This repo contains the source code for the following paper: Subspace Regularizers for Few-Shot Class Incremental Learning, Afra Feyza Akyürek, Ekin Akyürek, Derry Wijaya, Jacob Andreas. ICLR 2022. link
Few-shot class incremental learning (FSCIL) refers to the problem of updating a trained classifier to discriminate among an expanded set of classes with limited labeled data. We focus on the image classification problem. The two key challenges are catastrophic forgetting of old classes and overfitting to new classes due to small data. Subspace regularizarization while helps controlling forgetting, it remedies overfitting by regularizing the novel weights towards a feature space learned by old classes. Thanks to this simple fix, we are able to augment the last linear layer to accommodate more classes and continue fine-tuning the linear classifier on the new data as usual.
Please consider using sh setup.sh to create a conda environment and install required packages.
If running multi-session create and navigate to data/miniImageNet in this repo. Please download the file subspace-mini-data.tar.gz from this link, uncompress the contents under data/miniImageNet. You should have two files under data/miniImageNet: all.pickle and class_names.txt. You may consider using the command line utility gdrive for downloading files from Google Drive. This guide is also helpful for setting up gdrive. If using gdrive, you will need the fileIds, and they are usually visible in the shareable links.
Create a folder named dumped and cd into it. Download pretrained models for multi-session from the same link above (multi-session-resnet18.tar.gz), extract and place the contents under dumped/backbones/continual/resnet18. The numbered folder names refer to the seeds. Eventually you will have dumped/backbones/continual/resnet18/1, dumped/backbones/continual/resnet18/2, etc.
The repo currently supports only multi-session experiments.
Sample scripts are under scripts/continual for multi-session. In every script, at the beginning we provide slurm options for batch jobs. Feel free to ignore this part if you're not using Slurm. you will see a nested for loop which we used for hyperparameter tuning while current values indicate the best parameters. Each experiment is run for 10 different seeds. If you are interested in only a single run, scroll down to the bottom for the respective command in every .sh file. Make sure to set the desired seed between 1-10.
The existing commands use memory (+M setup in the paper). In order to turn off memory replay, simply remove the options --n_base_support_samples 1 and --memory_replay 1 which will set them to zero. When switching between +/-M, make sure to edit your EXP_FOLDER parameter in the sh scripts to avoid overriding your previous experiments.
If you would like to train the backbones from scratch, you may use scripts/continual/slurm_run_backbone.sh. After training your backbones, please refer to the above section Using pretrained models, making sure you point to your trained backbones in respective scripts.
All experiments were run on 32GB NVIDIA Tesla V100 nodes, however memory requirement is likely less.
Feel free to reach out to with questions, or open issues.
Afra Feyza Akyürek (akyurek@bu.edu)
Ekin Akyürek (akyurek@mit.edu)
A significant portion of this repository is based on RFS source code.
@inproceedings{akyurek2022subspace,
title={Subspace Regularizers for Few-Shot Class Incremental Learning},
author={Aky{\"u}rek, Afra Feyza and Aky{\"u}rek, Ekin and Wijaya, Derry and Andreas, Jacob},
booktitle={International Conference on Learning Representations},
year={2022}
}