DeepSpectrum working but hangs if a CPU Core maxes out

[larryzhang95]

I'm using DeepSpectrum to extract feature(s) from a series of audio file(s). It loads the model and weights without any issue. However once a CPU core maxes out to 100% and hangs at 100%, and cannot process any further files.

I was wondering if there's one or two work-arounds:

1. Can DeepSpectrum be completely executed on GPU? If so -- how do I configure this with the deepspectrum cli or codebase?
2. Is there a time delay I could obey to make sure the processes are killed properly each time so I don't run into the CPU issue.

I am using the same command line arguments as the example.

I am using a system with 64GB RAM ; i9 processor

[mauricege]

Hi!

I'm sorry to hear that this happens. Could you provide further information on the files you are trying to extract features from?

1. Can DeepSpectrum be completely executed on GPU? If so -- how do I configure this with the deepspectrum cli or codebase?

Unfortunately, this is not possible, as input pre-processing is handled with librosa.

2. Is there a time delay I could obey to make sure the processes are killed properly each time so I don't run into the CPU issue.

I'm not exactly sure what you mean by this? Are your CPU-Cores still maxed out after you have already killed the deepspectrum process?

I am using a system with 64GB RAM ; i9 processor

A thing worth trying with your hardware is to manually limit the number of workers created by deepspectrum with the --number-of-processes option, as by default, the tool uses as many workers as you have CPU cores in your system.

If the files you are trying to process are not confidential, you could provide me with them so that I can further investigate the problem.

Cheers,
Maurice

[larryzhang95]

Hi Maurice,

I realized that the reason that the processes are hanging is due to the fact that I am wrapping the execution within a python script -- meaning that when I kill the current execution I'm not killing the process itself.

When directly executed from cmdline, when I kill the execution, the processes themselves are killed too.

The files I'm processing are confidential. However I can describe some of the nature.

I've been processing files of different sizes (e.g. a few seconds to several minutes). When I try the smaller files (order of seconds), deepspectrum works perfectly fine for around 10-15 of them, but then maxes out the cores. When I try a larger file (order of minutes), it immediately maxes out the cores.