PEFT_UTILS 是Parameter-Efficient Fine-Tuning (PEFT) 方法的一个工具库。目前收集了LoRA、GLoRA、SSF 和 RepAdapter等PEFT方法。

以下是一个使用repadapter加载和训练模型的简单示例：

0. 导入模块 可以从repadapter.py，或peft_utils包中导入需要的模块。

1. 插入PEFT层

   在模型中插入PEFT层。

   set_repadapter(model=model)

   若需要只训练特定线性层，可修改set_repadapter使用正则表示式匹配特定的name

   import re
   import torch.nn as nn
   def set_repadapter(model, pattern):
   # 编译正则表达式模式
   regex = re.compile(pattern)
   for name, module in model.named_modules():
   # 检查模块是否是线性层并且名称匹配正则表达式
   if isinstance(module, nn.Linear) and regex.match(name):

2. 设置模型参数的requires_grad属性

   根据需要, 设置模型参数的requires_grad属性来确定哪些参数需要训练。在使用不同方法的时候'SSF'需要随所使用的方法调整，

   trainable = []
   for n, p in model.named_parameters():
       if any([x in n for x in ['adapter']]):
           trainable.append(p)
           p.requires_grad = True
       else:
           p.requires_grad = False

3. 保存模型repadapter部分

   训练完成后，一般仅保存模型的repadapter参数，这样可以节省非常多的硬盘空间占用，是repadapter的优势之一。

   import os
   save_repadapter(os.path.join(output_dir,"final.pt"), model=model)

4. 加载模型repadapter部分

   若需要加载训练完成后保存的模型。model需要set_repadapter后再加载。

   load_repadapter(load_path, model=model)

5. 重参数化模型 merge_repadapter在模型训练后用于简化模型结构，减少模型大小和推理时间。 merge_repadapter传入模型，和repadapter的保存路径，会进行重参数化。

   merge_repadapter(model,load_path=None,has_loaded=False)

This project uses methods and code from multiple sources:

• The method "Towards Efficient Visual Adaption via Structural Re-parameterization" by Luo et al. is incorporated, with the implementation found at the RepAdapter GitHub repository. We are grateful to Luo, Gen and the co-authors for their contributions.

• Additionally, we leverage "One-for-All: Generalized LoRA for Parameter-Efficient Fine-tuning" by Arnav Chavan et al. The corresponding code is available at the GLoRA GitHub repository. Our thanks go to Arnav Chavan and his colleagues for making their work accessible.

If you use the methods from Luo et al. or Chavan et al. in your work, please consider citing their papers:

@article{luo2023towards,
  title={Towards Efficient Visual Adaption via Structural Re-parameterization},
  author={Luo, Gen and Huang, Minglang and Zhou, Yiyi  and Sun, Xiaoshuai and Jiang, Guangnan and Wang, Zhiyu and Ji, Rongrong},
  journal={arXiv preprint arXiv:2302.08106},
  year={2023}
}

@misc{chavan2023oneforall,
  title={One-for-All: Generalized LoRA for Parameter-Efficient Fine-tuning},
  author={Arnav Chavan and Zhuang Liu and Deepak Gupta and Eric Xing and Zhiqiang Shen},
  year={2023},
  eprint={2306.07967},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}