Learning Debiased Representation via Disentangled Feature Augmentation (Neurips 2021, Oral): Official Project Webpage
This repository provides the official PyTorch implementation of the following paper:
Learning Debiased Representation via Disentangled Feature Augmentation
Jungsoo Lee* (KAIST AI, Kakao Enterprise), Eungyeup Kim* (KAIST AI, Kakao Enterprise),
Juyoung Lee (Kakao Enterprise), Jihyeon Lee (KAIST AI), and Jaegul Choo (KAIST AI)
(* indicates equal contribution. The order of first authors was chosen by tossing a coin.)
NeurIPS 2021, Oral
Paper: Arxiv
Abstract: Image classification models tend to make decisions based on peripheral attributes of data items that have strong correlation with a target variable (i.e., dataset bias). These biased models suffer from the poor generalization capability when evaluated on unbiased datasets. Existing approaches for debiasing often identify and emphasize those samples with no such correlation (i.e., bias-conflicting) without defining the bias type in advance. However, such bias-conflicting samples are significantly scarce in biased datasets, limiting the debiasing capability of these approaches. This paper first presents an empirical analysis revealing that training with "diverse" bias-conflicting samples beyond a given training set is crucial for debiasing as well as the generalization capability. Based on this observation, we propose a novel feature-level data augmentation technique in order to synthesize diverse bias-conflicting samples. To this end, our method learns the disentangled representation of (1) the intrinsic attributes (i.e., those inherently defining a certain class) and (2) bias attributes (i.e., peripheral attributes causing the bias), from a large number of bias-aligned samples, the bias attributes of which have strong correlation with the target variable. Using the disentangled representation, we synthesize bias-conflicting samples that contain the diverse intrinsic attributes of bias-aligned samples by swapping their latent features. By utilizing these diversified bias-conflicting features during the training, our approach achieves superior classification accuracy and debiasing results against the existing baselines on both synthetic as well as a real-world dataset.
Code Contributors
Jungsoo Lee [Website] [LinkedIn] [Google Scholar] (KAIST AI, Kakao Enterprise)
Eungyeup Kim [Website] [LinkedIn] [Google Scholar] (KAIST AI, Kakao Enterprise)
Juyoung Lee [Website] (Kakao Enterprise)
Pytorch Implementation
Installation
Clone this repository.
git clone https://github.com/kakaoenterprise/Learning-Debiased-Disentangled.git
cd Learning-Debiased-Disentangled
pip install -r requirements.txt
Datasets
We used three datasets in our paper.
Download the datasets with the following url. Note that BFFHQ is the dataset used in "BiaSwap: Removing Dataset Bias with Bias-Tailored Swapping Augmentation" (Kim et al., ICCV 2021). Unzip the files and the directory structures will be as following:
cmnist
└ 0.5pct / 1pct / 2pct / 5pct
└ align
└ conlict
└ valid
└ test
cifar10c
└ 0.5pct / 1pct / 2pct / 5pct
└ align
└ conlict
└ valid
└ test
bffhq
└ 0.5pct
└ valid
└ test
How to Run
CMNIST
Vanilla
python train.py --dataset cmnist --exp=cmnist_0.5_vanilla --lr=0.01 --percent=0.5pct --train_vanilla --tensorboard --wandb
python train.py --dataset cmnist --exp=cmnist_1_vanilla --lr=0.01 --percent=1pct --train_vanilla --tensorboard --wandb
python train.py --dataset cmnist --exp=cmnist_2_vanilla --lr=0.01 --percent=2pct --train_vanilla --tensorboard --wandb
python train.py --dataset cmnist --exp=cmnist_5_vanilla --lr=0.01 --percent=5pct --train_vanilla --tensorboard --wandb
bash scripts/run_cmnist_vanilla.sh
Ours
python train.py --dataset cmnist --exp=cmnist_0.5_ours --lr=0.01 --percent=0.5pct --curr_step=10000 --lambda_swap=1 --lambda_dis_align=10 --lambda_swap_align=10 --use_lr_decay --lr_decay_step=10000 --lr_gamma=0.5 --train_ours --tensorboard --wandb
python train.py --dataset cmnist --exp=cmnist_1_ours --lr=0.01 --percent=1pct --curr_step=10000 --lambda_swap=1 --lambda_dis_align=10 --lambda_swap_align=10 --use_lr_decay --lr_decay_step=10000 --lr_gamma=0.5 --train_ours --tensorboard --wandb
python train.py --dataset cmnist --exp=cmnist_2_ours --lr=0.01 --percent=2pct --curr_step=10000 --lambda_swap=1 --lambda_dis_align=10 --lambda_swap_align=10 --use_lr_decay --lr_decay_step=10000 --lr_gamma=0.5 --train_ours --tensorboard --wandb
python train.py --dataset cmnist --exp=cmnist_5_ours --lr=0.01 --percent=5pct --curr_step=10000 --lambda_swap=1 --lambda_dis_align=10 --lambda_swap_align=10 --use_lr_decay --lr_decay_step=10000 --lr_gamma=0.5 --train_ours --tensorboard --wandb
bash scripts/run_cmnist_ours.sh
Corrupted CIFAR10
Vanilla
python train.py --dataset cifar10c --exp=cifar10c_0.5_vanilla --lr=0.001 --percent=0.5pct --train_vanilla --tensorboard --wandb
python train.py --dataset cifar10c --exp=cifar10c_1_vanilla --lr=0.001 --percent=1pct --train_vanilla --tensorboard --wandb
python train.py --dataset cifar10c --exp=cifar10c_2_vanilla --lr=0.001 --percent=2pct --train_vanilla --tensorboard --wandb
python train.py --dataset cifar10c --exp=cifar10c_5_vanilla --lr=0.001 --percent=5pct --train_vanilla --tensorboard --wandb
bash scripts/run_cifar10c_vanilla.sh
Ours
python train.py --dataset cifar10c --exp=cifar10c_0.5_ours --lr=0.0005 --percent=0.5pct --curr_step=10000 --lambda_swap=1 --lambda_dis_align=1 --lambda_swap_align=1 --use_lr_decay --lr_decay_step=10000 --lr_gamma=0.5 --train_ours --tensorboard --wandb
python train.py --dataset cifar10c --exp=cifar10c_1_ours --lr=0.001 --percent=1pct --curr_step=10000 --lambda_swap=1 --lambda_dis_align=5 --lambda_swap_align=5 --use_lr_decay --lr_decay_step=10000 --lr_gamma=0.5 --train_ours --tensorboard --wandb
python train.py --dataset cifar10c --exp=cifar10c_2_ours --lr=0.001 --percent=2pct --curr_step=10000 --lambda_swap=1 --lambda_dis_align=5 --lambda_swap_align=5 --use_lr_decay --lr_decay_step=10000 --lr_gamma=0.5 --train_ours --tensorboard --wandb
python train.py --dataset cifar10c --exp=cifar10c_5_ours --lr=0.001 --percent=5pct --curr_step=10000 --lambda_swap=1 --lambda_dis_align=1 --lambda_swap_align=1 --use_lr_decay --lr_decay_step=10000 --lr_gamma=0.5 --train_ours --tensorboard --wandb
bash scripts/run_cifar10c_ours.sh
BFFHQ
Vanilla
python train.py --dataset bffhq --exp=bffhq_0.5_vanilla --lr=0.0001 --percent=0.5pct --train_vanilla --tensorboard --wandb
bash scripts/run_bffhq_vanilla.sh
Ours
python train.py --dataset bffhq --exp=bffhq_0.5_ours --lr=0.0001 --percent=0.5pct --lambda_swap=0.1 --curr_step=10000 --use_lr_decay --lr_decay_step=10000 --lambda_dis_align 2. --lambda_swap_align 2. --dataset bffhq --train_ours --tensorboard --wandb
bash scripts/run_bffhq_ours.sh
Pretrained Models
In order to test our pretrained models, run the following command.
python test.py --pretrained_path=
--dataset=
--percent=
We provide the pretrained models in the following urls.
CMNIST 0.5pct
CMNIST 1pct
CMNIST 2pct
CMNIST 5pct
CIFAR10C 0.5pct
CIFAR10C 1pct
CIFAR10C 2pct
CIFAR10C 5pct
Citations
Bibtex coming soon!
Contact
Acknowledgments
This work was mainly done when both of the first authors were doing internship at Vision Team/AI Lab/Kakao Enterprise. Our pytorch implementation is based on LfF. Thanks for the implementation.
579 Dec 22, 2022
4 Jul 02, 2022
27 Dec 18, 2022
37 Dec 02, 2022
82 Dec 29, 2022
36 Nov 23, 2022
604 Dec 14, 2022
11 Oct 17, 2022
1 Jan 10, 2022
35 Nov 21, 2022
3 Dec 26, 2022
5 Aug 25, 2020
315 Jan 01, 2023
6 Oct 26, 2022
845 Nov 29, 2022
2 Aug 02, 2022
200 Jan 01, 2023
1.1k Dec 30, 2022
19 Nov 13, 2022
246 Dec 13, 2022