AI Composer is a student project with a simple task: Create an AI using a keras LSTM model that composes new instrumental music.

Our project is divided into three parts:

• Data processing
• LSTM model Training
• Pygame Application / song generation

• python 3.8

Optional:

• Cuda + cuDNN for GPU training

Run pip install -r requirements.txt

The generator our model uses for training can only process .h5 files. To facilitate this you first need a dataset of midi files. You then need to set up config.py with the correct values.

You can change these values in config.py.

# Where the midi files process_data.py uses are stored
MIDI_FOLDER = your dataset path

# Where to store the files converted by process_data.py
CONVERTED_PATH = your converted dataset path

Run process_data.py to start the processing.

We strongly advise training on a GPU, as it greatly accelerates the training speed.

Our training process uses these config values when training.

# Where the trained models are saved
MODELS_FOLDER = your model path

# How many time steps to train on at a time
SEQUENCE_LENGTH = your length

# The batch size to train with
BATCH_SIZE = your batch size

# The training learning rate
LEARNING_RATE = your learning rate

# The number of epochs to train
EPOCHS = your number of epochs

# The number of steps pr epochs
STEPS_PR_EPOCHS = your number of steps pr epochs

# Where to store the files converted by process_data.py
CONVERTED_PATH = your converted files

Run training.py to start the training.

The model will be saved to MODELS_FOLDER/model when done. It will also at the end of every epoch save the most accurate model to MODELS_FOLDER/most_accurate and the model with the least loss to MODELS_FOLDER/least_loss.

Run test_model.py to test the model

You can alter the python file to choose whether to create a new song from scratch or continue from a song in your CONVERTED_PATH dataset.

You can also change the temperature and note frequency of the predicted songs. A higher temperature will mean a more creative and chaotic end song, while a higher note frequency will simply mean more notes.

https://soundcloud.com/espen-fosseide/sets/ai-generated-music?utm_source=clipboard&utm_medium=text&utm_campaign=social_sharing