BS-RoFormer

实现了Band Split Roformer，这是字节跳动AI实验室开发的音乐源分离最先进的注意力网络。他们大幅超越了之前的第一名。该技术在频率（因此是多频带）和时间上使用轴向注意力。他们还进行了实验，证明旋转位置编码比学习绝对位置带来了巨大的改进。

它还支持立体声训练和输出多个音轨。

如果你有兴趣在公开场合复现最先进的音乐源分离器，请加入

更新：这篇论文已被Roman复现，权重已在这里开源

更新2：用于这个凯蒂·佩里的混音!

更新3：Kimberley Jensen已在这里开源了一个经过人声训练的MelBand Roformer!

致谢

感谢StabilityAI和🤗 Huggingface的慷慨赞助，以及我的其他赞助商，让我能够独立地开源人工智能。
感谢Roee和Fabian-Robert分享他们的音频专业知识并修复音频超参数
感谢@chenht2010和Roman解决了默认频带分割超参数的问题！
感谢Max Prod报告了Mel-Band Roformer在立体声训练中的一个重大bug！
感谢Roman成功训练了模型并在这个仓库中开源了他的训练代码和权重！
感谢Christopher修复了Mel-Band Roformer中多个音轨的问题
感谢Iver Jordal发现默认的stft窗口函数不正确

安装

$ pip install BS-RoFormer

使用方法

import torch
from bs_roformer import BSRoformer

model = BSRoformer(
    dim = 512,
    depth = 12,
    time_transformer_depth = 1,
    freq_transformer_depth = 1
)

x = torch.randn(2, 352800)
target = torch.randn(2, 352800)

loss = model(x, target = target)
loss.backward()

# 经过大量训练后

out = model(x)

要使用最近一篇后续论文中提出的Mel-Band Roformer，只需导入MelBandRoformer即可

import torch
from bs_roformer import MelBandRoformer

model = MelBandRoformer(
    dim = 32,
    depth = 1,
    time_transformer_depth = 1,
    freq_transformer_depth = 1
)

x = torch.randn(2, 352800)
target = torch.randn(2, 352800)

loss = model(x, target = target) loss.backward()

经过大量训练后

out = model(x)


## 待办事项

- [x] 加入多尺度短时傅里叶变换损失
- [x] 确定`n_fft`应该是多少
- [x] 审查频带分割和掩模估计模块

## 引用

```bibtex
@inproceedings{Lu2023MusicSS,
    title   = {基于频带分割RoPE Transformer的音乐源分离},
    author  = {卢韦宗 and 王巨江 and 孔秋强 and 洪韵宁},
    year    = {2023},
    url     = {https://api.semanticscholar.org/CorpusID:261556702}
}

@inproceedings{Wang2023MelBandRF,
    title   = {用于音乐源分离的梅尔频带RoFormer},
    author  = {王巨江 and 卢韦宗 and Minz Won},
    year    = {2023},
    url     = {https://api.semanticscholar.org/CorpusID:263608675}
}

@misc{ho2019axial,
    title  = {多维Transformer中的轴向注意力},
    author = {Jonathan Ho and Nal Kalchbrenner and Dirk Weissenborn and Tim Salimans},
    year   = {2019},
    archivePrefix = {arXiv}
}

@misc{su2021roformer,
    title   = {RoFormer：具有旋转位置嵌入的增强型Transformer},
    author  = {苏剑林 and 陆宇 and 潘胜峰 and 文博 and 刘云峰},
    year    = {2021},
    eprint  = {2104.09864},
    archivePrefix = {arXiv},
    primaryClass = {cs.CL}
}

@inproceedings{dao2022flashattention,
    title   = {Flash{A}ttention：具有{IO}感知的快速且内存高效的精确注意力},
    author  = {Dao, Tri and Fu, Daniel Y. and Ermon, Stefano and Rudra, Atri and R{\'e}, Christopher},
    booktitle = {神经信息处理系统进展},
    year    = {2022}
}

@article{Bondarenko2023QuantizableTR,
    title   = {可量化的Transformer：通过帮助注意力头什么都不做来消除异常值},
    author  = {Yelysei Bondarenko and Markus Nagel and Tijmen Blankevoort},
    journal = {ArXiv},
    year    = {2023},
    volume  = {abs/2306.12929},
    url     = {https://api.semanticscholar.org/CorpusID:259224568}
}

@inproceedings{ElNouby2021XCiTCI,
    title   = {XCiT：交叉协方差图像Transformer},
    author  = {Alaaeldin El-Nouby and Hugo Touvron and Mathilde Caron and Piotr Bojanowski and Matthijs Douze and Armand Joulin and Ivan Laptev and Natalia Neverova and Gabriel Synnaeve and Jakob Verbeek and Herv{\'e} J{\'e}gou},
    booktitle = {神经信息处理系统},
    year    = {2021},
    url     = {https://api.semanticscholar.org/CorpusID:235458262}
}