WinML Dashboard (preview)

WinML Dashboard is a tool for viewing, editing, converting, and validating machine learning models for Windows ML inference engine. The engine is built into Windows 10 and evaluates trained models locally on Windows devices using hardware optimizations for CPU and GPU to enable high performance inferences. (Download Preview Version of the tool)

Today there are several different frameworks available for training and evaluating machine learning models, which makes it difficult for app developers to integate models into their product. Windows ML uses ONNX machine learning model format that allows conversion from one framework format to another, and this Dashboard makes it easy to convert models from different frameworks to ONNX.

This tool supports converting to ONNX from the following source frameworks:

Apple Core ML
TensorFlow (subset of models convertible to ONNX)
Keras
Scikit-learn (subset of models convertible to ONNX)
Xgboost
LibSVM

The tool also allows validation of the converted model by evaluating the model with built-in Windows ML inference engine using synthetic data (default) or real input data on CPU or GPU.

Viewing and Editing Models

The Dashboard uses Netron for viewing machine learning models. Although WinML uses ONNX format, the Netron viewer supports viewing several different framework formats.

Many times a developer may need to update certain model metadata or modify model input and output nodes. This tool supports modifying model properties, metadata and input/output nodes of an ONNX model.

Selecting Edit tab (center top as shown in the snip below) takes you to viewing and editing panel. The left pane in the panel allows editing model input and output nodes, and the right pane allows editing Model properties. The center portion shows the graph. At this time, editing support to limited to model input/output node (and not inner nodes), model properties and model metadata.

The Edit/View button switches from Edit mode to View-only mode and vice versa. View-only mode doesn't allow editing and enables Netron viewer's native features such as the ability to see detailed information for each node.

Converting Models

The Convert tab (see snip below) helps convert models from several different frameworks (as listed above) to ONNX format by WinMLTools. Also you can select the ONNX version to convert to and apply Quantization.

In order to do the conversion, the tool installs a separate Python environment and a set of converter tools. This helps alleviate one of the big pain points of a typical developer - installing the right Python environment and tool chain for conversion and validating the model.

WinMLTools provides quantization feature to reduce the memory footprint of the model.

For quantization:

Type	Usage
Type1	Do not use dequantize_linear option, the converted model can only be run in Win10 released before Version 1809 (OS build 17763)
Type2	Use dequantize_linear option, the converted model can also be run in Win10 release after Version 1809 (OS build 17763)

For ONNX Version:

Version	For Windows 10 Build
1.2	1809 (OS build 17763)
1.3	after 1809 (OS build 17763)

Validating Models

Once you have an ONNX model, you can validate whether the conversion has happened successfully and that the model can be evaluated in Windows ML inference engine. This is done using the Run tab (see snip below).

You can choose various options such as CPU (default) vs GPU, real input vs synthetic input (default) etc. The result of model evaluation appears in the console window at the bottom.

Note that model validation feature is only available on Windows 10 October 2018 Update or newer version of Windows 10 as the tool relies on built-in Windows ML inference engine.

Debugging Inference

You can utilize the debug feature of WinML Dashboard to gain insight into how raw data is flowing through operators in your model. You can also choose to visualize this data for computer vision inference.

To debug your model follow these steps:

Navigate to the Edit tab and click on the operator for which you wish to capture intermediate data. On the left side panel there will be a Debug menu where you can select the formats of intermediate data you wish to capture. The options are currently text and png. Text will output a text file containing the dimensions, data type and raw tensor data produced by this operator. Png will format this data into an image file which can be useful for computer vision applications.
Navigate to the Run tab and select the model you wish to debug.
For the Capture field, select Debug from the dropdown.
Select an input image or csv to supply to your model at execution. Note that this is required when capturing Debug data.
Select an output folder to export debug data.
Click Run. Once execution is complete you can navigate to this selected folder to view your Debug capture.

Intermediate PNG convolutional output from an image of a kitten

Install

You can install an early preview version of WinML Dashboard from here.

Download Preview Version

Note that the prerelease version is not currently signed so you may see a warning about the binary being untrusted when you try to download and run the installer.

Alternatively, you can build the app from source following the instructions below.

Build from source

Prerequisites

Requirements	Version	Download	Command to check
Python3	3.4+	here	`python --version`
Yarn	latest	here	`yarn --version`
Node.js	latest	here	`node --version`
Git	latest	here	`git --version`
MSBuild	latest	here	`msbuild -version`
Nuget	latest	here	`nuget help`

All six prerequisites should be added to Environment Path. Note that MSBuild and Nuget will be included in a Visual Studio 2017 installation.

Steps to build and run

git clone https://github.com/Microsoft/Windows-Machine-Learning
cd Tools/WinMLDashboard
Run Git submodule update --init --recursive to update Netron.
Run yarn lock-netron-deps to execute necessary modifications to Netron dependency versions.
Run yarn to download dependencies.
Then, run yarn electron-prod to build and start the desktop application, which will launch the Dashboard.

All available commands can be seen at package.json.

Debugging

To open the debug view in the Electron app

Run it with flag --dev-tools
Or select View -> Toggle Dev Tools in the application menu
Or press Ctrl + Shift + I.

Getting Models

ONNX Model Zoo

Feedback

For issues, file a bug on GitHub Issues.
For feature requests, file an issue on GitHub Issues
Ask questions on Stack Overflow.

License

This tool is under the MIT license.