pieeatingninjas / winmlplayground

UWP Project for trying out WinML with different ONNX models

WinMLPlayground

UWP Project for trying out WinML with different ONNX models

Models

• SqueezeNet
• ResNet
• AlexNet (ONNX model not included as is exceeds GitHub's max file size. You can download it from Azure AI Gallery)
• ShuffleNet
• ZfNet (ONNX model not included as is exceeds GitHub's max file size. You can download it from Azure AI Gallery)
• DenseNet

DISCLAIMER

This sample project demonstrates how to do image classification using different ONNX models on WinML. Currently this is still work in progress, there is no guarantee that any of this code (like preprocessing) is 100% as it should be. Over time this will get validated and will be updated in this repository. That being said, all of the models integrated in this sample give correct results for the sample images and some manual tests (although the probability might be lower than expected).

winmlplayground's Issues

Look at FER

Test out and add FER+ Emotion Recognition Model to the sample app

Look at Inception v2

Test out and add Inception V2 Model to the sample app

Postprocessing DenseNet not OK

Adjust postprocessing of DenseNet as the probability doesn't seem to be OK. The classification itself seems to be good, but the probability pct not.

Crop image

When an image exceeds the expected input dimensions, we need to automatically do a center crop on the image.

Refactor ViewModels

Currently a lot of code is being repeated in the ViewModels.
Move as much code as possible to a base ViewModel that when adding new models, as less work as possible should be done.

Look at MNIST

Test out and add MNIST Handwritten Digit Recognition Model to the sample app

Add support for Object Detection models

Currenlty the sample app only shows classification models, hence the 5 most likely classifications are shown.
Models like YOLO can identify multiple objects in a picture so these kinds of object detection models should be handled differently and the results need to visualized as well.

Look at CaffeNet 1.2

Test out and add CaffeNet 1.2 Model to the sample app

Look at Tiny YOLOv2

Test out and add Tiny YOLOv2 Model (object detection) to the sample app

Refactor Views

Currently the View are exact copies of each other. Move the current controls to one UserControl which can be added on each separate view.

Move preprocessing to separate class

Move preprocessing logic to separate helper class, passing in shape, mean, std

Look at Inception v1

Test out and add Inception V1 Model to the sample app