Music_Genre_Transfer

Train & Test (only piano)

• Set up dataset

conda create -n {env_converter_name} -y python=3.5.4
source activate {env_converter_name}
pip install -r requirements_converter.txt
bash make_dataset.sh

• Set up environment

conda create -n {env_dl_name} -y python=3.5.4
source activate {env_dl_name}
pip install -r requirements.txt

• Download checkpoint file

gdown https://drive.google.com/uc?id=1dgbup3s6DfoZxREoZl8TmRTs9Djl_bZe
tar -zxvf checkpoint.tar.gz

• Train

python main.py --dataset_A_dir='BD_B' --dataset_B_dir='BD_D' --type='cyclegan' --model='full' --sigma_d=0 --phase='train' --epoch=20 --continue_train=True

• Test

python main.py --dataset_A_dir='BD_B' --dataset_B_dir='BD_D' --type='cyclegan' --model='full' --sigma_d=0 --phase='test' --which_direction='AtoB'
python main.py --dataset_A_dir='BD_B' --dataset_B_dir='BD_D' --type='cyclegan' --model='full' --sigma_d=0 --phase='test' --which_direction='BtoA'

Train & Test (piano & enssemble)

Start with clean state

• Make pretrain dataset

rm -rf datasets
source activate {env_converter_name}
bash make_pretrain_dataset.sh

• Start pretrain

source activate {env_dl_name}
patch -p0 < utils_patch.patch
python main.py --dataset_A_dir='classic' --dataset_B_dir='pop' --type='cyclegan' --model='full' --sigma_d=0 --phase='train' --input_nc=2 --output_nc=2 --epoch=100

• Set checkpoint directory

rm -rf checkpoint/BD_B2BD_D_full_0.0
mkdir -p checkpoint/BD_B2BD_D_full_0.0
mv checkpoint/classic2pop_full_0.0/* checkpoint/BD_B2BD_D_full_0.0

• Set up dataset

rm -rf datasets
source activate {env_converter_name}
bash make_dataset_2.sh

• Train

python main.py --dataset_A_dir='BD_B' --dataset_B_dir='BD_D' --type='cyclegan' --model='full' --sigma_d=0 --phase='train' --epoch=20 --continue_train=True --input_nc=2 --output_nc=2

• Test

python main.py --dataset_A_dir='BD_B' --dataset_B_dir='BD_D' --type='cyclegan' --model='full' --sigma_d=0 --phase='test' --which_direction='AtoB' --input_nc=2 --output_nc=2
python main.py --dataset_A_dir='BD_B' --dataset_B_dir='BD_D' --type='cyclegan' --model='full' --sigma_d=0 --phase='test' --which_direction='BtoA' --input_nc=2 --output_nc=2

Demo

important !! Should turn off cache of your browser

source activate {env_converter_name}
gdown https://drive.google.com/uc?id=1i_0AFIxCT-tOl8eauVt_mwiFwJK43xmz
tar -zxvf checkpoint.tar.gz
python demo.py

Demo video https://drive.google.com/file/d/1n0j9QHHsXLae96PVtpYj5Wg_yeY2bOMY/view?usp=sharing