Add masked MSE loss

[njellinas]

I noticed that you mask the outputs of the decoder, so there was no need to apply any more 'masking code'. Just in the calculation of the MSE I added a counter of the non-padded elements, changed the loss reduction to 'sum' and then divided with the counter.

Signed-off-by: njellinas nickgreek54@gmail.com

[hadaev8]

Just tried, get error TypeError: forward() missing 1 required positional argument: 'output_lengths'

[njellinas]

In which point did it output this error? I put an extra argument in the Tacotron2Loss forward function, but I give it as x[-1] which are the output_lengths from parse_batch.

[hadaev8]

Trace

TypeError                                 Traceback (most recent call last)

<ipython-input-12-1564558022d7> in <module>()
    307 create_mels()
    308 train(output_directory, log_directory, checkpoint_path,
--> 309       n_gpus, rank, group_name, hparams, log_directory2, checkpoint_path_vanilla)

1 frames

<ipython-input-12-1564558022d7> in train(output_directory, log_directory, checkpoint_path, n_gpus, rank, group_name, hparams, log_directory2, checkpoint_path_vanilla)
    244             y_pred = model(x)
    245 
--> 246             loss = criterion(y_pred, y)
    247             if hparams.distributed_run:
    248                 reduced_loss = reduce_tensor(loss.data, n_gpus).item()

/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    491             result = self._slow_forward(*input, **kwargs)
    492         else:
--> 493             result = self.forward(*input, **kwargs)
    494         for hook in self._forward_hooks.values():
    495             hook_result = hook(self, input, result)

TypeError: forward() missing 1 required positional argument: 'output_lengths'

[njellinas]

I guess you have a different codebase because this line 307 create_mels() does not exist in the master branch. And wherever there is a criterion function I have put the last argument.
Or should I pull from a different branch than master?

[hadaev8]

Yes, I have my own modification (minor) of this repo.
(and I'm not a contributor or something, just tried to use your loss function).
Maybe couz I use fp16 train.
Here is the code.

https://colab.research.google.com/drive/1tBOXMBNbAkS-zHDIvP3hfFEUNf-kGosd

[njellinas]

Oh OK then that makes sense. Look at the changes in the diffs, and apply them to your code. You are missing the last argument at criterion which must be the Mel lengths.

[hadaev8]

Oh, sorry, somehow I missed what i copy-pasted train file and just pulled your changes to my repo.

[rafaelvalle]

@njellinas Does adding the masked MSE loss improve the model?
For example, time to learn attention or lower validation loss?