sungho-CoolG / FSCL

Fair Contrastive Learning

This paper is presented at CVPR 2022. The implementation is based on the source code of supervised contrastive learning. We are still clearing up it and will provide a more extended version including implementation of other comparable models. In this version, we cover implementation of SupCon and SimCLR on CelebA and

Requirements

To install requirements:

conda env update -n env_name --file requirement.yaml

Training and Evaluation

To train models (i.e., FSCL, FSCL+, FSCL*, SupCon, SimCLR) in this paper on CelebA dataset, run this command:

(1) Representation Learning

Fair Supervised Contrastive Loss (FSCL)

python3 main_contrastive.py  --cosine --dataset celeba --method FSCL \
--group_norm 0 --name FSCL_CelebA --target_attribute_1 3 \
--sensitive_attribute_1 21 --data_folder ./datasets/CelebA/

Fair Supervised Contrastive Loss with Group-wise Normalization (FSCL+)

python3 main_contrastive.py  --cosine --dataset celeba --method FSCL \
--group_norm 1 --name FSCL_GroupNorm_CelebA --target_attribute_1 3 \
--sensitive_attribute_1 21 --data_folder ./datasets/CelebA/

Fair Supervised Contrastive Loss without Target Attribute Labels (FSCL*)

python3 main_contrastive.py  --cosine --dataset celeba --method FSCL* \
--name FSCL_WithoutTarget_CelebA --target_attribute_1 3 \
--sensitive_attribute_1 21 --data_folder ./datasets/CelebA/

Supervised Contrastive Loss (SupCon)

python3 main_contrastive.py  --cosine --dataset celeba --method SupCon \
--name SupCon_CelebA --target_attribute_1 3 \
--sensitive_attribute_1 21 --data_folder ./datasets/CelebA/

Self-supervised Contrastive Loss (SimCLR)

python3 main_contrastive.py  --cosine --dataset celeba --method SimCLR \
--name SimCLR_CelebA --target_attribute_1 3 \
--sensitive_attribute_1 21 --data_folder ./datasets/CelebA/

You might select if you apply the group-wise normalization by --group_norm . Target attributes (--target_attribute_1 -- target_attribute_2) and sensitive attributes (--sensitive_attribute_1 --sensitive_attribute_2) are set by the number of attributes (e.g., 3: attractiveness and 21: male) in CelebA dataset or by the name of attributes (e.g., gender, age, ethnicity) in UTK Face dataset.

(2) Classifier Training and Evaluation

python3 main_classifier.py --cosine --dataset celeba --target_attribute_1 3 --sensitive_attribute_1 21 --ckpt model.pth  --data_folder ./datasets/CelebA/

Results

**Image Classification on CelebA ** (TA: attractiveness SA: male)

Model	Supervision	Accuracy(%)	Equalized Odds(%)
FSCL	Supervised	79.1	11.5
FSCL+	Supervised	79.1	6.5
SupCon	Supervised	80.5	30.5
FSCL*	Semi-supervised	74.6	14.8
SimCLR	Self-supervised	75.7	29.4

**Image Classification on UTK FACE ** (TA: gender SA: ethnicity)

Data imbalance ($\alpha$)=4

Model	Supervision	Accuracy(%)	Equalized Odds(%)
FSCL	Supervised	90.1	2.7
FSCL+	Supervised	90.1	1.6
SupCon	Supervised	89.8	10.6

Data imbalance ($\alpha$)=3

Model	Supervision	Accuracy(%)	Equalized Odds(%)
FSCL	Supervised	92.3	1.7
FSCL+	Supervised	92.2	1.0
SupCon	Supervised	91.6	8.4

Data imbalance ($\alpha$)=2

Model	Supervision	Accuracy(%)	Equalized Odds(%)
FSCL	Supervised	91.6	1.0
FSCL+	Supervised	91.5	0.6
SupCon	Supervised	92.0	4.5

Contributing

This repo is licensed under the terms of the MIT license.