Task-Agnostic Guided Feature Expansion for Class-Incremental Learning

This is the code repo for Task-Agnostic Guided Feature Expansion for Class-Incremental Learning (CVPR 2025).

TL;DR: We propose TagFex framework to address feature collision in Class-Incremental Learning. TagFex features in continuously capturing task-agnostic features through a separate unsupervised model. TagFex achieves superior performance by a large margin, comparing to other expansion-based methods training models from scratch.

[arxiv]

To Run the Code

Package Requirements

pytorch torchvision torchmetrics loguru tqdm

Single-node training (for small datasets like CIFAR100)

CUDA_VISIBLE_DEVICES={gpu_ids} python main.py train --exp-configs {exp_config_files}

Multi-node training (for large datasets like ImageNet)

./trainddp.sh {gpu_ids} --exp-configs {exp_config_files}

Specify cuda device available in {gpu_ids}. Specify config files in {exp_config_files}, this argument accepts multiple values, the loading order is the same as the argument order. (see load_configs function)

Example:

python main.py train --exp-configs configs/all_in_one/cifar100_10-10_tagfex_resnet18.yaml --log-dir ./logs/exp_cifar100_10-10
./trainddp.sh 0,1,2,3 --exp-configs configs/all_in_one/cifar100_10-10_tagfex_resnet18.yaml --log-dir ./logs/exp_cifar100_10-10

Other useful arguments:

--exp-name # name the experiment
--log-dir # the folder to output log files
--ckpt-dir # the folder to checkpoint
--output-file-prefix # the filename prefix of output file

For more arguments, see argument.py.

Acknowledgements

This repository is inspired by PyCIL.