search

axeber01 / ngcc

Neural Generalized Cross Correlations https://arxiv.org/abs/2208.04654
MIT License
23 stars 11 forks source link

Is this training speed normal? #3

Open ywang411 opened 4 months ago

ywang411 commented 4 months ago
Pasted Graphic
ywang411 commented 4 months ago

I do not know why this is slow, thank you! My graphic card is A4500, and I run this code in a Windows operating system.

axeber01 commented 4 months ago

Generating the impulse responses at high reverberation takes quite a long time. In cfg.py, try lowering the upper bound of t60, or use anechoic = True (this will make the simulations reverberation-free). You could also add functionality to pre-generate the dataset instead of generating new data each epoch.

ywang411 commented 4 months ago

Generating the impulse responses at high reverberation takes quite a long time. In cfg.py, try lowering the upper bound of t60, or use anechoic = True (this will make the simulations reverberation-free). You could also add functionality to pre-generate the dataset instead of generating new data each epoch.

Thank you for your help! I found that lower the upper bound of t60 can reduce the training time, and could you please explain how to pre-generate the dataset? Because I thought that we have already prepared Dataloader before training. And I also tried to modify code to avoid do Labelsmoothing every time I run the main.py, but failed.