deezer / spleeter

Deezer source separation library including pretrained models.
https://research.deezer.com/projects/spleeter.html
MIT License
25.99k stars 2.85k forks source link

[Bug] Spleeter Training is not working #764

Open MohammedMehdiTBER opened 2 years ago

MohammedMehdiTBER commented 2 years ago

Description

Spleeter Training stucks but spleeter separating works.

Step to reproduce

Installed FFmpeg using PATH variable Installed Spleeter using python3.9 pip Downloaded MuseDB uncompressed version Set 22Khz (F=2048) in the config files. Downloaded CSV files from this respository and put them in their proper places. Extracted MuseDB to a folder named configs containing the two CSVs and musedb.config file Started the action using the spleeter train command

Output

Continued to load the musedb like a loop until It stucks at: INFO:spleeter:Audio data loaded successfully

Environment

OS Windows 11
Installation type PIP / FFmpeg (PATH)
RAM available 12GB
Hardware spec Intel Core i5 gen 2

Additional context

jlncrml commented 3 months ago

It's very likely training. You can easily check if it's training by calling tf.get_logger().setLevel('DEBUG') before the tf.estimator.train_and_evaluate(estimator, train_spec, evaluation_spec) call.

xuzhang12 commented 2 months ago

You can try adding the --verbose parameter when using the training command, for example: spleeter train -p configs/musdb_config.json -d E:\AI_Study\spleeter --verbose. This will give you detailed information about the training process. I got stuck because the train_max_steps parameter in musdb_config.json was too large, and since I’m using a CPU, the steps are increasing very slowly. Try reducing the train_max_steps to see if it resolves the issue.