Setup Guide

1. Download MusicNet data in .npz format from https://drive.google.com/drive/folders/11vtgfqEJukUNYvqZ6q9zFej7-S06QUSi?usp=sharing.

2. Open "2. Sentiment_based_MusicComposer (All data).ipynb" file in python notebook and set the directory path for data and other experiments (output).

3. Execute the "Read music files" cell. This step would generate a huge (~50GB) .hdf5 file which includes the audio, spectrogram and score features from the data. 

4. To train the model, verify if CUDA (GPU) is available. If available, change the cuda = torch.device("cpu") to cuda = torch.device("cuda") in the code and execute.
If GPU is not available, execute the code as is. 

Note: The best epoch for 170 songs in our case was, 290 and BATCH_SIZE = 128. Incase, the datasize is changed, experiment with these values for hyper-paramter tuning. 

5. Execute Model building (step-3 in code). verify that model.py is available in the same directory as your code for the project. 

6. For model inference, change the path of input file and enter the emotion to genererate the output on. 

The generated music is in .wav format which can be analyzed using the 3. EVALUATIONS.ipynb code.