This is a standalone pytorch implementation of Counterpoint by Convolution (Coconet) . This implementation is based on kevindonoghue's implementation, while changed few parameters and network structures for reproducing the network in the paper.
The code in this repo is modified from kevindonoghue's implementation by Yusong Wu and Kyle Kastner.
Although the code is modified to reproduce the original paper, the loss in here is just plain cross-entropy between input and output. Because the use of cross-entropy, it filters out all the training data with rest notes.
In the original paper (eq.3) and original implementation , the loss is not counted for back-propagation where input is not masked (because the objective would be simply to copy input to the output), and the loss is scaled by 1/(T-num_unmasked+1). In this implementation we found the un-scaled loss still produce decent output.
pytorch, pretty_midi, midi2audio, pyfluidsynth
To generate random Bach chorales in a length of 8 bars, simply clone the repo and run python generate.py.
The pre-trained network is trained using batch size of 64 and 50000 steps of updates.
To train the network, simply download the "Jsb16thSeparated.npz"
in here, and run the train.py.