Home

Awesome

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

Namecontrol_input_dimcontrol_n_layercontrol_typedecoder_activationencoder_activationfilm_typefilters_layer_1hop_lengthinput_channelskernel_sizelast_activationlrn_fftn_layersnum_frameoptimizerstridetest_result/agg/bass_ISRtest_result/agg/bass_SARtest_result/agg/bass_SDRtest_result/agg/bass_SIRtest_result/agg/drums_ISRtest_result/agg/drums_SARtest_result/agg/drums_SDRtest_result/agg/drums_SIRtest_result/agg/other_ISRtest_result/agg/other_SARtest_result/agg/other_SDRtest_result/agg/other_SIRtest_result/agg/vocals_ISRtest_result/agg/vocals_SARtest_result/agg/vocals_SDRtest_result/agg/vocals_SIR
complex_2048_512_128eval44densereluleaky_relucomplex245122[5,5]sigmoid0.00120486128adam[2,2]8.848354.813252.7954654.1146159.690444.29792253.4923654.635266.934553.878711.853761.08556256.06474752.20809252.497498.3487875
complex_32eval_44densereluleaky_relucomplex322562[5,5]sigmoid0.00110246256adam[2,2]8.08655754.795292.11459752.645902510.0199054.91580753.7952754.923337.51220254.586831.7054151.074067.4706953.633712.4158656.5487125
cunet_mme_sigmoid_32-eval44densereluleaky_relusimple322562[5,5]sigmoid0.00110246256adam[2,2]7.64625254.901941.849561.93136259.49972254.66947253.3271254.1132357.6484054.6598251.5004950.55410256.7109853.6021052.122355.72728

How to use

Training

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

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