Examples

These examples are designed to illustrate how PyTorch-accelerated can be used in different contexts.

Example	Location	Description	Dataset
Train MNIST (Vision classification)	`core/train_mnist.py`	Demonstrates how to train a custom model using the default Trainer	MNIST (included with TorchVision)
Integrate Metrics using a Callback	`core/train_with_metrics_in_callback.py`	Demonstrates how to integrate `TorchMetrics` metrics using a custom callback with the default Trainer	MNIST (included with TorchVision)
Integrate Metrics using a Custom Trainer	`core/train_with_metrics_in_loop.py`	Demonstrates how to integrate `TorchMetrics` metrics, using a custom Trainer	MNIST (included with TorchVision)
Finetune BERT (NLP Sequence classification)	`nlp/hf_bert_glue_mrpc.py`	Finetunes a BERT model from `Transformers` using a custom Trainer	GLUE MRPC (included with `Datasets`)
Train Faster-RCNN (Vision object detection)	`vision/faster_rcnn/train_cars.py`	Demonstrates how to create a custom Trainer for use with TorchVision's Faster-RCNN model	Cars (object detection)
Finetune a CNN (Vision classifier)	`vision/transfer_learning/pets_finetune.py`	Demonstrates how to use gradual unfreezing when finetuning a model from `timm`	Pets
Finetune a CNN (Vision classifier)	`vision/transfer_learning/pytorch_tutorial_finetune.py`	Demonstrates how to use `TorchVision` models and datasets when finetuning a model	Ants and Bees
Progressive resizing (Vision classifier)	`vision/transfer_learning/progressive_resizing.py`	Demonstrates how to use progressive resizing, to gradually increase image size	Imagenette
Using all components from `timm` (Vision classifier)	`vision/using_timm_components/all_timm_components.py`	Demonstrates how to create a custom `Trainer` to use Datasets, DataLoaders, models, optimizers and schedulers from `timm`	Imagenette
Using some components from `timm` (Vision classifier)	`vision/using_timm_components/train_mixup_ema.py`	Demonstrates how to create a custom `Trainer` to use data augmentation and ModelEMA from `timm` alongside torch DataLoaders	Imagenette

Running the examples

Each example has an associated DockerFile and Dev Container definition, which is located at the root of the folder, these define the execution environment for the example.; some examples share execution environments.

The easiest way to run an example is to open the relevant example in a dev container using Visual Studio Code, as described here.

Datasets Used

Some examples require external datasets to be downloaded; it is expected that all datasets are located in the data folder. Details of the datasets used are included below:

MNIST: A vision dataset included with TorchVision, this will be downloaded when the example script is executed.
GLUE MRPC: A NLP dataset included with datasets, this will be downloaded when the example script is executed.
Ants & Bees: A vision dataset which must be manually downloaded before use, a helper script is located in the data folder.
Imagenette: A vision dataset which must be manually downloaded before use, a helper script is located in the data folder.
Pets: A vision dataset which must be manually downloaded before use, a helper script is located in the data folder.
Cars: An object detection dataset, which can be downloaded from Kaggle here