Also infer dynamic axes for every model input and output

[NeoLegends]

I think this could be a fix for #510 (comment) and https://bitbucket.org/omnifluent/apptek_asr/pull-requests/1173/overview.

[albertz]

I think you should not hardcode ...:size1 here. You can just check the entry in extern_data/model_outputs, it will tell you what dynamic dims are there.

[NeoLegends]

How so? This is not within RETURNN (where this info might be available), and there are cases where there's simply extern_data = {"data": {"dim": 60}} and no more info.

[albertz]

there are cases where there's simply extern_data = {"data": {"dim": 60}}

It applies extern_data_template_from_config_opts. This will call Tensor(**opts). I.e. for your example, Tensor(dim=60). It follows the defaults from Tensor. The shape by default is (None, dim). shape is without the batch dim, for historical reasons. But there is by default always a batch dim, with default batch_dim_axis=0. The None in shape refers to a dynamic dim. As it comes after the batch dim, you get the axis 1 here.

You could replicate the defaults from Tensor here. But to really cover all cases, I think the easiest would be to just from returnn.tensor import Tensor and then Tensor(name=name, **opts).