Skip to content

"RDA: Reciprocal Distribution Alignment for Robust Semi-supervised Learning" by Yue Duan (ECCV 2022)

License

Notifications You must be signed in to change notification settings

NJUyued/RDA4RobustSSL

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

85 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

RDA4RobustSSL

This repo is the official Pytorch implementation of our paper:

RDA: Reciprocal Distribution Alignment for Robust Semi-supervised Learning
Authors: Yue Duan, Lei Qi, Lei Wang, Luping Zhou and Yinghuan Shi

  • đź”— Quick links: [PDF/Abs-arXiv | PDF/Abs-Published | Poster/Slides/Video | Code Download]
  • đź“° Latest news:
    • Slides and video explaining our paper are now online!
    • Our paper is accepted by European Conference on Computer Vision (ECCV) 2022 🎉🎉. Thanks to users.
  • đź“‘ Related works:
    • đź“Ť [MOST RELEVANT] Interested in more scenarios of SSL with mismatched distributions? 👉 Check out our ICCV'23 paper PRG [PDF-arXiv | Code].
    • 🆕 [LATEST] Interested in the SSL in fine-grained visual classification (SS-FGVC)? 👉 Check out our AAAI'24 paper SoC [PDF-arXiv | Code].
    • Interested in the conventional SSL or more application of complementary label in SSL? 👉 Check out our TNNLS paper MutexMatch [PDF-arXiv | Code].

Introduction

Reciprocal Distribution Alignment (RDA) is a semi-supervised learning (SSL) framework working with both the matched (conventionally) and the mismatched class distributions. Distribution mismatch is an often overlooked but more general SSL scenario where the labeled and the unlabeled data do not fall into the identical class distribution. This may lead to the model not exploiting the labeled data reliably and drastically degrade the performance of SSL methods, which could not be rescued by the traditional distribution alignment. RDA achieves promising performance in SSL under a variety of scenarios of mismatched distributions, as well as the conventional matched SSL setting.

Requirements

  • numpy==1.19.2
  • pandas==1.1.5
  • Pillow==9.0.1
  • torch==1.4.0+cu92
  • torchvision==0.5.0+cu92

How to Train

Important Args

  • --num_labels : Amount of labeled data used.
  • --mismatch [none/rda/darp/darp_reversed] : Select the type of mismatched distributions. none means the conventional balanced setting. rda means our protocol for constructing dataset with mismatched distributions, which is described in Sec. 4.2; darp means DARP's protocol described in Sec. C.1 of Supplementary Materials; darp_reversed means DARP's protocol for CIFAR-10 with reversed version of mismatched distributions.
  • --n0 : When --mismatch rda, this arg means the imbalanced ratio $N_0$ for labeled data; When --mismatch [darp/darp_reversed], this arg means the imbalanced ratio $\gamma_l$ for labeled data.
  • --gamma : When --mismatch rda, this arg means the imbalanced ratio $\gamma$ for unlabeled data; When --mismatch DARP/DARP_reversed, this arg means the imbalanced ratio $\gamma_u$ for unlabeled data.
  • --net : By default, Wide ResNet (WRN-28-2) are used for experiments. If you want to use other backbones for tarining, set --net [resnet18/preresnet/cnn13]. We provide alternatives as follows: ResNet-18, PreAct ResNet and CNN-13.
  • --dataset [cifar10/cifar100/stl10/miniimage] and --data_dir : Your dataset name and path. We support four datasets: CIFAR-10, CIFAR-100, STL-10 and mini-ImageNet. When --dataset stl10, set --fold [0/1/2/3/4].
  • --num_eval_iter : After how many iterations, we evaluate the model. Note that although we show the accuracy of pseudo-labels on unlabeled data in the evaluation, this is only to show the training process. We did not use any information about labels for unlabeled data in the training. Additionally, when you train model on STL-10, the pseudo-label accuracy will not be displayed normally, because we don't have the ground-truth of unlabeled data.

Training with Single GPU

To better reproduce our experimental results, it is recommended to follow our experimental environment using a single GPU for training.

python train_rda.py --world-size 1 --rank 0 --gpu [0/1/...] @@@other args@@@

Training with Multi-GPUs

  • Using DataParallel
python train_rda.py --world-size 1 --rank 0 @@@other args@@@
  • Using DistributedDataParallel with single node
python train_rda.py --world-size 1 --rank 0 --multiprocessing-distributed @@@other args@@@

Examples of Running

By default, the model and dist&index.txt will be saved in \--save_dir\--save_name. The file dist&index.txt will display detailed settings of mismatched distributions. This code assumes 1 epoch of training, but the number of iterations is 2**20. For CIFAR-100, you need set --widen_factor 8 for WRN-28-8 whereas WRN-28-2 is used for CIFAR-10. Note that you need set --net resnet18 for STL-10 and mini-ImageNet. Additionally, WRN-28-2 is used for all experiments under DARP's protocol.

Conventional Setting

Matched and balanced $C_x$, $C_u$ for Tab. 1 in Sec. 5.1

  • CIFAR-10 with 20 labels | Result of seed 1 (Acc/%): 93.40 | Weight: here
python train_rda.py --world-size 1 --rank 0 --lr_decay cos --seed 1 --num_eval_iter 1000 --overwrite --save_name cifar10 --dataset cifar10 --num_classes 10 --num_labels 20  --gpu 0

Mismatched Distributions

Imbalanced $C_x$ and balanced $C_u$ for Tab. 2 in Sec. 5.2

  • CIFAR-10 with 40 labels and $N_0=10$ | Result of seed 1 (Acc/%): 93.06 | Weight: here
python train_rda.py --world-size 1 --rank 0 --lr_decay cos --seed 1 --num_eval_iter 1000 --overwrite --save_name cifar10 --dataset cifar10 --num_classes 10 --num_labels 40 --mismatch rda --n0 10 --gpu 0
  • CIFAR-100 with 400 labels and $N_0=40$ | Result of seed 1 (Acc/%): 33.54 | Weight: here
python train_rda.py --world-size 1 --rank 0 --lr_decay cos --seed 1 --num_eval_iter 1000 --overwrite --save_name cifar100 --dataset cifar100 --num_classes 100 --num_labels 400 --mismatch rda --n0 40 --gpu 0 --widen_factor 8
  • mini-ImageNet with 1000 labels and $N_0=40$ | Result of seed 1 (Acc/%): 43.59 | Weight: here
python train_rda.py --world-size 1 --rank 0 --lr_decay cos --seed 1 --num_eval_iter 1000 --overwrite --save_name miniimage --dataset miniimage --num_classes 100 --num_labels 1000 --mismatch rda --n0 40 --gpu 0 --net resnet18 

Imbalanced and mismatched $C_x$, $C_u$ for Tab. 3 in Sec. 5.2

  • CIFAR-10 with 40 labels, $N_0=10$ and $\gamma=5$ | Result of seed 1 (Acc/%): 80.68 | Weight: here
python train_rda.py --world-size 1 --rank 0 --lr_decay cos --seed 1 --num_eval_iter 1000 --overwrite --save_name cifar10 --dataset cifar10 --num_classes 10 --num_labels 40 --mismatch rda --n0 10 --gamma 5 --gpu 0

Balanced $C_x$ and imbalanced $C_u$ for Tab. 5 in Sec. 5.2

  • CIFAR-10 with 40 labels and $\gamma=200$ | Result of seed 1 (Acc/%): 45.57 | Weight: here
python train_rda.py --world-size 1 --rank 0 --lr_decay cos --seed 1 --num_eval_iter 1000 --overwrite --save_name cifar10 --dataset cifar10 --num_classes 10 --num_labels 40 --mismatch rda --gamma 200 --gpu 0

DARP's protocol for Tab. 5 in Sec. 5.2

  • CIFAR-10 with $\gamma_l=100$ and $\gamma_u=1$ | Result of seed 1 (Acc/%): 93.11 | Weight: here
python train_rda.py --world-size 1 --rank 0 --lr_decay cos --seed 1 --num_eval_iter 1000 --overwrite --save_name cifar10 --dataset cifar10 --num_classes 10 --mismatch darp --n0 100 --gamma 1 --gpu 0
  • CIFAR-10 with $\gamma_l=100$ and $\gamma_u=100$ (reversed) | Result of seed 1 (Acc/%): 78.53 | Weight: here
python train_rda.py --world-size 1 --rank 0 --lr_decay cos --seed 1 --num_eval_iter 1000 --overwrite --save_name cifar10 --dataset cifar10 --num_classes 10 --mismatch darp_reversed --n0 100 --gamma 100 --gpu 0
  • For STL-10 in DARP's protocol, set --fold -1.

    STL-10 with $\gamma_l=10$ | Result of seed 1 (Acc/%): 87.21 | Weight: here

python train_rda.py --world-size 1 --rank 0 --lr_decay cos --seed 1 --num_eval_iter 1000 --overwrite --save_name stl10 --dataset stl10 --num_classes 10 --mismatch darp --n0 10 --gpu 0 --fold -1 

Resume Training and Evaluation

If you restart the training, please use --resume --load_path @your_path_to_checkpoint. Each time you start training, the evaluation results of the current model will be displayed. If you want to evaluate a model, use its checkpoints to resume training.

Results (e.g. seed=1)

Dateset Labels $N_0$ / $\gamma_l$ $\gamma$ / $\gamma_u$ Acc (%) Note
CIFAR-10 20 - - 93.40 Conventional setting
40 - - 94.13
80 - - 94.24
100 - - 94.66
40 10 - 93.06 Imbalanced $C_x$ and balanced $C_u$
40 20 - 81.51
100 40 - 94.42
100 80 - 78..99
40 10 2 81.60 Mismatched imbalanced $C_x$ and $C_u$
40 10 5 80.68
100 40 5 79.54
40 - 100 47.68 Balanced $C_x$ and imbalanced $C_u$
40 - 200 45.57
DARP 100 1 93.11 DARP's protocol
DARP 100 50 79.84
DARP 100 150 74.71
DARP (reversed) 100 100 78.53
CIFAR-100 400 40 - 33.54 Imbalanced $C_x$ and balanced $C_u$
1000 80 - 42.87
STL-10 1000 - - 82.53 Conventional setting
DARP 10 - 87.21 DARP's protocol
DARP 20 - 83.71
mini-ImageNet 1000 - - 47.73 Conventional setting
1000 40 - 43.59 Imbalanced $C_x$ and balanced $C_u$
1000 80 - 38.16
1000 40 10 25.91 Mismatched imbalanced $C_x$ and $C_u$

Citation

Please cite our paper if you find RDA useful:

@inproceedings{duan2022rda,
  title={RDA: Reciprocal Distribution Alignment for Robust Semi-supervised Learning},
  author={Duan, Yue and Qi, Lei and Wang, Lei and Zhou, Luping and Shi, Yinghuan},
  booktitle={European Conference on Computer Vision},
  pages={533--549},
  year={2022},
  organization={Springer}
}

or

@article{duan2022rda,
  title={RDA: Reciprocal Distribution Alignment for Robust Semi-supervised Learning},
  author={Duan, Yue and Qi, Lei and Wang, Lei and Zhou, Luping and Shi, Yinghuan},
  journal={arXiv preprint arXiv:2208.04619},
  year={2022}
}

Acknowledgement

Our code is based on open source code: LeeDoYup/FixMatch-pytorch.