Bachelor-Thesis | Bachelor Thesis - Design and Implementation

This might actually be a bug. Via register_as_extern_data, I'm not exactly sure that the logic of translate_to_common_search_beam is correct. I think the code currently expects that a target never has a beam.

So, to answer your question: I think you are doing it already correct (so we can close this StackOverflow question).

You should open a GitHub issue about this (and then link it here). It would be good to come up with a simple test case:

• I.e. where there is some beam (you don't even need RecLayer for that, just a single ChoiceLayer would be enough I think),
• then register_as_extern_data on that,
• and then some other layer which uses this target in some way (e.g. just with "loss": "ce").

Probably this will cause exactly your problem.

It does not only validate the batch length. It will collapse the batch and time (it has used flatten_with_seq_len_mask, see code of Loss.init and that function) and then calculate the loss on that flattened tensor. So also the seq length need to match. This might be a problem but I'm not sure. As you have the same target also for the rec layer itself, it should have the same seq length in training.

You can debug this by carefully checking the output of debug_print_layer_output_template, i.e. check the Data (batch-shape-meta) output, if the axes are all correct as you expect them to be. (debug_print_layer_output_template can and should always be enabled. It will not make it slower.) You can also temporarily enable debug_print_layer_output_shape, which will really print the shape of all tensors. That way you can verify how it looks like.

Your usage of ReinterpretDataLayer looks very wrong. You should never ever explicitly set the axes by integer (like "set_axes": {"B": 1, "T": 0}). Why are you doing this at all? This could be the reason why it is messed up in the end.

Your repeat function is not very generic. You are using hard coded axes integers there as well. You should never do that. Instead, you would write sth like:

You can try with compile = False parameter in your load_model call. This will remove any metadata related to optimizers and loss function and since you have a reinitialize function and if you don't need to train it from where you stopped last time, that won't be a problem I guess.

As you said it works with compile = False, I think the problem might be in your custom functions for loss/optimizer etc, I can't pin point what exactly is the problem, you can try tf. instead of tf.keras.backend. if you want.

Just replace