Tensorflow implementation of folk-rnn (https://github.com/IraKorshunova/folk-rnn). The original folk-rnn is implemented using Theano and Lasagne.

Written in python==3.6.8 and tensorflow==1.12

Run python train_rnn.py config5 data/allabcworepeats_parsed (or other training datas) for model training.

Checkpoints and metadata such as token2idx dict are stored in metadata folder. Relaunch training will automatically restore the checkpoints.

Run python sample_rnn.py metadata --seed "<your seed note sequence>" for model inference. The first two token of the seed are the measure and the key. A seed example M:3/4 K:Cmaj. It will load the latest checkpoints in metadata folder.

Support CudnnLSTM for CUDA GPU training, and normal LSTM for CPU inference as well.

The model results are not compared with the original implementation since I don't have a GPU. Comparison test is welcome!

Rong Gong, 2019 MIT License