The architecture of the encoder of this VAE was adapted from the approach for a denoising autoencoder by Bretan et al. (2017). The decoder consists of the inverse mapping without reparametrization.
To use the examples please activate virtual environment and start jupyter notebook from root of this project by:
jupyter notebookThen navigate to this folder and open VAE_Examples notebook. Note that you need a song in MIDI format to feed to the network. I have recorded short MIDI files to use, they are located in the utils folder midi_files.
-
Download a MIDI dataset of your choice and collect all files in one folder.
-
Transpose all files in this folder using the Transposer.py in Preprocessors folder. If you want to transpose your MIDI files to all possible pitches (recommended), try:
python Transposer.py --file_path /path/to/files --tp_by 60 --tp_step_size 1- Cut MIDI files using MIDI_file_cutter.py. MIDI file cutter will create a new folder sequences in your MIDI data folder which you can use for training.
python MIDI_file_cutter.py --file_path /path/to/transposed/files-
Create validation with different dataset. Works as depicted in 2. and 3., but for validation you don't necessarily need as many files. So you can use a higher tp_step_size in step 2.
-
Train the VAE with your created MIDI files
python VAE_Train.py --file_path /path/to/transposed/files/sequences --validation_path /path/to/valid/files --model_name my_model_123- Track your progress with tensorboard by running:
tensorboard --logdir=/path/to/vae/folder