This is the official PyTorch implementation of the following paper:
https://arxiv.org/abs/2306.09025
COVERHUNTER: COVER SONG IDENTIFY WITH REFINED ATTENTION AND ALIGNMENTS Feng Liu, Deyi Tuo, Yinan Xu, Xintong Han
Abstract: Cover song identification (CSI) focuses on finding the same music with different versions in reference anchors given a query track. In this paper, we propose a novel system named CoverHunter that overcomes the shortcomings of existing detection schemes by exploring richer features with refined attention and alignments. CoverHunter contains three key modules: 1) A convolution-augmented transformer (i.e., Conformer) structure that captures both local and global feature interactions in contrast to previous methods mainly relying on convolutional neural networks; 2) An attention-based time pooling module that further exploits the attention in the time dimension; 3) A novel coarse-to-fine training scheme that first trains a network to roughly align the song chunks and then refines the network by training on the aligned chunks. At the same time, we also summarize some important training tricks used in our system that help achieve better results. Experiments on several standard CSI datasets show that our method significantly improves over state-of-the-art methods with an embedding size of 128 (2.3% on SHS100K-TEST and 17.7% on DaTacos).
This year(2024) I went to compete on the MIREX CSI track because of being invited by the organizer. Due to time and energy constraints, I did not change my model and code, and used the previously trained model(which can be downloaded from google drive). Competition results can be found on mirex-2024. The final result is third place(almost the same as second place). Perhaps limited by the size of the training data(only open source data is used) and the algorithm, we are still far behind ByteCoverII. I am generally satisfied with this result.
In addition, I regret to say that in the past year, I have focused more on music generation. I am developing music generation algorithms like Suno and Mureka, so I may not update this project anymore. Thank you for your attention.
more code and model will be released soon.
- add main code
- add egs demo for covers80
- add code about course-to-fine training
- release model mentioned in paper
- train model again for popular song with chinese, with dataset released by TME.
Maybe CoverHunterMps is a better project, the author fixes some bugs of my code and make it run, add some new features and more detailed document. Thanks for the @alanngnet!
We take Covers80 as an example to show the whole process.
Before extracting features and training, we need to prepare our dataset as json-format. Every line of data file is a json-format contains information about wav file path, duration, speaker id and so on. An example is as below:
{"utt": "cover80_00000000_0_0", "wav": "data/covers80/wav_16k/annie_lennox+Medusa+03-A_Whiter_Shade_Of_Pale.wav", "dur_s": 316.728, "song": "A_Whiter_Shade_Of_Pale", "version": "annie_lennox+Medusa+03-A_Whiter_Shade_Of_Pale"}
{"utt": "cover80_00000000_0_1", "wav": ...}
...
Only audio with sample rate of 16k is enough for CSI task. And Covers80 dataset can be download from http://labrosa.ee.columbia.edu/projects/coversongs/covers80/, and make sure audio file is available in the dataset.
As mentioned in the paper, We use cqt feature extracted from signal. The script to extract features is as below, and it is worth noting that augmentation will be implement at this stage(very important for improving MAP).
python3 -m tools.extract_csi_features data/covers80/It is easy to run the train stage as:
python3 -m tools.train egs/covers80/
And our code supports run on multi gpus. Just use:
torchrun -m --nnodes=1 --nproc_per_node=2 tools.train egs/covers80/
For the convenience of viewing during training, we plot MAP at tensorboard after every epoch. A script is also offered to calculate MAP, top10, and rank1 with pre-trained model. The features of test-set data needs to be extracted first with tools.extract_csi_features. Note for test-set feature, we do not need augmentation, so the hparams file should be as one at data/covers80_testset.
After feature extracting, we can run as below:
python3 -m tools.eval_testset pretrain-model-dir query_path ref_path
Another choice is to use pre-trained model. Download pre-train model from https://drive.google.com/file/d/1rDZ9CDInpxQUvXRLv87mr-hfDfnV7Y-j/view. It is trained with SHS100k-train and Covers80 is not included in the train-set. After unzip it, you can run to eval Covers80 and get results shown like this:
2023-07-05 16:38:46,621 INFO Test, map:0.9266781356046699 rank1:3.0853658536585367
As stated in the paper, the CSI task can be better accomplished with additional alignment information. In order to keep the training code simple and readable, these alignment information is not contained in the dataloader code. However, a basic code is offered to explain our process of obtaining alignment information. And you can run the following code:
python3 -m tools/alignment_for_frames pretrained-model-dir, data-path, output-alignment-pathThe video of our oral presentation is here.
Hope that this project can help beginners to get started in this CSI field more quickly. If you have any questions, feel free to send me an email([email protected]) or ask in issue. Good luck!