DJtransGAN: Automatic DJ Transitions with Differentiable Audio Effects and Generative Adversarial Networks
This repository contains the code for "Automatic DJ Transitions with Differentiable Audio Effects and Generative Adversarial Networks" 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2022) Bo-Yu Chen, Wei-Han Hsu, Wei-Hsiang Liao, Marco A. Martínez-Ramírez, Yuki Mitsufuji, Yi-Hsuan Yang
The reop is nearly complete; we have already open source all the code you need for not only use (provide a pre-trained model) but also train the DJtransGAN. We will keep improving the document and add more visualization soon. Currently, the repo contains 3/4 of the DJtransGAN code include
- Differentiable DJ mixer includes differentiable fader and differentiable equalizer in the time-frequency domain
- DJtransGAN architecture and its training code
- DJtransGAN pre-trained model and its inference code
The remaining 1/4 of DJtransGANs' is implemented in another repo DJtransGAN-dg-pipeline to make the codebase clean. Moreover, These two repos are totally independent and can be executed individually, and we will detail the process below.
Furthermore, if you want to hear more audio example, please check our demo page here.
We collected two datasets to train our proposed approaches: DJ mixset from Livetracklist and individual EDM tracks from MTG-Jamendo-Dataset.
To be more specific, We in-house collect long DJ mixsets from Livetracklist and only consider the mixset with mix tag to ensure the quality of the mixset. Furthermore, we select the individual EDM track from MTG-Jamendo-Dataset, which means the track with an EDM tag in the total collection. The detailed information about using these two datasets is described in Section 3.1; please check it if you are interested in.
Unfortunately, we can not provide our training dataset for reproducing the results because of license issues. However, we release the training code and pre-trained model for you to try. Contact me or open the pull request if you have any other issues.
pip install -r requirements.txt
You should first clone the DJtransGAN-dg-pipeline and refer to its README.md to generate dataset include mixable pairs and mixes made by professional DJ.
git clone https://github.com/ChenPaulYu/DJtransGAN-dg-pipeline
Next, you should set the configuration in djtransgan/config/settings.py for global usage of the repo, Most important of all, you should set the path of PAIR_DIR, MIX_DIR and STORE_DIR.
PAIR_DIR: the directory contain a collection of mixable pair and its cue points which is generate by DJtransGAN-dg-pipeline.MIX_DIR: the directory contain a collection of mix segment and its cue points which is generate by DJtransGAN-dg-pipeline.STORE_DIR: the directory conatian the training and inferring result of DJtransGAN.
We release several examples in examples/ and script/ for not only training and infering but also individual usage of each component (e.g: differentiable fader, equalizer and mixer). Detail describe below.
You can choose to use fader and equalizer indivdually or use both in the same time. If you want to use it individually, please check examples/mixer/mask.ipynb. If you want to use both in the same time, please check examples/mixer/mixer.ipynb.
To train a DJtransGAN, you need to run the script in script/train.py
python script/train.py [--lr=(list, ex: [1e-5, 1e-5])] [--log=(int, ex: 20)] [--n_gpu=(int, ex: 0)] [--epoch=(int, ex: 20)] [--out_dir=(str, ex: 'gan')] [--n_sample=(int, ex: 4)] [--n_critic=(int, ex: 1)] [--cnn_type=(str, e.g: res_cnn, cnn)] [--loss_type=(str, e.g: minmax, least_square)] [--bath_size=(int, ex: 4)]
--lr: learning rate of generator and discriminator (should provide two value).--log: log interval during the GAN training.--n_gpu: speicify which gpu you want to use.--epoch: number of epoch which indicate how many time of dataset you want to train.--out_dir: the output directory which is going to save the training result.--n_sample: the number of sample (mix) the model will generate in the end of every epoch.--n_critic: how many time the discriminator training over generator training.--cnn_type: the cnn type of encoder (e.g:res_cnnorcnn) .--loss_type: the loss function during the GAN training supportminmaxandleast squareloss.--batch_size: the batch size of dataloader during the GAN training.
To generate the mix by trained generator, you need to run the script in script/inference.py. We provide two tracks in test/ for your reference.
python script/inference.py [--g_path=(str, ex:'./pretrained/djtransgan_minmax.pt')] [--out_dir=(str, ex: 'results/inference')] [--prev_track=(str, ex: './test/Breikthru ft Danny Devinci-Touch.mp3')] [--next_track=(str, ex: './test/Jameson-Hangin.mp3')] [--prev_cue=(float, ex:96)] [--next_cue=(float, ex:30)] [--download=(bool, ex:1)]
--g_path: the path of trained generator, can be the pre-trained model provide by us or the model training by you.--out_dir: the output directory which is going to save the result.--prev_track: the path of the previous track (first track).--next_track: the path of the next track (second track).--prev_cue: the cue point of previous track (the point previous track totally fade out).--next_cue: the cue point of next track (the point next track totally fade in).--download: specify whether download the pre-trained model provided by us.
If you use any of our code in your work please consider citing us.
@inproceedings{chen2022djtransgan,
title={Automatic DJ Transitions with Differentiable Audio Effects and Generative Adversarial Networks},
author={Chen, B. Y., Hsu, W. H., Liao, W. H., Ramírez, M. A. M., Mitsufuji, Y., & Yang, Y. H.},
booktitle={ICASSP},
year={2022}}
This repo is done during the internship in the Sony Group Corporation with outstanding mentoring by my incredible mentors in Sony Wei-Hsiang Liao, Marco A. Martínez-Ramírez, and Yuki Mitsufuji and my colleague Wei-Han Hsu and advisor Yi-Hsuan Yang in Academia Sinica. The results are a joint effort with Sony Group Corporation and Academia Sinica. I sincerely appreciate all the support made by them to make this research happen. Moreover, please check the other excellent AI research made by Sony here and their recent work "FxNorm-automix" and "distortionremoval" which is going to present in ISMIR 2022.
Copyright © 2022 Bo-Yu Chen
Licensed under the MIT License (the "License"). You may not use this package except in compliance with the License. You may obtain a copy of the License at
https://opensource.org/licenses/MIT
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.