Optimize weighted embedding extraction with pyannote 3.1

[Juan Coria]

With pyannote 3.1, we could do only 1 forward pass of the audio instead of num_speakers when extracting embeddings with weights. This is probably at least one of the causes behind the pytorch version of the wespeaker embedding model being that much slower.

This optimization would also reduce the latency of pyannote/embedding so both would need to be re-computed in the README table.

Important: we should verify that this method is also compatible with masking (e.g. in speechbrain embeddings)

[Juan Coria]

After a quick run of the diarization pipeline with 5s latency, it seems that pyannote 3.1 is hurting performance: DER=27.3 -> DER=29.1 with the exact same models and hyper-parameters.

This is most certainly due to the new API for embedding models compatible with weighted stats pooling, so contrary to my initial idea of leaving this for v0.10, I think it's more a compatibility issue and less a matter of latency optimization.

[Juan Coria]

cc @hbredin, in case you can think of any other changes that could have caused this

[Hervé BREDIN]

cc @hbredin, in case you can think of any other changes that could have caused this

I can't think of any. I did not witness any change of performance in offline speaker diarization when switching from hbredin/wespeaker-voxceleb-resnet34-LM (ONNX) to pyannote/wespeaker-voxceleb-resnet34-LM(PyTorch).

[Juan Coria]

@hbredin I narrowed it down to a change in the interpolation method for the weights:

- weights = F.interpolate(
-     weights, size=num_frames, mode="linear", align_corners=False
- )
+ weights = F.interpolate(weights, size=num_frames, mode="nearest")

Any particular reason for this change? I guess interpolating before calling the model should do the trick but I'm curious.

[Hervé BREDIN]

MPS support. "linear" interpolation is not yet supported with MPS backend.

[Juan Coria]

@hbredin MPS is needed for training though, right? Any way we can make it apply a different strategy? Because to interpolate before calling the model I actually need to know the number of frames that the model will produce internally right before the stats pooling.

I suggest something like:

model = Model.from_pretrained("pyannote/embedding")
model.set_interpolation_method("linear")
embeddings = model(waveform, weights)

Inside the forward method of the model:

outputs = self.stats_pool(outputs, weights=weights, method=self.interpolation_method)

And to call the interpolation:

def forward(..., interpolation_method: str = "nearest"):
    ...
    weights = F.interpolate(weights, size=num_frames, mode=interpolation_method)
    ...

Notice that the interpolation method would need to be changed with a setter because StatsPool is (as it should) created during instantiation.

Also, the same thing could be implemented for the masks in other embedding models (i.e. speechbrain and nemo). Right now I can't take advantage of this optimization because I would need to keep track of which model is compatible with it, and the rule may very easily change in the future, leading to compatibility hell.

I'm willing to open PRs for both features.

[Juan Coria]

Since everything is working fine except for that performance loss on AMI, I think I'd prefer not to wait for this fix to release v0.9. I'll add a note to the reproducibility section of the README to indicate that pyannote<3.1 should be used to get the exact same results and link this issue.

In any case, probably re-tuning hyper-parameters for this new interpolation method will give similar results.