Music generation from public MIDI datasets using Markov chains, LSTM language models + sampling, beam search inference.

src/:

• sequence: class for dealing with MIDI and converting to useful representation used by models
• monte_carlo: class for markov-chain model
• basic_rnn, bidirectional_rnn, sequence_rnn, stacked_rnn: various model classes
• evaluation: utilities for computing basic statistics on generate sequences

runner.py: Runs scripts for training, inference from saved checkpoints:

Training:

python3 runner.py -m trnn -i ../data/BachChorales/ -o models/sequence_rnn_BachChorales_128/ --lr 0.0005 --epochOffset 0 --inputLen 1 --layerSize 128 --nepochs 50

Generation:

python3 tester.py -m grnn -i models/sequence_rnn_BachChorales_128/epoch_60/checkpoint.ckpt -o ../outputs/sequence_RNN_BachChorales/sequence_rnn_128_60.mid --inputLen 1 --layerSize 128 --lr 0

See our write-ups for more information.

• Poster
• Project Report

• Akash Mahajan - Masters, Management Science and Engineering
• Suraj Heereguppe - Masters, Institute for Computational and Mathematical Engineering
• Nathan Dalal - Undergraduate, Computer Science