debatrix/upcl

Rethinking Class Incremental Learning from a Dynamic Imbalanced Learning Perspective

Rethinking Class Incremental Learning from a Dynamic Imbalanced Learning Perspective

---

Introduction

This code is based on PyCIL with modifications to details such as log output. All the original logs of the experiments mentioned in the paper are located in the logs/. The code for our proposed UPCL is located in models/upcl.py.

If you use our method or code in your research, please consider citing the paper as follows:

@article{wang2024rethinking,
  title={Rethinking Class-Incremental Learning from a Dynamic Imbalanced Learning Perspective},
  author={Wang, Leyuan and Xiang, Liuyu and Wang, Yunlong and Wu, Huijia and He, Zhaofeng},
  journal={arXiv preprint arXiv:2405.15157},
  year={2024}
}

Dependencies

1. torch 1.81
2. torchvision 0.6.0
3. tqdm
4. numpy
5. scipy

Run experiment

1. Edit the [MODEL NAME].json file for global settings.
2. Edit the hyper-parameters in the corresponding [MODEL NAME].py file (e.g., models/icarl.py).
3. Run: python main.py --config=./exps/[MODEL NAME].json
4. hyper-parameters

When using PyCIL, you can edit the global parameters and algorithm-specific hyper-parameter in the corresponding json file.

These parameters include:

• memory-size: The total exemplar number in the incremental learning process. Assuming there are $K$ classes at the current stage, the model will preserve $\left[\frac{memory_size}{K}\right]$ exemplar per class.
• init-cls: The number of classes in the first incremental stage. Since there are different settings in CIL with a different number of classes in the first stage, our framework enables different choices to define the initial stage.
• increment: The number of classes in each incremental stage $i$, $i$ > 1. By default, the number of classes per incremental stage is equivalent per stage.
• convnet-type: The backbone network for the incremental model.
• seed: The random seed adopted for shuffling the class order. According to the benchmark setting, it is set to 1993 by default.

Datasets

We have implemented the pre-processing of CIFAR100, imagenet100, Tinyimagenet, and imagenet1000. When training on CIFAR100, this framework will automatically download it. When training on imagenet100/1000 and Tinyimagenet, you should specify the folder of your dataset in utils/data.py.

    def download_data(self):
        assert 0,"You should specify the folder of your dataset"
        train_dir = '[DATA-PATH]/train/'
        test_dir = '[DATA-PATH]/val/'

upcl's Issues

It won't run with the default settings

感谢您对 CIL 领域的贡献！
当我下载upcl代码后，用默认设置运行代码，出现了下面的问题，您能帮忙看看吗？
2024-06-12 21:13:55,771 [trainer.py] => 'UPCL_iCaRL' object has no attribute 'k'
Traceback (most recent call last):
File "/home/ps/dl/shuai/UPCL-main/trainer.py", line 82, in _train
model.incremental_train(data_manager)
File "/home/ps/dl/shuai/UPCL-main/models/icarl.py", line 74, in incremental_train
self._train(self.train_loader, self.test_loader)
File "/home/ps/dl/shuai/UPCL-main/models/upcl.py", line 273, in _train
k = min(fea_mask.sum(0).min(), self.k)
AttributeError: 'UPCL_iCaRL' object has no attribute 'k'