Hi, I have a specific case and need some help/pointer.
I have designed a specialized normalization layer (with learnable parms) from nn.Module and like to apply QAT to this one. But, I couldn’t find a doc on how to make a corresponding module for QAT: such as attaching processes weight_fake_quant, activation_post_process as well.
Hi, Jerry, thanks for sharing that. Yes, I also saw the mapping table, but (I should have been more clear on this): For a custom layer, I need to make a corresponding QAT version of that customer layer. Is there any particular requirement for such QAT versions (in doc or example)? such as a particular set of functions or attributes? One of them looks like ‘from_float’.