About the high cpu occupation

[czecze123]

Hi, Sorry to disturb you, first of all thanks for the implementation of the EEND system in pytorch. Recently, when i training the large model ,i found that the CPU occupation is really high. i want to know the reason of it and how to solve it.

[Xflick]

Actually heavy CPU/disk bandwidth usage is a known issue. Since loading wav files & mfcc feature extraction are implemented online, the code is quite unfriendly to CPU and disk. Decreasing dataloader num_workers may help. Also you may want to check whether your code is running on CPU or GPU.

[czecze123]

Really thank you for your reply. Can we use GPU for loading wav files or mfcc feature extraction?

[Xflick]

I guess PyTorch now supports GPU-based feature extraction (haven't used PyTorch for quite a while :) ). You may refer to torchaudio for more information.

[czecze123]

hi, sorry to disturb you again. when i was training the model using my own data, i found that the training loss can't go down at 0.16, i want to ask you two questions: (1) about training dataset audio number, i used 20000 audios made by the code. what is your training dataset number? (2) about your training loss, i want to ask what is your final training loss and dev loss. Thanks!