C-Flat

This repo contaions the official implementation of the NeurIPS2024 paper "Make Continual Learning Stronger via C-Flat" [paper].

🥳 Integration Notice 🥳

C-Flat has been fully integrated into the [INFTY] package, serving as a strong algorithm supporting generalizability within the continual learning ecosystem. For the latest version and extended functionalities, please refer to INFTY.

Acknowledgment

This repository is partially based on PyCIL and GAM.

Prerequisites

This code is implemented in PyTorch, and we have tested the code under the following environment settings:

• python = 3.9.13
• torch = 2.0.1
• torchvision = 0.15.2

Algorithm and results

Usage

Step 0: Datasets

We provide the source code on three benchmark datasets: CIFAR-100, ImageNet-100, and Tiny-ImageNet. For CIFAR-100, it will be downloaded automatically upon first use. For the other two datasets or self-made datasets, please modify the corresponding path in utils/data.py.

Step 1: Quick start

• Train CIFAR-100 on WA with C-Flat

python main.py --config=exps/wa.json --cflat

• Train CIFAR-100 on WA without C-Flat

python main.py --config=exps/wa.json

For other detailed model configurations, please refer to exps/[model_name].json.

Step 2: Custom application

2.1 Create a closure to calculate the loss

def create_loss_fn(self, inputs, targets):
    def loss_fn():
        logits = self._network(inputs)["logits"]
        loss_clf = F.cross_entropy(logits, targets)
        return logits, [loss_clf]
    return loss_fn

2.2 Introduce the cflat optimizer

optimizer = C_Flat(params=self._network.parameters(), base_optimizer=base_optimizer, model=self._network, cflat=True)

2.3 Model update

loss_fn = self.create_loss_fn(inputs, targets)
optimizer.set_closure(loss_fn)
logits, loss_list = optimizer.step()

Citation

If you find this repo useful for your research, please consider citing the paper.

@article{bian2024make,
  title={Make continual learning stronger via c-flat},
  author={Bian, Ang and Li, Wei and Yuan, Hangjie and Wang, Mang and Zhao, Zixiang and Lu, Aojun and Ji, Pengliang and Feng, Tao and others},
  journal={Advances in Neural Information Processing Systems},
  volume={37},
  pages={7608--7630},
  year={2024}
}

Contact

If there are any questions, please feel free to contact the corresponding author: Tao Feng (fengtao.hi@gmail.com).