Fiddler: An AI Music Composer

The goal of the project is to model music compositions by capturing temporal dependencies in classical music composi- tions and eventually generate novel music compositions from the learned model. We use approximately 24000 music com- positions transcribed in ABC notation. We implement count- based n-gram language model, use its results as a baseline for recurrent neural network based methods and assess their abil- ity to generate structurally coherent, human-pleasing music. We achieved test accuracy of 69.78% for char-RNN and 73% for seq2seq model with both methods generating valid music compositions.

Detailed experiments, methodlogies and results are discussed in paper.pdf.

Install

In order to install fiddler run the following command:

python setup.py install

This will install fiddler as a command line tool.

Commands

Training Recurrent Neural Network

fiddler train_rnn [options]

train_rnn command supports following options:

Usage: fiddler train_rnn [OPTIONS]
  Train neural network

Options:
  -f, --file PATH            Train Data File Path
  -b, --batch-size INTEGER   Batch size
  -l, --layers INTEGER       Number of layers in the network
  -r, --learning-rate FLOAT  Learning Rate
  -n, --num-steps INTEGER    No. of time steps in RNN
  -s, --cell-size INTEGER    Dimension of cell states
  -d, --dropout FLOAT        Dropout probability for the output
  -e, --epochs INTEGER       No. of epochs to run training for
  -c, --cell [lstm|gru]      Type of cell used in RNN
  -t, --test-seed TEXT       Seed input for printing predicted text after each
                             training step
  --delim / --no-delim       Delimit tunes with start and end symbol
  --save / --no-save         Save model to file
  --help                     Show this message and exit.

If fiddler is not installed as command-line tool, you can use the same command using python src/cli.py with same arguments.

Generate music using a trained RNN model

Usage: fiddler generate [OPTIONS]

Options:
  -m, --model_path PATH  Directory path for saved model
  --help                 Show this message and exit.

Contributors

Manthan Thakar - Character-level Recurrent Neural Network and Sequence to Sequence Implementation

Rashmi Dwarka - Character-level Recurrent Neural Network and Sequence to Sequence Implementation

Tirthraj Parmar - Character-level n-gram language model implementation