0b4uOLFy9S3bp1ycnJSZz4Ii0tjcaMGUPm5uZKPwrBz0h28Zaenk7NmzenGTNmEBFRcnIylS1b9qNNFX
5WnFj+BBYvXkyVKlWioKAgueSySZMmVLlyZdq0aZPSX71eu3aNdHV1afz48URENHbsWHHcNdlYbDKHDx+mF
i1aUMOGDalVq1ZKPeXfh86ePUtt2rQRhweRJZcf1lxGRUVR8eLFxZpoVZU3ubx48SJNnz6d9PT0lH5GnZ/
N52adefjwISUkJFB6ejpVqVKF/
Pz8KCkpiYYPH07ly5fPN84q+7G2bNlCFhYWZGFhQRUqVKAjR44QUe7QazVr1iQbGxuqXbs2NWzYkMzMzFSm
nejPYNasWXLt+6VSKb1+/ZpsbGzo2LFjlJKSIjZVkFm1atVPn/
hzYqlC8nZo+dDy5cvJzc2NevToIZdcNmvWjMzNzenQoUM/
MtRvcuXKFTI2NiZHR0dxWUpKCs2YMYPU1NTEMb6I/
q11lSXKytpO9FPWrl1Lrq6u1KlTJ7HDVN6ay61bt4pl+jChVlUJCQnUrFkzMjExIU1NTZUfSkMVfWnWmbJl
y1Ljxo1p27ZtZGtrS6VLlyYLCwtOUhREdpy/
du0a2djY0MyZM+nIkSPk7e1NZcqUEWd6unLlCq1atYr69etHS5cupdu3bysy7F/K+/
fvaerUqaSnp0djx44Vl+fk5FDnzp1p0qRJZGlpSb179xbPWykpKdSlSxdas2aN0t0xLEicWKqo3bt3044dO
+jw4cPissjISDG5lA0vkZWVRcOGDVPaji2xsbFUpEgRql+/Ppmbm9PAgQPFda9evRJrLlevXk1Eucm17E/
2WFl9KrYNGzZQ7dq1qV27dmIHivT0dPL29iZbW1vauXPnZ5+vim7cuEEtWrSgq1evKjqUX9LXzDrj6OhIwc
HB9PTpUzpx4gTXVP5AH+utf/r0aVq1ahWNGDFCblt/
f38xuVT16RlV3cuXL2nevHlUrFgxGjNmjLh81KhRJAgCNWnShN69eye33N7eXm6Wrp8RJ5YqYODAgXJz8gY
HB5OJiQmVKlWKnJ2dadCgQeK6yMhIqlKlCgUFBdHZs2flXkfZksuYmBjS1NSk8PBwysnJoaVLl5KxsbFccv
n69WsKCwsjQRDEAd6V2cdqlOPj4/PVJKxbt47q1KlDbdu2FS8C0tLSyM/
P76dtz6ZsHcV+NZ+bdSYzM5MaNWpEXbp0UXCUv54Pe+vLjtuyjlM+Pj75fjv+/
v5ka2srNx4u+3HyVm68ffuWZs+eTcWKFaPQ0FBxmw4dOpCJiQkNGDCAxo8fT926dSNDQ8NfYkxRDTCl9urV
K2hoaGDv3r0oVqwYunbtinPnzuHAgQPQ1NTEvn37sGjRIqSnp2PZsmXo2bMn1NTUEB4eDmtra1StWhVEBEE
QoK6urujiyHn37h369u2L8ePHAwDatm0LABgzZgwAYN68eTA0NMTw4cOhrq6Ozp07Q0NDQ9xO2UilUqipqS
E5ORknTpyARCKBtrY2Fi9eDDs7O4wcORI2NjYAgA4dOiAnJwfBwcFQU1PD4MGDUa1aNWzfvl3BpSg8mpqai
g7hl9awYUPcuXMHKSkpKFu2LIyNjcV1GhoaMDQ0hKWlJYgIACAIgqJC/
WXIjhlXrlxB69atUaFCBZibmwMAYmJi0LRpU5w5cwbHjx9HvXr1xGP4li1b0LhxY8yYMQMtWrSAjo6OIovx
S5CdR4F/
fxsxMTEoVaoUunXrBkEQEBERAalUimnTpmHdunUICwvDjRs3cPbsWVSuXBknT55EhQoVFFmMH0PBiS37Csn
JyRQeHk7Ozs7UunVr6tatm1j7+Pr1a1q8eDHZ2tpSz549xefs2rVL6WooPyfvgLIfq7l8+fIlTZ48WWnHmp
PVOsTFxZGNjQ05OTmRpqYmVa1alVxcXMjb25sGDx6crzaydu3aZGJiQj179qT379//VLe/
mWpQtVlnfjZf6q1fq1YtsrKyouPHj+e7I8IjK/w4jx49IqJ/
+zokJiaSqamp2BHx5cuXNGfOHDIyMqKRI0eKz8vMzKSsrCyVOh9/
L04slVjeg0hycjKNHz+erK2tqWbNmnLbvX79mpYsWUL29vbUqlUruXWq+GXOm1x+2ONOGeVNKosUKUIjR46
k5ORk2rlzJzVp0oTq1q1L/fr1I1dXVxo8eLDYvub9+/fUq1cvmjx5Mp8gmEKo4qwzP5PP9da/
c+eOONyYj48PWVpa0smTJ1VivN6fzebNm8na2loczYOI6MWLF+Tg4CA3xW7e5DIsLEwRoSoFTiyVVN6Dh2y
g7CdPntD48eOpWLFi+b60qampNGvWLGrTps1PceBJTU2lyMhIEgSBQkJCFB3OF31sjmyi3KGgjIyM6OHDh7
Rw4UJyd3entm3b0qpVqygkJIScnJxUfo5spppUcdaZn82XeuuXKVOGWrduTUS5yaWhoaFccsN+jH379lHz5
s3Jw8NDbAObmJhI5cqVy3f8lnXoEQSBJk2apIhwFY7bWCohWbsbAJg4cSIuXryIyZMnw8nJCX379gUAREVF
QV1dHeHh4QAAAwMD9O7dG0OGDIEgCHKvoYoMDAwQEBAATU1N1KhRQ9HhfJFEIoG1tTUyMzNx4sQJ1K5dGwB
ga2sLAHj79i369euHokWLYsuWLRg9ejSMjY2xZs0alChRQpGhs1+Uvb09oqKioK2tDUNDQ0WH80t69+4dnj
17hsuXL+PmzZvYtm0bVq1aBWdnZ0ycOBF6enqYMGECJk2ahL1798LLy4uPFwrQuHFjaGtrY+7cuejTpw8WL
16MUqVK4c2bN5BIJHLbGhkZoUuXLtDU1ESDBg0UFLFiCUT/
a6nNlE5ISAjWrFmDadOmwdPTE6VLlwYAPH78GEuXLsXGjRvRoUMHjBs3Tu55lKeRsapTpbLcunULgwYNglQ
qxZw5c2BhYQEbGxsEBgZi+vTp4napqalIS0uDjo4OnyQY+8UdPnwY3t7eKF26NF6+fImZM2fC09MTdnZ2yM
7ORrNmzVCiRAmsX79e0aH+kvKeg6KjozF37lw8fPgQffr0wbp16+Dt7Q1ra2tIpVJkZ2cjMzMTFSpUUIkKk
cLCiaWSOnDgALp164Zt27ahWrVqICK8evUK9+7dQ7ly5SAIAmbNmoU5c+bg//7v/
9C9e3dFh8yQm1wOHjwY7969w+XLl9G1a1fMnj0bAJCTkwMNDb5JwBiT9+DBg4/21pdKpWjXrh3s7e0REREB
ACp9J+pncPDgQSxevBjHjh3Dixcv0KJFCyQkJEAQBGhpaUEikWDTpk1wcHBQdKgKw2c5JfXq1SuYm5ujatW
quHjxInbu3In169fjzZs3aNiwIebPn48ePXqgTJky6Nq1q6LDZf9Trlw58XaJgYEBWrZsKa5TtuGeGGPKwc
LCAhYWFnLLsrKyMHHiRJw8eRKTJ0/
mhPIHk9VUXrx4EU+fPoVUKoWvry+8vLwgCAJ0dHRw+fJlREREwMXFRXxeeno6ihYtqsDIFY9rLJXA+/
fvoaurK7csNjYWbm5u8PHxQUxMDJo1a4YGDRpAW1sb/
fr1w+7du+Wq2iUSCScuSuT27dsYOHAgiAhjx45FrVq1FB0SY0xFrF27FjExMYiKisLevXtRuXJlRYf0S9q6
dSu6deuGUqVK4dGjR/
D398fq1asBAIcOHcK8efPw5MkTTJs2TWxPqUrNtwoLJ5YKtmbNGiQmJiI0NBTa2togIkilUqirq+PkyZPYs
mULqlevjoYNG6JkyZJIT09HvXr1MGPGDDRs2FDR4bPPuHXrFoYOHYrnz59j9uzZqF69uqJDYowpuZs3b6JP
nz4wMjLC5MmT4ejoqOiQfimyxPDdu3do0qQJevbsidq1ayM+Ph5dunRBnTp1xIksjh49igkTJkAikeCff/
7hger/hxNLBfrjjz/Qp08f7NmzBz4+PnIzXpw/fx4mJiawtLQEAGRnZyMjIwNt2rRBamoqjh8/
zjWUKuDGjRsYO3YsZs2aJe5Lxhj7nJSUFO6tr0AHDhzAmjVroK6ujunTp8PExAQAcPLkSfj5+aF27drYtm0
bBEHA8ePHYW1tjTJlyig4auXBiaWCrFmzBj169MCOHTvQtGlTuaRy27ZtCAoKwtatW1GvXj1kZ2dj4cKF2L
JlC7KysnDy5Eloamqq/
JBCv4qsrCxoaWkpOgzGGGNfISoqCoGBgTAwMMD169dRvHhxsSbz5MmTCAgIgKOjIw4ePPjL3/
b+GM5KFGDlypXo2rUr6tevj6ZNmwLI7f0nCAJ27NiB1q1bY9KkSahXrx6A3GTTxcUFXl5eOHXqFDQ1NZGTk
8NJpYrgpJIxxpSLVCr95OOWLVti3bp1SE9PR1hYGIB/5wevVasW1q9fj/
v37yM5OfnHBaxCuMbyB4uMjESfPn3QvXt37NmzB61bt8bcuXMB5Lbt2LJlC169eoWgoKBPvgZ31GGMMca+z
40bN7BmzRoEBQXB0tJSrvYxOzsb27dvR7du3dCzZ0/
MmzdP7rkf63TLcnFi+QPNmTMHQ4cOxd9//40mTZpg6dKlCAsLQ4cOHcTkkjHGGGOFKzs7G7Vq1cL58+dhZ2
eH3377DVWrVkVAQIC4TUZGBnbu3Ilu3bqhT58+4pjE7PN4HMsfqHLlyli/
fj2aNGkCAGjXrh0EQcCYMWMAQEwuuUaSMcYYKzyampoICAhA+/
bt4ezsjJMnT6J3797YtWsXatSogT59+kBHRwdt27YFALRv3x5aWlpys6ixj+MaSwXIO87VmzdvsHHjRowZM
0au5pKTS8YYY6zwREdH47fffsOhQ4fg7u6Ox48f448//
sCMGTNQsWJF9OjRAw0aNICdnR22b98OR0fHX3pGna/FiaUSkCWXYWFh6NixI1e3M8YYYz/
AiBEj8PjxYyxbtgw6Ojpo164d4uLiUK1aNSQlJeH06dOYOXMmBg0axD3AvxLfClcCBgYG4m3x3r17w8rKCo
MHD1Z0WIwxxthPrVq1avj999+hpaWFnj17Ijo6GocOHUKFChVw8+ZN7Nu3D56enpxUfgOusVQir1+/
xtGjR9GsWTO+Dc4YY4z9APXq1cOJEydQqlQp7NmzR27ub/
btOLFUUjk5OdDQ4AplxhhjrDDI+jvs2bMHQ4YMwfTp0+Hn58fzfX8nHmFbSXFSyRhjjBUeWfJYpUoVSKVSX
LhwQW45+284sWSMMcbYL8vU1BTjx4/H7Nmzce7cOUWHo/
I4sWSMMcbYL61Bgwbw8PCAubm5okNRedzGkjHGGGO/
vIyMDOjo6Cg6DJXHiSVjjDHGGCsQfCucMcYYY4wVCE4sGWOMMcZYgeDEkjHGGGOMFQhOLBljjDHGWIHgxJI
xxhhjjBUITiwZY4wxxliB4MSSMcaUlJWVFebMmSM+FgQBO3bs+OFxhIeHw9XV9ZPro6OjIQgCXr9+/
dWvWb9+fQQHB39XXCtXrkSxYsW+6zUYYwWLE0vGGFMRjx8/
RpMmTb5q2y8lg4wxVhg0FB0AY4z9zLKysqClpVUgr1WqVKkCeR3GGCssXGPJGGNfqX79+hgwYAAGDBgAQ0N
DGBsbY+zYscg7gZmVlRUmTpyILl26wMDAAEFBQQCAEydOoE6dOtDV1YWFhQUGDRqE9PR08XkpKSlo3rw5dH
V1YW1tjXXr1uV7/w9vhT98+BDt27dH8eLFUbRoUbi7u+Ps2bNYuXIlIiIiEBcXB0EQIAgCVq5cCQB4/
fo1evbsiZIlS8LAwAANGzZEXFyc3PtMmzYNpqam0NfXR48ePZCRkfFNn9OLFy/
Qvn17lC5dGkWKFEHFihWxYcOGfNvl5OR89rPMzMzE8OHDUbp0aRQtWhTVqlVDdHT0N8XCGPuxOLFkjLFvsG
rVKmhoaODcuXOYO3cufv/9dyxbtkxum//7v/+Di4sLLl26hLFjxyIxMRE+Pj7w9/
fH5cuXERUVhRMnTmDAgAHic7p164YHDx7gyJEj2LJlCxYtWoSUlJRPxpGWloZ69eohOTkZu3btQlxcHEaOH
AmpVIq2bdti2LBhqFChAh4/
fozHjx+jbdu2AICAgACkpKRg7969uHDhAtzc3ODp6YmXL18CADZt2oTw8HBMmTIF58+fh5mZGRYtWvRNn1F
GRgaqVKmCv//+G1evXkVQUBA6d+6Mc+fOfdNnOWDAAJw+fRobN27E5cuXERAQAB8fH9y6deub4mGM/
UDEGGPsq9SrV48cHR1JKpWKy0JCQsjR0VF8XLZsWfLz85N7Xo8ePSgoKEhu2fHjx0lNTY3ev39PN2/
eJAB07tw5cX18fDwBoNmzZ4vLAND27duJiGjp0qWkr69PL168+Gis48ePJxcXl3zvaWBgQBkZGXLLbW1tae
nSpUREVKNGDerXr5/
c+mrVquV7rbyOHDlCAOjVq1ef3MbX15eGDRsmPv7SZ3nv3j1SV1en5ORkudfx9PSk0NBQIiJasWIFGRoafv
I9GWM/
HrexZIyxb1C9enUIgiA+rlGjBmbNmgWJRAJ1dXUAgLu7u9xz4uLicPnyZbnb20QEqVSKpKQkJCQkQENDA1W
qVBHXOzg4fLbHc2xsLCpXrozixYt/dexxcXFIS0tDiRIl5Ja/f/
8eiYmJAID4+Hj06dNHbn2NGjVw5MiRr34fiUSCKVOmYNOmTUhOTkZWVhYyMzNRpEgRue0+91leuXIFEokE5
cuXl3tOZmZmvvgZY8qDE0vGGCtgRYsWlXuclpaG3r17Y9CgQfm2tbS0REJCwje/h66u7jc/
Jy0tDWZmZh9tp1iQw/
bMnDkTc+fOxZw5c1CxYkUULVoUwcHByMrK+qZY1dXVceHCBTFhl9HT0yuwWBljBYsTS8YY+wZnz56Ve3zmz
BmUK1cuX/KTl5ubG65fvw47O7uPrndwcEBOTg4uXLgADw8PAMDNmzc/
Oy5kpUqVsGzZMrx8+fKjtZZaWlqQSCT54njy5Ak0NDRgZWX10dd1dHTE2bNn0aVLF7kyfouTJ0/
it99+Q6dOnQAAUqkUCQkJcHJyktvuc59l5cqVIZFIkJKSgjp16nzT+zPGFIc77zDG2De4f/
8+hg4dips3b2LDhg2YP38+Bg8e/NnnhISE4NSpUxgwYABiY2Nx69Yt7Ny5U+y8Y29vDx8fH/
Tu3Rtnz57FhQsX0LNnz8/
WSrZv3x6lSpWCn58fTp48iTt37mDr1q04ffo0gNze6UlJSYiNjcXz58+RmZkJLy8v1KhRA35+fti/fz/
u3r2LU6dOYcyYMTh//
jwAYPDgwfjzzz+xYsUKJCQkYPz48bh27do3fUblypXDgQMHcOrUKcTHx6N37954+vTpN32W5cuXR8eOHdGl
Sxds27YNSUlJOHfuHKZOnYq///77m+JhjP04nFgyxtg36NKlC96/f4+qVauif//
+GDx4sDik0KdUqlQJR48eRUJCAurUqYPKlStj3LhxMDc3F7dZsWIFzM3NUa9ePbRq1QpBQUEwMTH55GtqaW
lh//79MDExQdOmTVGxYkVMmzZNrDn19/
eHj48PGjRogJIlS2LDhg0QBAF79uxB3bp1ERgYiPLly6Ndu3a4d+8eTE1NAQBt27bF2LFjMXLkSFSpUgX37
t1D3759v+kzCgsLg5ubG7y9vVG/fn0xAf7Wz3LFihXo0qULhg0bBnt7e/
j5+SEmJgaWlpbfFA9j7McRiPIMGsYYY+yT6tevD1dXV7lpFhljjP2LaywZY4wxxliB4MSSMcYYY4wVCL4Vz
hhjjDHGCgTXWDLGGGOMsQLBiSVjjDHGGCsQnFgyxhhjjLECwYklY4wxxhgrEJxYMsYYY4yxAsGJJWOMMcYY
KxCcWDLGGGOMsQLBiSVjjDHGGCsQ/w8PVBS4AF4GjAAAAABJRU5ErkJggg==",
"text/plain": [
"<Figure size 1000x700 with 1 Axes>"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"from torchmetrics import ConfusionMatrix\n",
"from mlxtend.plotting import plot_confusion_matrix\n",
"\n",
"# 2. Setup confusion matrix instance and compare predictions to targets\n",
"confmat = ConfusionMatrix(num_classes=len(class_names), task='multiclass')\n",
"confmat_tensor = confmat(preds=y_pred_tensor,\n",
" target=test_data.targets)\n",
"\n",
"# 3. Plot the confusion matrix\n",
 "fig, ax = plot_confusion_matrix(\n",
" conf_mat=confmat_tensor.numpy(), # matplotlib likes working with NumPy \
n",
" class_names=class_names, # turn the row and column labels into class
names\n",
" figsize=(10, 7)\n",
");"
]
},
{
"cell_type": "markdown",
"id": "381c1c93-df30-451c-b65e-5d4c1680dc30",
"metadata": {
"id": "381c1c93-df30-451c-b65e-5d4c1680dc30"
},
"source": [
"Woah! Doesn't that look good?\n",
"\n",
"We can see our model does fairly well since most of the dark squares are down
the diagonal from top left to bottom right (and ideal model will have only values
in these squares and 0 everywhere else).\n",
"\n",
"The model gets most \"confused\" on classes that are similar, for example
predicting \"Pullover\" for images that are actually labelled \"Shirt\".\n",
"\n",
"And the same for predicting \"Shirt\" for classes that are actually
labelled \"T-shirt/top\".\n",
"\n",
"This kind of information is often more helpful than a single accuracy metric
because it tells use *where* a model is getting things wrong.\n",
"\n",
"It also hints at *why* the model may be getting certain things wrong.\n",
"\n",
"It's understandable the model sometimes predicts \"Shirt\" for images labelled
\"T-shirt/top\".\n",
"\n",
"We can use this kind of information to further inspect our models and data to
see how it could be improved.\n",
"\n",
"> **Exercise:** Use the trained `model_2` to make predictions on the test
FashionMNIST dataset. Then plot some predictions where the model was wrong
alongside what the label of the image should've been. After visualizing these
predictions do you think it's more of a modelling error or a data error? As in,
could the model do better or are the labels of the data too close to each other
(e.g. a \"Shirt\" label is too close to \"T-shirt/top\")?"
]
},
{
"cell_type": "markdown",
"id": "25818e83-89de-496d-8b56-af4fc9f2acc5",
"metadata": {
"id": "25818e83-89de-496d-8b56-af4fc9f2acc5"
},
"source": [
"## 11. Save and load best performing model\n",
"\n",
"Let's finish this section off by saving and loading in our best performing
model.\n",
"\n",
 "Recall from [notebook 01](https://www.learnpytorch.io/01_pytorch_workflow/#5-
saving-and-loading-a-pytorch-model) we can save and load a PyTorch model using a
combination of:\n",
"* `torch.save` - a function to save a whole PyTorch model or a model's
`state_dict()`. \n",
"* `torch.load` - a function to load in a saved PyTorch object.\n",
"* `torch.nn.Module.load_state_dict()` - a function to load a saved
`state_dict()` into an existing model instance.\n",
"\n",
"You can see more of these three in the [PyTorch saving and loading models
documentation](https://pytorch.org/tutorials/beginner/saving_loading_models.html).\
n",
"\n",
"For now, let's save our `model_2`'s `state_dict()` then load it back in and
evaluate it to make sure the save and load went correctly. "
]
},
{
"cell_type": "code",
"execution_count": 57,
"id": "d058e8fa-560f-4350-a154-49593ff403c9",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "d058e8fa-560f-4350-a154-49593ff403c9",
"outputId": "0156a518-dae2-4b25-999a-c0a77ef7ef7c"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Saving model to: models/03_pytorch_computer_vision_model_2.pth\n"
]
}
],
"source": [
"from pathlib import Path\n",
"\n",
"# Create models directory (if it doesn't already exist), see:
https://docs.python.org/3/library/pathlib.html#pathlib.Path.mkdir\n",
"MODEL_PATH = Path(\"models\")\n",
"MODEL_PATH.mkdir(parents=True, # create parent directories if needed\n",
" exist_ok=True # if models directory already exists, don't
error\n",
")\n",
"\n",
"# Create model save path\n",
"MODEL_NAME = \"03_pytorch_computer_vision_model_2.pth\"\n",
"MODEL_SAVE_PATH = MODEL_PATH / MODEL_NAME\n",
"\n",
"# Save the model state dict\n",
"print(f\"Saving model to: {MODEL_SAVE_PATH}\")\n",
"torch.save(obj=model_2.state_dict(), # only saving the state_dict() only saves
the learned parameters\n",
" f=MODEL_SAVE_PATH)"
]
},
 {
"cell_type": "markdown",
"id": "a1542284-8132-42ba-b00d-57e9b9037e4e",
"metadata": {
"id": "a1542284-8132-42ba-b00d-57e9b9037e4e"
},
"source": [
"Now we've got a saved model `state_dict()` we can load it back in using a
combination of `load_state_dict()` and `torch.load()`.\n",
"\n",
"Since we're using `load_state_dict()`, we'll need to create a new instance of
`FashionMNISTModelV2()` with the same input parameters as our saved model
`state_dict()`."
]
},
{
"cell_type": "code",
"execution_count": 58,
"id": "634a8f7a-3013-4b45-b365-49b286d3c478",
"metadata": {
"id": "634a8f7a-3013-4b45-b365-49b286d3c478"
},
"outputs": [],
"source": [
"# Create a new instance of FashionMNISTModelV2 (the same class as our saved
state_dict())\n",
"# Note: loading model will error if the shapes here aren't the same as the
saved version\n",
"loaded_model_2 = FashionMNISTModelV2(input_shape=1, \n",
" hidden_units=10, # try changing this to
128 and seeing what happens \n",
" output_shape=10) \n",
"\n",
"# Load in the saved state_dict()\n",
"loaded_model_2.load_state_dict(torch.load(f=MODEL_SAVE_PATH))\n",
"\n",
"# Send model to GPU\n",
"loaded_model_2 = loaded_model_2.to(device)"
]
},
{
"cell_type": "markdown",
"id": "feeaebf4-6040-4fa5-852d-5eb8d2bbb94c",
"metadata": {
"id": "feeaebf4-6040-4fa5-852d-5eb8d2bbb94c"
},
"source": [
"And now we've got a loaded model we can evaluate it with `eval_model()` to
make sure its parameters work similarly to `model_2` prior to saving. "
]
},
{
"cell_type": "code",
"execution_count": 59,
"id": "3e3bcd06-d99b-47bc-8828-9e3903285599",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
 "id": "3e3bcd06-d99b-47bc-8828-9e3903285599",
"outputId": "c0ee1d5f-9573-4e1a-8430-ee09fb4d72cd"
},
"outputs": [
{
"data": {
"text/plain": [
"{'model_name': 'FashionMNISTModelV2',\n",
" 'model_loss': 0.3285697102546692,\n",
" 'model_acc': 88.37859424920129}"
]
},
"execution_count": 59,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Evaluate loaded model\n",
"torch.manual_seed(42)\n",
"\n",
"loaded_model_2_results = eval_model(\n",
" model=loaded_model_2,\n",
" data_loader=test_dataloader,\n",
" loss_fn=loss_fn, \n",
" accuracy_fn=accuracy_fn\n",
")\n",
"\n",
"loaded_model_2_results"
]
},
{
"cell_type": "markdown",
"id": "c2b37855-c0da-4834-a2d4-a0faa8410b65",
"metadata": {
"id": "c2b37855-c0da-4834-a2d4-a0faa8410b65"
},
"source": [
"Do these results look the same as `model_2_results`?"
]
},
{
"cell_type": "code",
"execution_count": 60,
"id": "68544254-c99a-47ec-a32f-9816c21a993e",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "68544254-c99a-47ec-a32f-9816c21a993e",
"outputId": "74b8d4ca-d35a-4f70-e8b9-ed54f034358e"
},
"outputs": [
{
"data": {
"text/plain": [
"{'model_name': 'FashionMNISTModelV2',\n",
" 'model_loss': 0.3285697102546692,\n",
" 'model_acc': 88.37859424920129}"
 ]
},
"execution_count": 60,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"model_2_results"
]
},
{
"cell_type": "markdown",
"id": "0ee07f93-4344-4c7a-8b1d-92a56034e7b2",
"metadata": {
"id": "0ee07f93-4344-4c7a-8b1d-92a56034e7b2"
},
"source": [
"We can find out if two tensors are close to each other using `torch.isclose()`
and passing in a tolerance level of closeness via the parameters `atol` (absolute
tolerance) and `rtol` (relative tolerance).\n",
"\n",
"If our model's results are close, the output of `torch.isclose()` should be
true."
]
},
{
"cell_type": "code",
"execution_count": 61,
"id": "48dcf0ba-7e00-4406-8aaa-41918856361a",
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "48dcf0ba-7e00-4406-8aaa-41918856361a",
"outputId": "47324300-0d00-46de-d130-1283ad044ef8"
},
"outputs": [
{
"data": {
"text/plain": [
"tensor(True)"
]
},
"execution_count": 61,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Check to see if results are close to each other (if they are very far away,
there may be an error)\n",
"torch.isclose(torch.tensor(model_2_results[\"model_loss\"]), \n",
" torch.tensor(loaded_model_2_results[\"model_loss\"]),\n",
" atol=1e-08, # absolute tolerance\n",
" rtol=0.0001) # relative tolerance"
]
},
{
 "cell_type": "markdown",
"id": "c3969b7d-9955-4b6f-abf8-fe8eedf233a9",
"metadata": {
"id": "c3969b7d-9955-4b6f-abf8-fe8eedf233a9"
},
"source": [
"## Exercises\n",
"\n",
"All of the exercises are focused on practicing the code in the sections
above.\n",
"\n",
"You should be able to complete them by referencing each section or by
following the resource(s) linked.\n",
"\n",
"All exercises should be completed using [device-agnostic
code](https://pytorch.org/docs/stable/notes/cuda.html#device-agnostic-code).\n",
"\n",
"**Resources:**\n",
"* [Exercise template notebook for 03](https://github.com/mrdbourke/pytorch-
deep-learning/blob/main/extras/exercises/
03_pytorch_computer_vision_exercises.ipynb)\n",
"* [Example solutions notebook for 03](https://github.com/mrdbourke/pytorch-
deep-learning/blob/main/extras/solutions/
03_pytorch_computer_vision_exercise_solutions.ipynb) (try the exercises *before*
looking at this)\n",
"\n",
"1. What are 3 areas in industry where computer vision is currently being
used?\n",
"2. Search \"what is overfitting in machine learning\" and write down a
sentence about what you find. \n",
"3. Search \"ways to prevent overfitting in machine learning\", write down 3 of
the things you find and a sentence about each. **Note:** there are lots of these,
so don't worry too much about all of them, just pick 3 and start with those.\n",
"4. Spend 20-minutes reading and clicking through the [CNN Explainer website]
(https://poloclub.github.io/cnn-explainer/).\n",
" * Upload your own example image using the \"upload\" button and see what
happens in each layer of a CNN as your image passes through it.\n",
"5. Load the
[`torchvision.datasets.MNIST()`](https://pytorch.org/vision/stable/generated/
torchvision.datasets.MNIST.html#torchvision.datasets.MNIST) train and test
datasets.\n",
"6. Visualize at least 5 different samples of the MNIST training dataset.\n",
"7. Turn the MNIST train and test datasets into dataloaders using
`torch.utils.data.DataLoader`, set the `batch_size=32`.\n",
"8. Recreate `model_2` used in this notebook (the same model from the [CNN
Explainer website](https://poloclub.github.io/cnn-explainer/), also known as
TinyVGG) capable of fitting on the MNIST dataset.\n",
"9. Train the model you built in exercise 8. on CPU and GPU and see how long it
takes on each.\n",
"10. Make predictions using your trained model and visualize at least 5 of them
comparing the prediction to the target label.\n",
"11. Plot a confusion matrix comparing your model's predictions to the truth
labels.\n",
"12. Create a random tensor of shape `[1, 3, 64, 64]` and pass it through a
`nn.Conv2d()` layer with various hyperparameter settings (these can be any settings
you choose), what do you notice if the `kernel_size` parameter goes up and down?\
n",
"13. Use a model similar to the trained `model_2` from this notebook to make
predictions on the test
[`torchvision.datasets.FashionMNIST`](https://pytorch.org/vision/main/generated/
torchvision.datasets.FashionMNIST.html) dataset. \n",
" * Then plot some predictions where the model was wrong alongside what the
label of the image should've been. \n",
" * After visualizing these predictions do you think it's more of a
modelling error or a data error? \n",
" * As in, could the model do better or are the labels of the data too close
to each other (e.g. a \"Shirt\" label is too close to \"T-shirt/top\")?\n",
"\n",
"## Extra-curriculum\n",
"* **Watch:** [MIT's Introduction to Deep Computer
Vision](https://www.youtube.com/watch?v=iaSUYvmCekI&list=PLtBw6njQRU-
rwp5__7C0oIVt26ZgjG9NI&index=3) lecture. This will give you a great intuition
behind convolutional neural networks.\n",
"* Spend 10-minutes clicking through the different options of the [PyTorch
vision library](https://pytorch.org/vision/stable/index.html), what different
modules are available?\n",
"* Lookup \"most common convolutional neural networks\", what architectures do
you find? Are any of them contained within the
[`torchvision.models`](https://pytorch.org/vision/stable/models.html) library? What
do you think you could do with these?\n",
"* For a large number of pretrained PyTorch computer vision models as well as
many different extensions to PyTorch's computer vision functionalities check out
the [PyTorch Image Models library `timm`](https://github.com/rwightman/pytorch-
image-models/) (Torch Image Models) by Ross Wightman."
]
}
],
"metadata": {
"accelerator": "GPU",
"colab": {
"gpuType": "A100",
"machine_shape": "hm",
"provenance": []
},
"gpuClass": "standard",
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.8.13"
},
"vscode": {
"interpreter": {
"hash": "3fbe1355223f7b2ffc113ba3ade6a2b520cadace5d5ec3e828c83ce02eb221bf"
}
}
},
 "nbformat": 4,
"nbformat_minor": 5
}