src/third_party/MusicSourceSeparationTraining/docs/changes.md

Changes

v1.0.2

• Added multi GPU validation (earlier validation was performed on single GPU)
• training.batch_size in config now must be set for single GPU (if you use multiple GPUs it will be automatically multiplied by number of GPUs)

v1.0.3

• Added "spawn" fix for multiprocessing
• Function get_model_from_config now takes path of config as input.
• On latest version of pytorch some problems with torch.backends.cudnn.benchmark = True - big slow down. Fixed version 2.0.1 in requirements.txt
• --valid_path parameter for train.py now can accept several validation folders instead of one. Added warning if validation folder is empty.
• Small fix for AMP usage in Demucs models taken from config
• Support for Demucs3 mmi model was added
• GPU memory consumption was reduced during inference and validation.
• Some changes to repair click problems on the edges of segment.
• Added support to train on FLAC files. Some more error checks added.
• viperx's Roformer weights and configs added
• --extract_instrumental argument added to inference.py