Hi!
I cannot find how input_scale and zero_point are set to qnn layers that support them (qnn.conv2d or qnn.requantize). The constructors of these layers appear to take them as arguments - so I would assume that there is a forward pass of the model and they are deduced?
For example, the scale of the output of a qnn.conv2d is input_scale * kernel_scale. How is this result loaded in the input_scale attribute of the next layer?
Thank you!