News 2: A new repository containing the code of this paper as well as its extensions to video representation learning is available here.

News 1: SCE has been extended to video and the paper is available here.

You should now use eztorch, this repo is deprecated.

This repository contains a Pytorch implementation of Similarity Contrastive Estimation for Self-Supervised Soft Contrastive Learning (SCE) that has been published in IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2023.

It relies on several packages to optimize training and simplify the code such as:

• Hydra for configuration.
• Pytorch-Lightning for simplifying the code.
• Kornia to apply augmentations on GPU and speed-up training.

The following results are the main ones reported in our paper:

You can find below the command lines to launch the configs to retrieve those results and above the checkpoint links.

To install this repository you first need to install a recent version of Pytorch (> 1.8).

Finally launch the following command:

cd sce
pip install -e .

The argument -e makes a dev installation that allows you to make changes in the repository without needing to install the package again. It is optional.

Warning If you want to manually install the dependencies, then, you should install Kornia from the Julien Denize's fork that aligned transformations of Kornia with expected results from Torchvision.

We launched our experiments on a computational cluster configured via SLURM using for two crops 8 A100-80G GPUs and for multi-crop 16 A100-80G GPUs.

We provide below the commands using the srun command from SLURM that was inside a SLURM script. Pytorch-Lightning directly detects SLURM is used and configure accordingly the distributed training. We strongly suggest you to refer to the Pytorch-Lightning's documentation to correctly set up a command line without srun if you do not have access to a slurm cluster.

output_dir=...
dataset_dir=...

config_path="../sce/configs/run/pretrain/sce/resnet50"
config_name="resnet50_imagenet"
seed=42

cd sce/run

srun --kill-on-bad-exit=1 python pretrain.py -cp $config_path -cn $config_name\
    dir.data=$dataset_dir dir.root=$output_dir \
    dir.exp='pretrain' seed.seed=$seed \
    datamodule.train.loader.num_workers=8 \
    datamodule.val.loader.num_workers=8 \
    trainer=gpu_ddp_sbn_amp trainer.gpus=8 \
    trainer.max_epochs=100 model.optimizer.initial_lr=0.6

srun --kill-on-bad-exit=1 python pretrain.py -cp $config_path -cn $config_name\
    dir.data=$dataset_dir dir.root=$output_dir \
    dir.exp='pretrain' seed.seed=$seed \
    datamodule.train.loader.num_workers=8 \
    datamodule.val.loader.num_workers=8 \
    trainer=gpu_ddp_sbn_amp trainer.gpus=8

srun --kill-on-bad-exit=1 python pretrain.py -cp $config_path -cn $config_name\
    dir.data=$dataset_dir dir.root=$output_dir \
    dir.exp='pretrain' seed.seed=$seed \
    datamodule.train.loader.num_workers=8 \
    datamodule.val.loader.num_workers=8 \
    trainer=gpu_ddp_sbn_amp trainer.gpus=8 \
    trainer.max_epochs=300

srun --kill-on-bad-exit=1 python pretrain.py -cp $config_path -cn $config_name\
    dir.data=$dataset_dir dir.root=$output_dir \
    dir.exp='pretrain' seed.seed=$seed \
    datamodule.train.loader.num_workers=8 \
    datamodule.val.loader.num_workers=8 \
    trainer=gpu_ddp_sbn_amp trainer.gpus=8 \
    trainer.max_epochs=1000 \
    model.optimizer.params.weight_decay=1.5e-6

config_name="resnet50_imagenet_multi_crop"

srun --kill-on-bad-exit=1 python pretrain.py -cp $config_path -cn $config_name \
    dir.data=$dataset_dir dir.root=$output_dir \
    dir.exp='pretrain' seed.seed=$seed \
    datamodule.train.loader.num_workers=8 \
    datamodule.val.loader.num_workers=8 \
    trainer=gpu_ddp_sbn_amp +trainer.num_nodes=2 \
    trainer.gpus=8

Same as for pretraining we launched our experiment on a SLURM cluster.

eval_config_path="../sce/configs/run/evaluation/linear_classifier/mocov3/resnet50"
eval_config_name="resnet50_imagenet"
checkpoint=...

srun --kill-on-bad-exit=1 python linear_classifier_evaluation.py \
    -cp $eval_config_path -cn $eval_config_name \
    dir.data=$dataset_dir dir.root=$output_dir \
    dir.exp='linear_classifier_evaluation' \
    model.pretrained_trunk_path=$checkpoint seed.seed=$seed \
    datamodule.train.loader.num_workers=8 \
    datamodule.val.loader.num_workers=8 trainer.gpus=8

For Transfer Learning evaluation we used code provided by several authors that we would like to thank for sharing their work. Below we redirect you to their GitHub for every transfer we have done.

All evaluations are based on the multi-crop checkpoint.

To evaluate the transferability of our pretrained checkpoints on various datasets by training a linear classifier, we used the ssl-transfer repository.

To evaluate the transferability of our pretrained checkpoints on PASCAL VOC by training an SVM classifier, we used the PCL repository.

To evaluate the transferability of our pretrained checkpoints on COCO by training a Mask R-CNN for object detection and mask segmentation, we used the triplet repository.

This work was made possible by the use of the Factory-AI supercomputer, financially supported by the Ile-de-France Regional Council.

CeCILL licence version 2.1.

See LICENSE for details.

If you found our work useful, please consider citing us:

@InProceedings{Denize_2023_WACV,
    author    = {Denize, Julien and Rabarisoa, Jaonary and Orcesi, Astrid and H\'erault, Romain and Canu, St\'ephane},
    title     = {Similarity Contrastive Estimation for Self-Supervised Soft Contrastive Learning},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2023},
    pages     = {2706-2716}
}