search

rwth-i6 / pytorch-to-returnn

Make PyTorch code runnable within RETURNN
3 stars 6 forks source link

add RETURNN behavior_version #93

Closed vieting closed 2 years ago

vieting commented 2 years ago

For #92, we need RETURNN's behavior_version to have the correct behavior of the DotLayer. There might be fixes necessary due to this change.

albertz commented 2 years ago

One fix: Change

net_dict={name: {"class": "subnetwork", "from": "data", "subnetwork": {"output": {"class": "copy"}}}}

to

net_dict={name: {"class": "subnetwork", "from": "data", "subnetwork": {"output": {"class": "copy", "from": "data"}}}}
albertz commented 2 years ago

Ah we have to change get_returnn_axis_description. When you remove the if dim_tag.dyn_size is not None: and just always use this code, does this already work?

Edit There is also another heuristic we can use here: We can use the "dim:..." syntax when the static dim is unique.

albertz commented 2 years ago

You also need to adapt BatchNorm1d.import_params_torch_to_returnn for the new param version. Maybe just:

out_shape = [layer.output.dim]
vieting commented 2 years ago

out_shape = [layer.output.dim]

That seems to work. So we don't neet _expand_dims at all anymore, I guess.