Lofi Transformers

Installation

git clone --recurse-submodules -j8 <this-package>

conda create -y -n lofi_transformers python=3.8
conda activate lofi_transformers
conda install -y -c conda-forge cudatoolkit-dev
conda install -y pytorch torchvision torchaudio -c pytorch-lts -c nvidia
pip install -r requirements.txt
pip install -e fast-transformers

gdown 14vXsPerw02e7YBrxKMlgkvSB9wD48qVx
unzip processed_lofi_dataset.zip

gdown 19Seq18b2JNzOamEQMG1uarKjj27HJkHu --output exp/pretrained_transformer.zip
unzip exp/pretrained_transformer.zip -d exp/ 

# Train (finetune)
python emopia/workspace/transformer/main_cp.py --task_type ignore --data_root lofi_dataset --path_train_data train --load_dict dictionary.pkl --exp_name finetune_lofi --load_ckt pretrained_transformer --load_ckt_loss 25

# Inference
python emopia/workspace/transformer/main_cp.py --mode inference --load_dict dictionary.pkl --data_root lofi_dataset --path_train_data train --exp_name finetune_lofi --load_ckt finetune_lofi --task_type ignore --num_songs 5 --load_ckt_loss 8 --out_dir gen/finetune_lofi/8

• Need to manually copy utils, saver, models from emopia/workspace/transformer to make generate.py workable.
• You'll also need to put a .sf2 soundfont file into soundfonts folder to render the midi file to audio.
  • For example A320U.sf2 from signal

Discord bot

• Set your own token.

token = os.environ["BOT_TOKEN"]

• Comment line 453 in emopia/workspace/transformer/main_cp.py to prevent saving .npy files.