Perceiver Music Transformer Toolkit

Toolset for easy development with Perceive-Music-Transformer (https://github.com/asigalov61/Perceiver-Music-Transformer). Only supports multi-instrumental, currently.

Paper

• https://arxiv.org/abs/2202.07765

Links

• https://github.com/asigalov61/Perceiver-Music-Transformer
• https://github.com/asigalov61/Euterpe
• https://github.com/asigalov61/tegridy-tools

Install

python setup.py install

• writing requirements.txt is still in progress.

Tensorboard

tensorboard --logdir ./tensorboard_logs

Example

• Create dataset

# load ./mids/*.mid and create dataset to ./dataset dir`
create_dataset("mids", "dataset")

• load dataset

# load dataset from ./dataset dir`
train_data = load_train_data("dataset")

• train

pmt = PerceiverMusicTransformerToolkit() # you can specify params here to train
pmt.init_model()

# prepare training data
train_loader, val_loader = pmt.prepare_train_data(train_data)

# train and save checkpoints to ./checkpoints dir
pmt.train(train_loader, val_loader, "checkpoints")

• eval

# generate continue
input_data = midifile_to_input("seed.mid")
output_data = pmt.generate_continue(input_data)

# output to midi
input_to_midi(output_data, "output.mid")

# save toolkit params
save_toolkit_params(pmt, "params.dump")

• load

pmt = load_toolkit_params("params.dump")
# load ckpt
pmt.load_model("checkpoints/...")

Note for parameters tuning

• SEQ_LEN: Perceiver-AR inputs are fixed length. You can specify max length of its sequence.
• PREFIX_SEQ_LEN: Length of Key and Value for cross-attention.
• HEADS: Number of heads of multi-head-attention.
• DEPTH: Number of latents-self-attention.