Examining the Mapping Functions of Denoising Autoencoders in Music Source Separation

Support material for the article: "Examining the Mapping Functions of Denoising Autoencoders in Music Source Separation".
It contains the code for:

• Analyzing an open-access dataset
• Implementations of the Denoising Autoencoder and the variants for music source separation
• Loss functions employed in the article
• The algorithm that is denoted as the Neural Couplings Algorithm (NCA)
• Reproducing the figures of the corresponding article
• Toy-examples for the gradient calculation

It also contains binary files used in:

• Creating the segments for the testing of the NCA and creating the corresponding visualizations
• The optimized models' weights
• Python requirements text file

How to use:

1. Install the requirements
2. Set the data-set paths in "processes_scripts/run_me.py" and "processes_scripts/analyze_data.py"
3. Run "process_scripts/analyze_data.py" to analyze and store the spectral data for training the models
4. Run "process_scripts/run_me" to train (in the case that re-training is necessary with "training_flag = True") and to compute the linearly composed matrices for reproducing  one of the figures
5. Run "process_scripts/run_couplings.py" to plot the average linear composition, evaluate the strategies of the NCA algorithm and make plots on specific test data.

TBD:

• Github gist for the NCA algorithm
• Pytorch and python updates