I use pytorch 2.1 with QAT training in FX mode and save model in the ONNX format. When I check model by the Netron, I find that in flatten the output scale is different from the input scale, which I don’t think is normal, because the flatten should only change the dimensions and not the values, what went wrong here?
could you print the model after prepare_qat_fx and training? and check the input and output fake quant for the flatten op?