Music Source Separation Training

The repository is organized as follows:

• configs/config_*.yaml: Configuration files for model hyperparameters and training settings.
• models/*: Implementation of available models for training and inference.
• train.py: Main entry point for single-GPU training.
• train_ddp.py: Entry point for Multi-GPU training (Distributed Data Parallel). Use this for 2 or more GPUs for faster training.
• inference.py: Script to process a folder of music files for source separation.
• dataset.py: Logic for creating training samples.
• valid.py: Validation logic and metric calculation.
• utils.py: Common utility functions.
• ensemble.py: Script to combine results from different models to improve separation quality.
• gui-wx.py: Graphical user interface for the code.

When using ensemble.py, consider the following guidelines to optimize the Signal-to-Distortion Ratio (SDR):

• Model Quality: It is best to ensemble models of equal quality. If one model is significantly worse than the others, ensembling it will reduce the overall quality of the output.
• Method Selection: In experimental testing, avg_wave consistently performs better than or equal to other methods in terms of SDR score.
• Aggressiveness Control: If your models are too aggressive, use min_fft to perform a more conservative ensemble and reduce their influence.

Type 7 is designed to reduce class frequency bias and improve the learning of rare instruments in sparse multi-instrument datasets.

How it works:

1. A random instrument (class) is selected.
2. A random track containing that instrument is chosen.
3. An aligned chunk is loaded from that track.
4. The dataset returns the active_stem_ids (indices of instruments actually present).

Class Frequency Filtering: Instruments that appear in more than max_class_presence_ratio (default: 0.4) of the total tracks are excluded to prevent dominant classes (like vocals) from overwhelming the training.

Returned Values:

• Stems tensor
• Mixture tensor (from mixture.wav if available, otherwise sum of stems)
• active_stem_ids: indices of instruments present in the current sample.

--- Song 1:
------ flute.wav
------ violin.wav
------ mixture.wav

The admin_test.py script is designed for bulk testing of all configurations and models. Unlike tests.py, it does not require you to specify model weights or datasets manually.

By default, it performs validation and inference. You can modify the configurations and parameters being tested by editing the MODEL_CONFIGS variable within the script. This is useful for ensuring that multiple model/config combinations are valid before committing to full-scale training or detailed testing.

In this structure, each folder represents a single instrument (stem). The folder contains all audio files belonging to that specific stem.

Structure:

--- vocals:
------ vocals_1.wav
------ vocals_2.wav
--- bass:
------ bass_1.wav
------ bass_2.wav

--- vocals:
------ vocals_1.wav
------ vocals_2.wav
------ vocals_3.wav
------ vocals_4.wav
------ ...
--- bass:
------ bass_1.wav
------ bass_2.wav
------ bass_3.wav
------ bass_4.wav
------ ...

The MUSDB structure is used when each song is contained within its own folder. Each folder must contain all required stems as individual audio files (e.g., .wav or .flac).

Structure:

--- Song 1:
------ vocals.wav
------ bass.wav
------ drums.wav
------ other.wav
--- Song 2:
------ vocals.wav
------ bass.wav
------ drums.wav
------ other.wav

Type 6 is an extension of Type 4 designed for scenarios where the mixture is treated as a separate signal (e.g., for distillation or consistency losses) rather than just the sum of stems.

Key Features:

• Stems are loaded aligned from the same song position.
• If a mixture.wav file is present in the song folder, it is loaded explicitly.
• If mixture.wav is missing, the mixture is computed as the sum of the stems.
• Supports precomputed random chunks.

Recommended Structure:

--- Song 1:
------ vocals.wav
------ bass.wav
------ drums.wav
------ other.wav
------ mixture.wav

Instead of running via Python and CMD, you can use a pre-compiled Windows executable.

1. Download the EXE file from the provided Mega link.
2. Place the .exe file directly inside the root folder of the Music-Source-Separation-Training repository.
3. You can then create a desktop shortcut to the .exe for easy access.

Regardless of the type used for training, the validation dataset must follow the Type 1 (MUSDB) structure and must include a mixture.wav file for every song. The mixture.wav should be the sum of all stems for that song.

Required Structure:

--- Song 1:
------ vocals.wav  
------ bass.wav  
------ drums.wav  
------ other.wav  
------ mixture.wav
--- Song 2:
------ vocals.wav  
------ bass.wav  
------ drums.wav  
------ other.wav  
------ mixture.wav