IEMOCAP recipe - missing lengths in call to embedding module ?

Hello Speechbrain team,

While trying to reproduce the results mentioned in https://github.com/speechbrain/speechbrain/tree/develop/recipes/IEMOCAP I figured out that the accuracy on the test set varies w.r.t the batch size in the test set dataloader, sometimes reducing the accuracy by up to 7% when batch_size = 1.

Further investigations in the code made me think that the ‘lengths’ parameter might be missing in the call to the embedding module (train.py line 37).

I tried to change it to: embeddings = self.modules.embedding_model(feats, lengths=lens) and ran the train and test stages again. Then the variations are less important but are still present : around 1% difference from one test to another (train stage done only once).

So I have 2 questions:

1. Do you think the modification hereabove is correct?
2. What could explain the remaining fluctuations with respect to the batch size of the test set?

Thank you in advance

[wikong]

Hi @aheba @TParcollet @mravanelli Thank you for the fix, and also for the additional work with wav2vec2. Very interesting indeed…

[mravanelli]

I think the correct way is to pass the lenghts parameter to the embedding model (this way zero padded elements are removed from the statistical pooling operation). @aheba, can we do it in your ongoing PR?