MutexMatch4SSL

This repo is the official Pytorch implementation of our paper:

MutexMatch: Semi-Supervised Learning with Mutex-Based Consistency Regularization
Authors: Yue Duan, Lei Qi, Lei Wang, Luping Zhou and Yinghuan Shi

• 🔗 Quick links:

  • Code Download
  • [PDF/Abs-arXiv | PDF&Abs-Published]

• 📰 Latest news:

  • The final version is published in: IEEE Transactions on Neural Networks and Learning Systems (Volume: 35, Issue: 6, June 2024).
  • Our paper is accepted by IEEE Transactions on Neural Networks and Learning Systems (TNNLS) 📕📕.

• 📑 Related works:

  • 🆕 [LATEST] Interested in the cross-modal retrieval with noisy correspondence? 👉 Check out our ACMMM'24 paper PC2 [PDF-arXiv | Code].
  • [SSL] Interested in the SSL in fine-grained visual classification (SS-FGVC)? 👉 Check out our AAAI'24 paper SoC [PDF-arXiv | Code].
  • [SSL] Interested in more scenarios of SSL with mismatched distributions? 👉 Check out our ICCV'23 paper PRG [PDF-arXiv | Code].
  • [SSL] Interested in robust SSL with mismatched distributions or more applications of complementary label in SSL? 👉 Check out our ECCV'22 paper RDA [PDF-arXiv | Code].

• 🏆 Ranks in Papers With Code:

  • 
  • 
  • 
  • 
  • 
  • 
  • 
  • 

Introduction

The core issue in semi-supervised learning (SSL) lies in how to effectively leverage unlabeled data, whereas most existing methods tend to put a great emphasis on the utilization of high-confidence samples yet seldom fully explore the usage of low-confidence samples. In this article, we aim to utilize low-confidence samples in a novel way with our proposed mutex-based consistency regularization, namely MutexMatch. Specifically, the high-confidence samples are required to exactly predict "what it is" by the conventional true-positive classifier (TPC), while low-confidence samples are employed to achieve a simpler goal — to predict with ease "what it is not" by the true-negative classifier (TNC). In this sense, we not only mitigate the pseudo-labeling errors but also make full use of the low-confidence unlabeled data by the consistency of dissimilarity degree.

Requirements

• matplotlib==3.3.2
• numpy==1.19.2
• pandas==1.1.5
• Pillow==9.0.1
• torch==1.4.0+cu92
• torchvision==0.5.0+cu92

Training

Important Args

• --k : Control the intensity of consistency regularization on TNC. By default, $k$=--num_classes.
• --num_classes : Number of classes in your dataset.
• --num_labels : Amount of labeled data used.
• --net [wrn/resnet18/cnn13] : By default, Wide ResNet (WRN-28-2) is used for experiments. You can use --widen_factor 8 for WRN-28-8. We provide alternatives as follows: ResNet-18 and CNN-13.
• --dataset [cifar10/cifar100/svhn/stl10/miniimage/tinyimage] and --data_dir : Your dataset name and path. We support five datasets: CIFAR-10, CIFAR-100, SVHN, STL-10, mini-ImageNet and Tiny-ImageNet. When --dataset stl10, set --fold [0/1/2/3/4] and --num_labels [1000/5000].
• --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 ground-truth of unlabeled data.

Training with Single GPU

python train_mutex.py --rank 0 --gpu [0/1/...] @@@other args@@@

Training with Multi-GPUs

• Using DataParallel

python train_mutex.py --world-size 1 --rank 0 @@@other args@@@

• Using DistributedDataParallel and single node

python train_mutex.py --world-size 1 --rank 0 --multiprocessing-distributed @@@other args@@@

Examples of Running

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.

WideResNet

• CIFAR-10 with 40 labels | result of seed 1 (Acc/%): 94.91 | weight: here

python train_mutex.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  --gpu 0

• CIFAR-100 with $k$=60 and 200 labels | result of seed 1 (Acc/%): 43.84 | weight: here

python train_mutex.py --world-size 1 --rank 0 --lr_decay cos --k 60 --widen_factor 8 --seed 1 --num_eval_iter 1000 --overwrite --save_name cifar100 --dataset cifar100 --num_classes 100 --num_labels 200  --gpu 0

• SVHN with 40 labels | result of seed 1 (Acc/%): 97.24 | weight: here

python train_mutex.py --world-size 1 --rank 0 --lr_decay cos --seed 1 --num_eval_iter 1000 --overwrite --save_name svhn --dataset svhn --num_classes 10 --num_labels 40  --gpu 0

---

CNN-13

• CIFAR-10 with 1000 labels | result of seed 1 (Acc/%): 93.01 | weight: here

python train_mutex.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 1000 --net cnn13 --gpu 0

---

ResNet-18

• mini-ImageNet with 1000 labels | result of seed 1 (Acc/%): 47.90 | weight: here

python train_mutex.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 --net resnet18 --gpu 0

Resume Training and Evaluation

Resume

If you restart the training, please use --resume --load_path @checkpoint path@.

Evaluation

python eval_mutex.py --data_dir @dataset path@ --load_path @checkpoint path@ --dataset [cifar10/cifar100/svhn/stl10/miniimage/tinyimage] 

Use --net [resnet18/cnn13] for different backbones.

Results (e.g., CIFAR-10)

• $k$=--num_classes

seed	10 labels	20 labels	40 labels	80 labels
1	15.73	93.43	94.91	93.69
2	71.47	93.24	94.76	93.64
3	93.07	93.42	92.96	92.05
4	86.32	87.56	88.41	93.43
5	65.66	91.18	95.05	93.32
avg	66.45	91.77	93.22	93.23

• $k$=0.6*--num_classes

seed	10 labels	20 labels	40 labels	80 labels
1	65.01	93.85	94.50	94.77
2	19.99	92.95	94.01	94.67
3	70.46	92.86	94.74	94.83
4	68.89	86.63	94.95	95.43
5	63.23	94.84	92.84	95.30
avg	57.52	92.23	94.21	95.00

Citation

Please cite our paper if you find MutexMatch useful:

@article{duan2022mutexmatch,
  title={MutexMatch: Semi-supervised Learning with Mutex-based Consistency Regularization},
  author={Duan, Yue and Zhao, Zhen and Qi, Lei and Wang, Lei and Zhou, Luping and Shi, Yinghuan and Gao, Yang},
  journal={arXiv preprint arXiv:2203.14316},
  year={2022}
}

or

@article{9992211,
  author={Duan, Yue and Zhao, Zhen and Qi, Lei and Wang, Lei and Zhou, Luping and Shi, Yinghuan and Gao, Yang},
  journal={IEEE Transactions on Neural Networks and Learning Systems}, 
  title={MutexMatch: Semi-Supervised Learning With Mutex-Based Consistency Regularization}, 
  year={2024},
  volume={35},
  number={6},
  pages={8441-8455},
  keywords={Training;Predictive models;Semisupervised learning;Data models;Task analysis;Entropy;Labeling;Mutex-based consistency regularization;semi-supervised classification},
  doi={10.1109/TNNLS.2022.3228380}}

Acknowledgement

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