Can I know which version of pytorch & CUDA was this evaluated on? about amongthesleep HOT 3 OPEN

Hi, we were running this on a cluster which was fairly outdated, this might explain the issues you are running into.

Here are the versions:
Python: 3.5
CUDA: 8.0.44.
Pytorch: 0.2.0_3

If possible could you send the error you received.

from amongthesleep.

Hi, I changed my build as suggested by you and the model loads comfortably. However with the spectrogram transform section of the pipeline makes the CPU utilization shoot up to 100% causing the OS tasks to freeze and crash. I've tried changing the process priority using NICE and throttling using cpuutils but no progress.

I'll try to run the spectrogram transform and store the data in a hdf5 format and try to use the network. Thanks for the quick reply. I'll post updates regarding this.

from amongthesleep.

This makes sense as the spectrogram transform is not performed by the gpu and is a very expensive operation.

Pre computing the spectrograms is a good idea. An alternative is to only transform smaller sections of the data into spectrograms for each step. The downside is that this is mor complicated to normalize.

There might also be a gpu based library which can perform the transformation.

Thanks for offering to keep us updated, good luck.

from amongthesleep.