Self-Damaging Contrastive Learning

Introduction

The recent breakthrough achieved by contrastive learning accelerates the pace for deploying unsupervised training on real-world data applications. However, unlabeled data in reality is commonly imbalanced and shows a long-tail distribution, and it is unclear how robustly the latest contrastive learning methods could perform in the practical scenario. This paper proposes to explicitly tackle this challenge, via a principled framework called Self-Damaging Contrastive Learning (SDCLR), to automatically balance the representation learning without knowing the classes. Our main inspiration is drawn from the recent finding that deep models have difficult-to-memorize samples, and those may be exposed through network pruning [1]. It is further natural to hypothesize that long-tail samples are also tougher for the model to learn well due to insufficient examples. Hence, the key innovation in SDCLR is to create a dynamic self-competitor model to contrast with the target model, which is a pruned version of the latter. During training, contrasting the two models will lead to adaptive online mining of the most easily forgotten samples for the current target model, and implicitly emphasize them more in the contrastive loss. Extensive experiments across multiple datasets and imbalance settings show that SDCLR significantly improves not only overall accuracies but also balancedness, in terms of linear evaluation on the full-shot and few-shot settings.

[1] Hooker, Sara, et al. "What Do Compressed Deep Neural Networks Forget?." arXiv preprint arXiv:1911.05248 (2019).

Method

The overview of the proposed SDCLR framework. Built on top of simCLR pipeline by default, the uniqueness of SDCLR lies in its two different network branches: one is the target model to be trained, and the other "self-competitor" model that is pruned from the former online. The two branches share weights for their non-pruned parameters. Either branch has its independent batch normalization layers. Since the self-competitor is always obtained and updated from the latest target model, the two branches will co-evolve during training. Their contrasting will implicitly give more weights on long-tail samples.

Environment

Requirements:

pytorch 1.7.1 
opencv-python
scikit-learn 
matplotlib

Recommend installation cmds (linux)

conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=10.2 -c pytorch # change cuda version according to hardware
pip install opencv-python
conda install -c conda-forge scikit-learn matplotlib

Details about and Imagenet-100-LT Imagenet-LT-exp

Imagenet-100-LT sampling list

Imagenet-LT-exp sampling list

Pretrained models downloading

CIFAR10: pretraining, fine-tuning

CIFAR100: pretraining, fine-tuning

Imagenet100/Imagenet: pretraining, fine-tuning

Train and evaluate pretrained models

Before all

chmod +x  cmds/shell_scrips/*

CIFAR10

SimCLR on balanced training datasets

# pre-train and finetune
for split_num in 1 2 3 4 5
do
./cmds/shell_scrips/cifar-10-LT.sh -g 1 -w 8 --split split${split_num}_D_b
done

# evaluate pretrained model (after download and unzip the pretrained model)
for split_num in 1 2 3 4 5
do
./cmds/shell_scrips/cifar-10-LT.sh -g 1 -w 8 --split split${split_num}_D_b  --only_finetuning True  --test_only True
done

# summery result (after "pre-train and finetune" or "evaluate pretrained model")
# linear separability
python exp_analyse.py --dataset cifar10
# few shot
python exp_analyse.py --dataset cifar10 --fewShot

SimCLR on long tail training datasets

# pre-train and finetune
for split_num in 1 2 3 4 5
do
./cmds/shell_scrips/cifar-10-LT.sh -g 1 -w 8 --split split${split_num}_D_i
done

# evaluate pretrained model (after download and unzip the pretrained model)
for split_num in 1 2 3 4 5 
do
./cmds/shell_scrips/cifar-10-LT.sh -g 1 -w 8 --split split${split_num}_D_i --only_finetuning True --test_only True
done

# summery result (after "pre-train and finetune" or "evaluate pretrained model")
# linear separability
python exp_analyse.py --dataset cifar10 --LT
# few shot
python exp_analyse.py --dataset cifar10 --LT --fewShot

SDCLR on long tail training datasets

# pre-train and finetune
for split_num in 1 2 3 4 5
do
./cmds/shell_scrips/cifar-10-LT.sh -g 1 -w 8 --split split${split_num}_D_i --prune True --prune_percent 0.9 --prune_dual_bn True
done

# evaluate pretrained model (after download and unzip the pretrained model)
for split_num in 1 2 3 4 5 
do
./cmds/shell_scrips/cifar-10-LT.sh -g 1 -w 8 --split split${split_num}_D_i --prune True --prune_percent 0.9 --prune_dual_bn True --only_finetuning True --test_only True
done

# summery result (after "pre-train and finetune" or "evaluate pretrained model")
# linear separability
python exp_analyse.py --dataset cifar10 --LT --prune
# few shot
python exp_analyse.py --dataset cifar10 --LT --prune --fewShot

CIFAR100

SimCLR on balanced training datasets

# pre-train and finetune
for split_num in 1 2 3 4 5
do
./cmds/shell_scrips/cifar-100-LT.sh -g 1 -p 4867 -w 8 --split cifar100_split${split_num}_D_b
done

# evaluate pretrained model (after download and unzip the pretrained model)
for split_num in 1 2 3 4 5
do
./cmds/shell_scrips/cifar-100-LT.sh -g 1 -p 4867 -w 8 --split cifar100_split${split_num}_D_b --only_finetuning True --test_only True
done

# summery result (after "pre-train and finetune" or "evaluate pretrained model")
# linear separability
python exp_analyse.py --dataset cifar100
# few shot
python exp_analyse.py --dataset cifar100 --fewShot

SimCLR on long tail training datasets

# pre-train and finetune
for split_num in 1 2 3 4 5
do
./cmds/shell_scrips/cifar-100-LT.sh -g 1 -p 4867 -w 8 --split cifar100_split${split_num}_D_i
done

# evaluate pretrained model (after download and unzip the pretrained model)
for split_num in 1 2 3 4 5
do
./cmds/shell_scrips/cifar-100-LT.sh -g 1 -p 4867 -w 8 --split cifar100_split${split_num}_D_i --only_finetuning True --test_only True
done

# summery result (after "pre-train and finetune" or "evaluate pretrained model")
# linear separability
python exp_analyse.py --dataset cifar100 --LT
# few shot
python exp_analyse.py --dataset cifar100 --LT --fewShot

SDCLR on long tail training datasets

# pre-train and finetune
for split_num in 1 2 3 4 5
do
./cmds/shell_scrips/cifar-100-LT.sh -g 1 -p 4867 -w 8 --split cifar100_split${split_num}_D_i --prune True --prune_percent 0.9 --prune_dual_bn True
done

# evaluate pretrained model (after download and unzip the pretrained model)
for split_num in 1 2 3 4 5
do
./cmds/shell_scrips/cifar-100-LT.sh -g 1 -p 4867 -w 8 --split cifar100_split${split_num}_D_i --prune True --prune_percent 0.9 --prune_dual_bn True --only_finetuning True --test_only True
done

# summery result (after "pre-train and finetune" or "evaluate pretrained model")
# linear separability
python exp_analyse.py --dataset cifar100 --LT --prune
# few shot
python exp_analyse.py --dataset cifar100 --LT --prune --fewShot

Imagenet-100-LT

SimCLR on balanced training datasets

# pre-train and finetune
./cmds/shell_scrips/imagenet-100-res50-LT.sh --data \path\to\imagenet -g 2 -p 4867 -w 10 --split imageNet_100_BL_train

# evaluate pretrained model (after download and unzip the pretrained model)
./cmds/shell_scrips/imagenet-100-res50-LT.sh --data \path\to\imagenet -g 2 -p 4867 -w 10 --split imageNet_100_BL_train --only_finetuning True --test_only True

# summery result (after "pre-train and finetune" or "evaluate pretrained model")
# linear separability
python exp_analyse.py --dataset imagenet100
# few shot
python exp_analyse.py --dataset imagenet100 --fewShot

SimCLR on long tail training datasets

# pre-train and finetune
./cmds/shell_scrips/imagenet-100-res50-LT.sh --data \path\to\imagenet -g 2 -p 4867 -w 10 --split imageNet_100_LT_train

# evaluate pretrained model (after download and unzip the pretrained model)
./cmds/shell_scrips/imagenet-100-res50-LT.sh --data \path\to\imagenet -g 2 -p 4860 -w 10 --split imageNet_100_LT_train --only_finetuning True --test_only True

# summery result (after "pre-train and finetune" or "evaluate pretrained model")
# linear separability
python exp_analyse.py --dataset imagenet100 --LT
# few shot
python exp_analyse.py --dataset imagenet100 --LT --fewShot

SDCLR on long tail training datasets

# pre-train and finetune
./cmds/shell_scrips/imagenet-100-res50-LT.sh --data \path\to\imagenet -g 2 -p 4867 -w 10 --split imageNet_100_LT_train --prune True --prune_percent 0.3 --prune_dual_bn True --temp 0.3

# evaluate pretrained model (after download and unzip the pretrained model)
./cmds/shell_scrips/imagenet-100-res50-LT.sh --data \path\to\imagenet -g 2 -p 4860 -w 10 --split imageNet_100_LT_train --prune True --prune_percent 0.3 --prune_dual_bn True --temp 0.3 --only_finetuning True --test_only True

# summery result (after "pre-train and finetune" or "evaluate pretrained model")
# linear separability
python exp_analyse.py --dataset imagenet100 --LT --prune
# few shot
python exp_analyse.py --dataset imagenet100 --LT --prune --fewShot

Imagenet-Exp-LT

SimCLR on balanced training datasets

# pre-train and finetune
./cmds/shell_scrips/imagenet-res50-LT.sh --data \path\to\imagenet -g 2 -p 4867 -w 10 --split imageNet_BL_exp_train

# evaluate pretrained model (after download and unzip the pretrained model)
./cmds/shell_scrips/imagenet-res50-LT.sh --data \path\to\imagenet -g 2 -p 4867 -w 10 --split imageNet_BL_exp_train --only_finetuning True --test_only True

# summery result (after "pre-train and finetune" or "evaluate pretrained model")
# linear separability
python exp_analyse.py --dataset imagenet
# few shot
python exp_analyse.py --dataset imagenet --fewShot

SimCLR on long tail training datasets

# pre-train and finetune
./cmds/shell_scrips/imagenet-res50-LT.sh --data \path\to\imagenet -g 2 -p 4867 -w 10 --split imageNet_LT_exp_train

# evaluate pretrained model (after download and unzip the pretrained model)
./cmds/shell_scrips/imagenet-res50-LT.sh --data \path\to\imagenet -g 2 -p 4868 -w 10 --split imageNet_LT_exp_train --only_finetuning True --test_only True

# summery result (after "pre-train and finetune" or "evaluate pretrained model")
# linear separability
python exp_analyse.py --dataset imagenet --LT
# few shot
python exp_analyse.py --dataset imagenet --LT --fewShot

Citation

@inproceedings{
jiang2021self,
title={Self-Damaging Contrastive Learning},
author={Jiang, Ziyu and Chen, Tianlong and Mortazavi, Bobak and Wang, Zhangyang},
booktitle={International Conference on Machine Learning},
year={2021}
}

[ICML 2021] “ Self-Damaging Contrastive Learning”, Ziyu Jiang, Tianlong Chen, Bobak Mortazavi, Zhangyang Wang

Related tags

Overview

Self-Damaging Contrastive Learning

Introduction

Method

Environment

Details about and Imagenet-100-LT Imagenet-LT-exp

Pretrained models downloading

Train and evaluate pretrained models

Before all

CIFAR10

CIFAR100

Imagenet-100-LT

Imagenet-Exp-LT

Citation

Owner

VITA

improvement of CLIP features over the traditional resnet features on the visual question answering, image captioning, navigation and visual entailment tasks.

Repo for "Benchmarking Robustness of 3D Point Cloud Recognition against Common Corruptions" https://arxiv.org/abs/2201.12296

Image Deblurring using Generative Adversarial Networks

Calibrated Hyperspectral Image Reconstruction via Graph-based Self-Tuning Network.

The official codes of our CVPR2022 paper: A Differentiable Two-stage Alignment Scheme for Burst Image Reconstruction with Large Shift

PyTorch implementation of MICCAI 2018 paper "Liver Lesion Detection from Weakly-labeled Multi-phase CT Volumes with a Grouped Single Shot MultiBox Detector"

SwinTrack: A Simple and Strong Baseline for Transformer Tracking

Code for "PVNet: Pixel-wise Voting Network for 6DoF Pose Estimation" CVPR 2019 oral

The Turing Change Point Detection Benchmark: An Extensive Benchmark Evaluation of Change Point Detection Algorithms on real-world data

Collection of generative models in Pytorch version.

Official Pytorch implementation for 2021 ICCV paper "Learning Motion Priors for 4D Human Body Capture in 3D Scenes" and trained models / data

Official implementation of Monocular Quasi-Dense 3D Object Tracking

Automates Machine Learning Pipeline with Feature Engineering and Hyper-Parameters Tuning :rocket:

Canonical Capsules: Unsupervised Capsules in Canonical Pose (NeurIPS 2021)

Pytorch implementation for our ICCV 2021 paper "TRAR: Routing the Attention Spans in Transformers for Visual Question Answering".

Covid-19 Test AI (Deep Learning - NNs) Software. Accuracy is the %96.5, loss is the 0.09 :)

🚀 An end-to-end ML applications using PyTorch, W&B, FastAPI, Docker, Streamlit and Heroku

Code accompanying the paper on "An Empirical Investigation of Domain Generalization with Empirical Risk Minimizers" published at NeurIPS, 2021

A modification of Daniel Russell's notebook merged with Katherine Crowson's hq-skip-net changes

AbelNN: Deep Learning Python module from scratch