Skip to content
/ CASSLE Public
forked from hankook/AugSelf

Official implementation of "Augmentation-aware Self-supervised Learning with Conditioned Projector"

Notifications You must be signed in to change notification settings

gmum/CASSLE

 
 

Repository files navigation

Presented at NeurIPS 2023 Workshop: Self-Supervised Learning - Theory and Practice - check out the poster!

TL;DR: We condition the projector of self-supervised models with augmentation information and demonstrate that this improves their performance during transfer learning.

thumbnail

Overview

Self-supervised learning (SSL) is a powerful technique for learning robust representations from unlabeled data. By learning to remain invariant to applied data augmentations, methods such as SimCLR and MoCo are able to reach quality on par with supervised approaches. However, this invariance may be harmful to solving some downstream tasks which depend on traits affected by augmentations used during pretraining, such as color.

In this paper, we propose to foster sensitivity to such characteristics in the representation space by modifying the projector network, a common component of self-supervised architectures. Specifically, we supplement the projector with information about augmentations applied to images. In order for the projector to take advantage of this auxiliary conditioning when solving the SSL task, the feature extractor learns to preserve the augmentation information in its representations.

Our approach, coined Conditional Augmentation-aware Self-supervised Learning (CASSLE), is directly applicable to typical joint-embedding SSL methods regardless of their objective functions. Moreover, it does not require major changes in the network architecture or prior knowledge of downstream tasks. In addition to an analysis of sensitivity towards different data augmentations, we conduct a series of experiments, which show that CASSLE improves over various SSL methods, reaching state-of-the-art performance in multiple downstream tasks.

Checkpoints

We provide checkpoints for models pretrained with and without CASSLE in this Google Drive folder.

Dependencies

conda create -n CASSLE python=3.8 pytorch=1.7.1 torchvision=0.8.2 cudatoolkit=10.1 ignite -c pytorch
conda activate CASSLE
pip install scipy tensorboard kornia==0.4.1 sklearn wandb

Pretraining

We pretrain the ResNet-50 models on the ImageNet-100 subset of ImageNet. You can download ImageNet here, and create the ImageNet-100 subset using this repository.

We provide a command for pretraining MoCo-v2 + CASSLE. To train the Baseline model, replace the --aug-inj-type option to proj-none. To train AugSelf, use --aug-inj-type proj-none --ss-crop 0.5 --ss-color 0.5. For using other frameworks like SimCLR, use the --framework option.

python pretrain_cond.py  \
  --logdir $LOGDIR \
  --framework moco --dataset imagenet100 --datadir $IMGENET_100_FOLDER \
  --model resnet50 --batch-size 256 --max-epochs 500 \
  --aug-treatment mlp --aug-nn-depth 6 --aug-nn-width 64 --aug-inj-type proj-cat \
  --aug-cond crop color color_diff blur grayscale \
  --base-lr 0.03 --wd 1e-4 --ckpt-freq 50 --eval-freq 50 --num-workers 16 --seed 1 --distributed

Evaluation

Our main evaluation setups are linear evaluation on fine-grained classification datasets (Table 1).

Linear evaluation

CUDA_VISIBLE_DEVICES=0 python transfer_linear_eval.py \
    --pretrain-data imagenet100 \
    --ckpt CKPT \
    --model resnet50 \
    --dataset cifar10 \
    --datadir DATADIR \
    --metric top1

Acknowledgements

This codebase is built upon this code for the excellent Improving Transferability of Representations via Augmentation-Aware Self-Supervision by Hankook Lee, Kibok Lee, Kimin Lee, Honglak Lee, and Jinwoo Shin. We are very thankful to the authors for providing an excellent-quality reproducible code to the community and guiding us on running it.

Citation

If you find our work interesting, please cite it:

@misc{przewiezlikowski2023augmentationaware,
      title={Augmentation-aware Self-supervised Learning with Conditioned Projector}, 
      author={Marcin Przewięźlikowski and Mateusz Pyla and Bartosz Zieliński and Bartłomiej Twardowski and Jacek Tabor and Marek Śmieja},
      year={2023},
      eprint={2306.06082},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}