# Generating MIDI music

## Structure of this project
/data            - directory for all MIDI files
encoding.py      - functions that convert the musical composition between different formats
generate.py      - script for generating new music
/generated_midi  - directory that holds generated music
make_corpus.py   - script that creates a text corpus from the MIDI files
midi_report.pdf  - documentation of the project
models.py        - architectures of the used neural network models
/model_weights   - model weights are stored here during and after training; contains weights of pretrained models
requirements.txt - required libraries
/survey_samples  - generated samples that were used in a user survey for evaluation
test_corpus.txt  - corpus for final evaluation
train.py         - script for training the models
train_corpus.txt - training data for the models


## Usage
### Make corpus

$ ./make_corpus.py

Takes all MIDI files in all sub-directories of ./data and creates a single corpus file.

### Train models

$ ./train.py

Trains the three implemented models (simple lstm, lstm, cnn-lstm) on the train_corpus.
This may take several days to finish even on a powerful computer.

### Generate MIDI

$ ./generate.py

Generates new MIDI files which will be stored in the ./generated_midi folder.
Uses all three models and generates 20 samples with each by default.