dfdx
dfdx copied to clipboard
Added draft BatchNorm1D implementation
This is a WIP draft PR for batch normalization, wanted to discuss some potential design choices and issues.
Since the behavior of batchnorm is different for training and inference we need some way to test for both.
One possibility is to add a test that calls forward_mut multiple times and then a final forward call that uses the running statistics.
Also, I have been pretty busy recently, so I would probably wait for https://github.com/coreylowman/dfdx/issues/163 before taking this on again.