ECCV22-FOSTER
ECCV22-FOSTER copied to clipboard
G-U-N
The official implementation for ECCV22 paper: "FOSTER: Feature Boosting and Compression for Class-Incremental Learning" in PyTorch.
FOSTER: Feature Boosting and Compression for Class-Incremental Learning
The code repository for "Feature Boosting and Compression for Class-Incremental Learning " [paper] (ECCV22) in PyTorch. If you use any content of this repo for your work, please cite the following bib entry:
@article{wang2022foster,
title={FOSTER: Feature Boosting and Compression for Class-Incremental Learning},
author={Wang, Fu-Yun and Zhou, Da-Wei and Ye, Han-Jia and Zhan, De-Chuan},
journal={arXiv preprint arXiv:2204.04662},
year={2022}
}
Feature Boosting and Compression for Class-Incremental Learning
The ability to learn new concepts continually is necessary in this ever-changing world. However, deep neural networks suffer from catastrophic forgetting when learning new categories. Many works have been proposed to alleviate this phenomenon, whereas most of them either fall into the stability-plasticity dilemma or take too much computation or storage overhead. Inspired by the gradient boosting algorithm to gradually fit the residuals between the target and the current approximation function, we propose a novel two-stage learning paradigm FOSTER, empowering the model to learn new categories adaptively.
Gradient Boosting. we propose a novel perspective from gradient boosting to analyze and achieve the goal of class-incremental learning. Gradient boosting methods use the additive model to gradually converge the ground-truth target model where the subsequent one fits the residuals between the target and the prior one.
Results
Experimental results show that our method achieves state-of-the-art performance.
Results on CIFAR-100
| Protocols | Reproduced Avg | Reported Avg |
|---|---|---|
| B0 5 steps | 73.88 | 72.54 |
| B0 10 steps | 73.10 | 72.90 |
| B0 20 steps | 70.59 | 70.65 |
| B50 5 steps | 71.08 | 70.10 |
| B50 10 steps | 68.61 | 67.95 |
| B50 25 steps | 64.95 | 63.83 |
| B50 50 steps | 59.96 | - |
We visualize the grad-CAM before and after feature boosting. As shown in the figure~(top-left), the freeze CNN only focuses on the head of the birds, ignoring the rest of their bodies, while the new CNN learns that the whole body is important for classification, which is consistent with our claim. Similarly, the middle and right figures show that the new CNN also discovers some essential but ignored patterns of the mailbox, the dog, and the tennis.
Please refer to our [paper] for detailed results.
Prerequisites
The following packages are required to run the scripts:
Training scripts
-
Train CIFAR-100
python main.py --config=./configs/cifar/b0inc10.json -
Train ImageNet-100
python main.py --config=./configs/foster-imagenet100.json -
Train FOSTER-RMM
python main.py --config=./configs/foster-rmm.json
Remember to change YOURDATAROOT into your own data root, or you will encounter errors.
Acknowledgment
We thank the following repos for providing helpful components/functions in our work.
Contact
If there are any questions, please feel free to contact with the author: Fu-Yun Wang ([email protected]). Enjoy the code.
G-U-N