Music Transition

Prerequisites

Set up

git clone https://github.com/ToBigsSound-1516/transition.git
cd trainsition
pip install -r requirements.txt

Download Lakh Pianoroll Dataset

python3 prepare_data.py ./your_data_path -1

The last option means the number of pools. Use 0 to turn off multi-processing option. Use -1 to set the pools automatically.

Training

python3 main.py --train --ckpoint ./your_checkpoint_path --data_path ./your_data_path --n_epochs 100000

Base Arguments

--ckpoint: str, default = "./ckpoint". checkpoint will be saved here. If it is not .pt file, the latest checkpoint will be loaded.

--latent_dim: int, default = 128.

--cpu: Use this option to unable CUDA training. When there is no GPU in your environment, this option will be used automatically.

Training Arguments

--train: Use this option to run the train script.

--data_path: str, default = "./data". There must be a data.npy in your data_path. Please refer the prepare_data.py.

--lr: float, default = 0.0001. learning rate.

--batch_size: int, default = 8192.

--n_epochs: int, default = 1000.

--ckpoint_interval: int, default = 100.

--save_sample: Use this option to save some midi samples while training.

Mixing

python3 main.py --ckpoint ./your_checkpoint_path --midi_path1 ./first_midi_path --midi_path2 ./second_midi_path --start1 100 --start2 200 --midi_save_dir ./mixed_midi_will_be_saved_here

Mixing Arguments

--midi_path1, --midi_path2: str. The midi files to mix.

--midi_save_dir: str, default = "./data/mixed". Mixed midi file will be saved here.

--start1, --start2: int. Mixed point.

--mix_margin: int, default = 4 * measure_resolution. Mixed file will be saved with the margin.

Backend

How to run flask

python3 app.py

You can get some sample result HERE.