Redundant decoder in ESPnetASRModel when ctc_weight is 1.0

[pyf98]

Describe the bug
When ctc_weight=1.0, the ESPnetASRModel should not have a decoder in its parameters, regardless of the decoder config. However, now it will add a default RNN decoder.

Note that the typical way to use CTC only is to set ctc_weight: 1.0 and remove decoder and decoder_conf in the training config.

I believe this bug is introduced in a recent commit, because it worked well previously.

https://github.com/espnet/espnet/blob/master/espnet2/asr/espnet_model.py#L146

It should be something like this:

            if ctc_weight == 1.0:
                self.decoder = None
            else:
                self.decoder = decoder

[sw005320]

Thanks!
@ftshijt, can you fix it?
Also, is there a way to add a test to ensure this does not happen?

[ftshijt]

I changed the previous behavior mainly because the self.decoder = None is not user-aware from the config (e.g., user can still set decoder with anything, but eventually, there is no decoder is initialized).

For that purpose, I add a None decoder and ask the user to set decoder=None when they do the initialization. So probably, we may just throw an error (e.g., you should set decoder to none etc.).

What's your opinion on it?

[pyf98]

This will affect our previous CTC configs and models.

[ftshijt]

I see. So I will add a warning msg and recover the previous options. thanks for bringing up the issue.

[pyf98]

Actually, I think the decoder should be None by default. The user should specify a decoder when they really want to use one.

[ftshijt]

Yeah, I thought that would be better, but the only concern I had at the time is that, it may lead to some issues to other parts (e.g., test functions etc.) Anyway, since we brought up the issue, let's just do it and fixes issues with the change.