roger-tseng / conditioned-u-net-pytorch

An unofficial pytorch implementation of Conditioned-U-Net

Conditioned-U-Net-pytorch

An unofficial pytorch implementation of Conditioned-U-Net

News

An extension of this model was released.

Installation

conda install pytorch>=1.6 cudatoolkit=10.2 -c pytorch
conda install -c conda-forge ffmpeg librosa
conda install -c anaconda jupyter
pip install musdb museval pytorch_lightning effortless_config tensorboard wandb pydub
pip install https://github.com/PytorchLightning/pytorch-lightning/archive/0.9.0rc12.zip --upgrade

Evaluation Result

Name	control_input_dim	control_n_layer	control_type	decoder_activation	encoder_activation	film_type	filters_layer_1	hop_length	input_channels	kernel_size	last_activation	lr	n_fft	n_layers	num_frame	optimizer	stride	test_result/agg/bass_ISR	test_result/agg/bass_SAR	test_result/agg/bass_SDR	test_result/agg/bass_SIR	test_result/agg/drums_ISR	test_result/agg/drums_SAR	test_result/agg/drums_SDR	test_result/agg/drums_SIR	test_result/agg/other_ISR	test_result/agg/other_SAR	test_result/agg/other_SDR	test_result/agg/other_SIR	test_result/agg/vocals_ISR	test_result/agg/vocals_SAR	test_result/agg/vocals_SDR	test_result/agg/vocals_SIR
complex_2048_512_128eval	4	4	dense	relu	leaky_relu	complex	24	512	2	[5,5]	sigmoid	0.001	2048	6	128	adam	[2,2]	8.84835	4.81325	2.795465	4.114615	9.69044	4.2979225	3.492365	4.63526	6.93455	3.87871	1.85376	1.0855625	6.0647475	2.2080925	2.49749	8.3487875
complex_32eval_	4	4	dense	relu	leaky_relu	complex	32	256	2	[5,5]	sigmoid	0.001	1024	6	256	adam	[2,2]	8.0865575	4.79529	2.1145975	2.6459025	10.019905	4.9158075	3.795275	4.92333	7.5122025	4.58683	1.705415	1.07406	7.470695	3.63371	2.415865	6.5487125
cunet_mme_sigmoid_32-eval	4	4	dense	relu	leaky_relu	simple	32	256	2	[5,5]	sigmoid	0.001	1024	6	256	adam	[2,2]	7.6462525	4.90194	1.84956	1.9313625	9.4997225	4.6694725	3.327125	4.113235	7.648405	4.659825	1.500495	0.5541025	6.710985	3.602105	2.12235	5.72728

How to use

Training

• train.py

  • parameters related to dataset

    • --musdb_root your musdb path
    • --musdb_is_wav True
    • --filed_mode False

  • parameters for the model configuration

    • --model_name cunet
      • we only support cunet currently
    • stft parameters
      • --n_fft 1024
      • --hop_size 512
      • --num_frame 256
    • Condition generator parameters
      • --film_type simple
      • --filters_layer_1 32
      • --control_type dense
        • we only support dense currently.
        • TODO: conv control
      • --control_n_layer 4
    • U-Net parameters
      • --n_layers 6
      • --stride (2,2)
      • --kernel_size (5,5)
      • --last_activation sigmoid
      • --encoder_activation leaky_relu
      • --decoder_activation relu

  • parameters for the training env.

    • lr 0.001
    • optimizer adam
    • --gpus 1
      • warn 1 (important): if you want to use multi gpus, then we recommend you to use ddp for the distributed_backend, i.e., --distributed_backend ddp
      • warn 2 (important): however, it seems that lightning currently does not support synchronized on_validation_epoch_end so that some log operations might be lost when you try to append logs for every instance in on_validation_epoch_end in ddp mode.
    • --batch_size your batch size
    • --num_workers number of workders
    • --pin_memory True
    • --log_system True
      • or you can use wandb
    • --patience 20
      • for early stop
    • --checkpoints_path ```your_path``
      • audio checkpoints are stored in here.
    • --save_top_k
      • for audio checkpoint saving
    • --run_id run_id
      • if you want to name this run, then use this. default: time stamp
    • --dev_mode True
      • if True, then every dataset deals with 1~4 tracks, which are much smaller than those of counterparts.
    • --float16 True
      • if True, then 16 precision training enabled

example

/train.py --musdb_root ../repos/musdb18_wav --filed_mode True --n_fft 2048 --hop_length 512 --num_frame 128 --filters_layer_1 24 --last_activation sigmoid --film_type complex --num_workers 20 --pin_memory True --log_system wandb --float16 True --batch_size 128 --gpus 2 --distributed_backend ddp --save_top_k 20 --patience 20

Evaluation

• eval.py

  • parameters related to dataset

    • --musdb_root your musdb path
    • --musdb_is_wav True
    • --filed_mode False

  • parameters for the model configuration

    • --model_name cunet
      • we only support cunet currently
    • stft parameters
      • --n_fft 1024
      • --hop_size 512
      • --num_frame 256
    • Condition generator parameters
      • --film_type simple
      • --filters_layer_1 32
      • --control_type dense
        • we only support dense currently.
        • TODO: conv control
      • --control_n_layer 4
    • U-Net parameters
      • --n_layers 6
      • --stride (2,2)
      • --kernel_size (5,5)
      • --last_activation sigmoid
      • --encoder_activation leaky_relu
      • --decoder_activation relu

  • parameters for the Evaluation env.

    • --gpus 1
      • if use set gpus > 1, then automatically eval.py resets it to be 1 :(.
      • It seems that lightning currently does not support synchronized on_validation_epoch_end .
      • Although we have to log every single bbs metric for each track in musdb.test, -we found that some logs are lost when we use ddp.
      • I think that multiple-gpus with dp will work, but i have not tested it yet.
      • To prevent ghost logs, we currently set gpus = 1.
    • --batch_size your batch size
    • --num_workers number of workders
    • --pin_memory True
    • --log_system True
      • or you can use wandb
    • --checkpoints_path ```your_path``
      • audio checkpoints are stored in here.
    • --run_id run_id you want to eval
    • --epoch the epoch (int) you want to eval)
    • --dev_mode True
      • if True, then every dataset deals with 1~4 tracks, which are much smaller than those of counterparts.
    • --float16 True
      • if True, then 16 precision training enabled

example

/eval.py --musdb_root ../repos/musdb18_wav --filed_mode True --n_fft 2048 --hop_length 512 --num_frame 128 --filters_layer_1 24 --last_activation sigmoid --film_type complex --num_workers 20 --pin_memory True --log_system wandb --float16 True --batch_size 128 --gpus 1 --run_id complex_2048_512_128 --model_name cunet --epoch 52

Reference

[1] Meseguer-Brocal, Gabriel, and Geoffroy Peeters. "CONDITIONED-U-NET: INTRODUCING A CONTROL MECHANISM IN THE U-NET FOR MULTIPLE SOURCE SEPARATIONS." Proceedings of the 20th International Society for Music Information Retrieval Conference. 2019.

@inproceedings{Meseguer-Brocal_2019, Author = {Meseguer-Brocal, Gabriel and Peeters, Geoffroy}, Booktitle = {20th International Society for Music Information Retrieval Conference}, Editor = {ISMIR}, Month = {November}, Title = {CONDITIONED-U-NET: Introducing a Control Mechanism in the U-net For Multiple Source Separations.}, Year = {2019}}

[2] Official Github Repository, (Tensorflow-based): Conditioned-U-Net Conditioned-U-Net for multitask musical instrument source separations