Inconsistent output with "correct" classification

[neo-yuan-fit]

Ask a Question

Question

Hi! Thanks for the work. I was following tutorial here to successfully convert the fp32-tflite model into onnx with opset 13.

When i run the model with onnx and tf, the prediction looks correct as shown here.

However, if I use the np.testing.assert_allclose as suggested here, I get a mismatch rate of 71% on the same image input. Furthermore, if I use a randomly generated matrix input with the same shape, the mismatch will increase into 100%.

I would like to ask why this is happening and if there is any quantitative way to determine if the model conversion is successful.

Thank you!