Validation samples

[DeveloperNo579212]

Found this example and tried a model exported from teachablemachine to SavedModel and then tf2onnx.
With 10 images of two class each (left hand 10 pics, right hand 10 pics) I get 100% hitrate on website prior to export
using the training sample as prediction input.
Using tf2onnx 1.9.3 when converting to onnx I get a warning (Runtime warning) but it looks good else. (Tried 1.10 latest also)
Using the slight modified source code I do not get 100% hitrate with same input for prediction, about 50/60 % and unreliable.

No clue on how to validate onnx model.

[eNeRGy164]

I'm no expert in ONNX, but one thing I ran into in the past was different tools use different ways to handle RGB vs BGR.

If you swap the Blue and Red channels, does the score improve?
In my code I have to do this:

CvInvoke.CvtColor(screen, screen, ColorConversion.Bgr2Rgb);

[DeveloperNo579212]

Thanks verified my onnx is ok, I thought the alpha channel might be needed but it is only 3 dimensions in the onnx.

[MrBean2016]

There are also different options for resizing found in resizing: Microsoft.ML.Transforms.Image.ImageResizingEstimator.ResizingKind