This is a written project with regard to processing sound using keras deep learning library. These are the planned application:
- Genres classification
- Music auto-tagging
- Instrumental classification
- Tempo classification
- Deep Learning Jazz generator
- Sort music by genres, artist, tempo, tags
- Piano analysis
- Researching on processing Midi file format: tracks, messages, channels, etc.
- More research on processing Midi file format: tracks, messages, channels, etc.
- More research on MIDI file processing and feature engineering.
- Add data generator.
- Add code to read and parse midi file, converting them to structrured form.
- Add helper and utilization function.
- Edit reference for paper
- Add critical data preprocessing function: convert array of NoteEvent to a 2D array with shape (max_tick, note_range) where x axis is pitch and y axis is time.
- Update compressed dataset
- Finish Data Preprocessor and Data Generator. Working on building Convolutional Neural Network with analytic pixel. The paper can be found here
- Try building model with generative adversarial neural network.
- Build model with PixelCNN++ philosophy.
- Build model with Variational Autoencoder philosophy.
- Train 16000 epochs on CNN model, with dataset of single-track midi files.
- Log file surpassing (edit .gitignore)
- Switch from CNN model to seg-to-seg LSTM model. Do more research.
- Split the main notebook into one with music generator processing as image processing, the other as sequence processing (CNN and LSTM).
- Utilize a unique class for a note event for efficient data pre-processing.
- Add function to build note matrix from midi tracks.
- Succesfully train on Bach's prelude and fuge in C major BWV 846 piece, 2000 epoches, with accuracy of 89% and loss 12%.