subnetwork problem in group norm

rwth-i6 / pytorch-to-returnn

Make PyTorch code runnable within RETURNN

3 stars 6 forks source link

subnetwork problem in group norm #27

Closed vieting closed 3 years ago

vieting commented 3 years ago

I slightly modified the Fp32GroupNorm test case to reproduce the problem that I mentioned earlier in #18. If Fp32GroupNorm is wrapped inside a Sequential, the layer that is input to GroupNorm.import_params_torch_to_returnn will be a SubnetworkLayer. Do you have any ideas on this?

albertz commented 3 years ago

You should not modify an existing test case to test for sth different. Make a new test case.

vieting commented 3 years ago

I thought it would be easier to see the detail that differed but sure, it's better to have a new test case.