RAM usage

[kichel98]

Hello,

thanks for all your work 👍 Predicting using Leaderboard A model is working as expected on Linux and Windows.

However, I run into problems with memory during the separation of longer mixtures. When I try to separate a song from MUSDB18 (test part, song Georgia Wonder - Siren, 73 MB, 7m 10sec), I get the following error:

Linux:

Command terminated by signal 9
Exit code 137

Windows:

onnxruntime.capi.onnxruntime_pybind11_state.RuntimeException: [ONNXRuntimeError] : 6 : RUNTIME_EXCEPTION : Non-zero status code returned while running Conv node. Name:'470_nchwc' Status Message: bad allocation

Looks like an Out Of Memory error, it works fine for smaller mixtures. But a Linux machine, on which I test your solution has about 46 GB RAM, which is pretty big in my opinion. My question is: have you ever faced that kind of issue during your testing on MUSDB18? Do you have any workarounds for OOM errors? Do you know some hardware requirements?

[ws-choi]

Hi @kichel98 , Thank you for reporting this issue. I will resolve this by this month, hopefully.

[lyndonlauder]

Hi @ws-choi

did you get around to resolving this issue? Your model is very popular and the karaoke community would be very thankful for this.

[ws-choi]

Hi @kichel98, @lyndonlauder , Sorry for the delay. I forgot this thread for a while.

I pushed a new branch where you can control the batch_size.

TLDR; please switch to this branch and try this commend: python separate.py --batch_size 4

Usage:

    parser.add_argument('--model_name', type=str, default='tdf_extra')
    parser.add_argument('--demucs_name', type=str, default='demucs_extra')
    parser.add_argument('--onnx_name', type=str, default='onnx_B')
    parser.add_argument('--mixture_dir_path', type=str, default='./data/test')
    parser.add_argument('--results_data_path', type=str, default='./data/results')
    parser.add_argument('--batch_size', default=None, type=int)