ytsrt66589 / ddsp-piano-pytorch

Introduction

This is the project of implementing DaFX20in22 paper: DDSP-Piano: Differentiable Piano model for MIDI-to-Audio Performance Synthesis by the Pytorch framework.

Official information

• Official implementation is by TensorFlow: original implementation
• Official paper link: paper link

Future To-Do

• [?] Inference.py
• [?] Better data organization
• [?] Better code organization

Run

1. create data, data_cache, data_cache_val directory
2. download maestro v3.0.0 dataset
3. put it in the data directory ( data/maestro-v3.0.0 )
4. run bash full_training.sh -> start training

Training Strategy

In the original paper, there are 2-stage training and fine-tune training by the following three scripts. However, only running the first training strategy is good enough.

python3 train.py \
	--batch_size $phase_1_batch_size \
	--epochs $phase_1_n_epochs \
	--lr $phase_1_learning_rate \
	--phase 1 \
	$maestro_path \
	$maestro_cache_path \
	$exp_dir

python3 train.py \
	--batch_size $phase_2_batch_size \
	--epochs $phase_2_n_epochs \
	--lr $phase_2_learning_rate \
	--phase "2" \
	--restore "$exp_dir/phase_1/ckpts/" \
	$maestro_path \
	$maestro_cache_path \
	$exp_dir

python3 train.py \
	--batch_size $phase_3_batch_size \
	--epochs $phase_3_n_epochs \
	--lr $phase_3_learning_rate \
	--phase "3" \
	--restore "$exp_dir/phase_2/ckpts/" \
	$maestro_path \
	$maestro_cache_path \
	$exp_dir