fix: moving the operations to the GPU if GPU available #1522
Conversation
| @@ -606,7 +606,7 @@ def __call__( | |||
|
|
|||
| embeddings[f] = self.session_.run( | |||
| output_names=["embs"], | |||
| input_feed={"feats": masked_feature.numpy()[None]}, | |||
| input_feed={"feats": masked_feature.cpu().numpy()[None]}, | |||
There was a problem hiding this comment.
Here, you are actually sending the features back to CPU, before sending them to the inference. Therefore, I don't really understand why this should speed things up.
Can you please provide me with a bit more context?
Ideally, a Colab notebook illustrating the speed up that this brings?
There was a problem hiding this comment.
If we don't use the cpu() method, it will result in an error. You can refer to the official PyTorch documentation for this. What I've observed is that if we don't place the tensor on the GPU, both torch.vstack() and torch.fft.rfft() will consume a significant amount of CPU resources.
# Under the 'wavs_16k' directory, there are a total of 8 '.lst' files, each of which contains paths to audio files.
# After running the following code, 8 PyAnnote processes are launched on a computing node equipped
# with 8 V100 GPUs, with each process occupying one GPU.
for list in `ls wavs_16k/*.lst`
do
sleep 1s
echo $list
srun -N1 -n1 --gres=gpu:1 python scripts/pyannote_main.py ${list} > logs/`basename $list .lst`_${time}.log 2>&1 &
done
| @@ -338,6 +338,10 @@ def iter_waveform_and_mask(): | |||
| for i, batch in enumerate(batches, 1): | |||
| waveforms, masks = zip(*filter(lambda b: b[0] is not None, batch)) | |||
|
|
|||
| if torch.cuda.is_available(): | |||
There was a problem hiding this comment.
I think we should move to GPU only if self.device == 'cuda'. This can lead to some troubles, for example if user have gpu available, but he wants to run it on cpu. Also it don't match with .to behaviour
| @@ -338,6 +338,10 @@ def iter_waveform_and_mask(): | |||
| for i, batch in enumerate(batches, 1): | |||
| waveforms, masks = zip(*filter(lambda b: b[0] is not None, batch)) | |||
|
|
|||
| if torch.cuda.is_available(): | |||
| waveforms = tuple([x.cuda() for x in waveforms]) | |||
There was a problem hiding this comment.
I believe that we can cast waveform somewhere before. Because casting to cuda for every element in list seems not efficient. May be we can cast original file or waveform to cuda before splitting it into batch? So we don't need to make multiple casts to cuda for every element in batch instead of 1 cast
|
Closing this PR as I think #1529 (just merged) fixes (at least part of) the issue. |
According my testing, before the code modifications, the 8 speaker diarization processes running on 8 GPUs(A100) separately of the same node, with some operations still on the CPU, the RTF (Real-Time Factor) was 3.32, After moving the operations to the GPU, the RTF reached 0.015. On V100, the RFT reached 0.02.