Too many CPU resources for fine-tuning

[sausage-333]

Thanks for your great work.

By the way, I have found that CPU of my machines are overloaded compared to GPUs when I use "audio" modality.
It is okay when I use only "video" modality.
I mean the computation bottleneck is occurred in CPU not GPU. (i.e. GPU usages on 0~20%)

How can I dissolve this problem?
Maybe computing MFCCs in advance would help, but I am not sure.

[chevalierNoir]

Have you tried increasing the number of workers per task? Also, the filterbank extraction could be potentially replaced by something more efficient like these in torchaudio.

[sausage-333]

Do you mean that do I need to increase num_workersor nprocs_per_node? There is a bottleneck on CPU, so loading more work to CPU would not solve this issue I guess...