[Python] Different result probabilities between VGG16 in MATLAB And in Python when save in...

I re-trained the last layer of a VGG16 using MATLAB and saved in ONNX format. I asked chatGPT to create a small PyGUI with which I can select a test image, the onnx network and do the inference. It is all functional but the classes prediction probabilities are not the same as when I use MATLAB and the function classify.

Most python frameworks do some pre-processing (subtracting the mean and divide by the std before presenting the test image to the network for inference, so why we do not need to do such image pre-processing with MATLAB ?

When exporting the trained network in MATLAB using exportONNXNetwork does the version of ONNX operator set important ? Do I need to use the highest possible ?

Thank you for your time and help

Brahim

Continue reading...