/TorchSSL

A PyTorch-based library for semi-supervised learning (NeurIPS'21)

Primary LanguagePythonMIT LicenseMIT

TorchSSL is deprecated and no longer maintained. Please refer USB, an upgrade version of TorchSSL. Training in USB only takes 12.5% of training time of using TorchSSL, and produce better results.

TorchSSL

A Pytorch-based toolbox for semi-supervised learning. This is also the official implementation for FlexMatch: boosting semi-supervised learning using curriculum pseudo labeling published at NeurIPS 2021. [arXiv] [Zhihu article] [Video]

News and Updates

17/08/2022

  • TorchSSL (this repo) is no longer maintained and updated. We have created/updated a more comprehensive codebase and benchmark for Semi-Supervised Learning - USB. It is built upon TorchSSL but more flexible to use and more extensiable, containing datasets spanning Computer Vision, Natural Language Processing, and Audio Processing.

15/02/2021

  • The logs and model weights are shared! We notice that some model weights are missing. We will try to upload the missing model weights in the future.
  • The results of BestAcc have been updated! I use single P100 for CIFAR-10 and SVHN, single P40 for STL-10, signle V100-32G for CIFAR-100.

Introduction

TorchSSL is an all-in-one toolkit based on PyTorch for semi-supervised learning (SSL). Currently, we implmented 9 popular SSL algorithms to enable fair comparison and boost the development of SSL algorithms.

Supported algorithms: In addition to fully-supervised (as a baseline), TorchSSL supports the following popular algorithms:

  1. PiModel (NeurIPS 2015) [1]
  2. MeanTeacher (NeurIPS 2017) [2]
  3. PseudoLabel (ICML 2013) [3]
  4. VAT (Virtual adversarial training, TPAMI 2018) [4]
  5. MixMatch (NeurIPS 2019) [5]
  6. UDA (Unsupervised data augmentation, NeurIPS 2020) [6]
  7. ReMixMatch (ICLR 2019) [7]
  8. FixMatch (NeurIPS 2020) [8]
  9. FlexMatch (NeurIPS 2021) [9]

Besides, we implement our Curriculum Pseudo Labeling (CPL) method for Pseudo-Label (Flex-Pseudo-Label) and UDA (Flex-UDA).

Supported datasets: TorchSSL currently supports 5 popular datasets in SSL research:

  1. CIFAR-10
  2. CIFAR-100
  3. STL-10
  4. SVHN
  5. ImageNet

Main Results

The results are best accuracies with standard errors. In the results, "40", "250", "1000" etc. under the dataset row denote different numbers of labeled samples (e.g., "40" in CIFAR-10 means that there are only 4 labeled samples for each class). We use random seed 0,1,2 for all experiments. All configs are included under the config/ folder. You can directly cite these results in your own research.

Note FullySupervised results are from training the model using all training data in the dataset, regardless of the label amount denoted in the table.

CIFAR-10 and CIFAR-100

CIFAR-10 CIFAR100
40 250 4000 400 2500 10000
FullySupervised 95.38±0.05 95.39±0.04 95.38±0.05 80.7±0.09 80.7±0.09 80.73±0.05
PiModel [1] 25.66±1.76 53.76±1.29 86.87±0.59 13.04±0.8 41.2±0.66 63.35±0.0
PseudoLabel [3] 25.39±0.26 53.51±2.2 84.92±0.19 12.55±0.85 42.26±0.28 63.45±0.24
PseudoLabel_Flex [9] 26.26±1.96 53.86±1.81 85.25±0.19 14.28±0.46 43.88±0.51 64.4±0.15
MeanTeacher [2] 29.91±1.6 62.54±3.3 91.9±0.21 18.89±1.44 54.83±1.06 68.25±0.23
VAT [4] 25.34±2.12 58.97±1.79 89.49±0.12 14.8±1.4 53.16±0.79 67.86±0.19
MixMatch [5] 63.81±6.48 86.37±0.59 93.34±0.26 32.41±0.66 60.24±0.48 72.22±0.29
ReMixMatch [7] 90.12±1.03 93.7±0.05 95.16±0.01 57.25±1.05 73.97±0.35 79.98±0.27
UDA [6] 89.38±3.75 94.84±0.06 95.71±0.07 53.61±1.59 72.27±0.21 77.51±0.23
UDA_Flex [9] 94.56±0.52 94.98±0.07 95.76±0.06 54.83±1.88 72.92±0.15 78.09±0.1
FixMatch [8] 92.53±0.28 95.14±0.05 95.79±0.08 53.58±0.82 71.97±0.16 77.8±0.12
FlexMatch [9] 95.03±0.06 95.02±0.09 95.81±0.01 60.06±1.62 73.51±0.2 78.1±0.15

STL-10 and SVHN

STL-10 SVHN
40 250 1000 40 250 1000
FullySupervised None None None 97.87±0.02 97.87±0.01 97.86±0.01
PiModel [1] 25.69±0.85 44.87±1.5 67.22±0.4 32.52±0.95 86.7±1.12 92.84±0.11
PseudoLabel [3] 25.32±0.99 44.55±2.43 67.36±0.71 35.39±5.6 84.41±0.95 90.6±0.32
PseudoLabel_Flex [9] 26.58±2.19 47.94±2.5 67.95±0.37 36.79±3.64 79.58±2.11 87.95±0.54
MeanTeacher [2] 28.28±1.45 43.51±2.75 66.1±1.37 63.91±3.98 96.55±0.03 96.73±0.05
VAT [4] 25.26±0.38 43.58±1.97 62.05±1.12 25.25±3.38 95.67±0.12 95.89±0.2
MixMatch [5] 45.07±0.96 65.48±0.32 78.3±0.68 69.4±8.39 95.44±0.32 96.31±0.37
ReMixMatch [7] 67.88±6.24 87.51±1.28 93.26±0.14 75.96±9.13 93.64±0.22 94.84±0.31
UDA [6] 62.58±8.44 90.28±1.15 93.36±0.17 94.88±4.27 98.08±0.05 98.11±0.01
UDA_Flex [9] 70.47±2.1 90.97±0.45 93.9±0.25 96.58±1.51 97.34±0.83 97.98±0.05
FixMatch [8] 64.03±4.14 90.19±1.04 93.75±0.33 96.19±1.18 97.98±0.02 98.04±0.03
FlexMatch [9] 70.85±4.16 91.77±0.39 94.23±0.18 91.81±3.2 93.41±2.29 93.28±0.3

ImageNet

100k labels
top-1 top-5
FixMatch [8] 56.34 78.20
FlexMatch [9] 58.15 80.52

Logs and weights

You can download the shared logs and weights here.

https://1drv.ms/u/s!AlpW9hcyb0KvmyCfsCjGvhDXG5Nb?e=Xc6amH

Usage

Before running or modifing the code, you need to:

  1. Clone this repo to your machine.
  2. Make sure Anaconda or Miniconda is installed.
  3. Run conda env create -f environment.yml for environment initialization.

Run the experiments

It is convenient to perform experiment with TorchSSL. For example, if you want to run FlexMatch algorithm:

  1. Modify the config file in config/flexmatch/flexmatch.yaml as you need
  2. Run python flexmatch.py --c config/flexmatch/flexmatch.yaml

Customization

If you want to write your own algorithm, please follow the following steps:

  1. Create a directory for your algorithm, e.g., SSL, write your own model file SSl/SSL.py in it.
  2. Write the training file in SSL.py
  3. Write the config file in config/SSL/SSL.yaml

Citing TorchSSL

If you think this toolkit or the results are helpful to you and your research, please cite our paper:

@article{zhang2021flexmatch},
  title={FlexMatch: Boosting Semi-supervised Learning with Curriculum Pseudo Labeling},
  author={Zhang, Bowen and Wang, Yidong and Hou, Wenxin and Wu, Hao and Wang, Jindong and Okumura, Manabu and Shinozaki, Takahiro},
  booktitle={Neural Information Processing Systems (NeurIPS)},
  year={2021}
}

Maintainers

Yidong Wang1, Hao Chen2, Yue Fan3, Hao Wu1, Bowen Zhang1, Wenxin Hou1,4, Yuhao Chen5, Jindong Wang4

Tokyo Institute of Technology1

Carnegie Mellon University2

Max-Planck-Institut für Informatik3

Microsoft Research Asia4

Megvii5

Contributing

  1. You are welcome to open an issue on bugs, questions, and suggestions.
  2. If you want to join TorchSSL team, please e-mail Yidong Wang (646842131@qq.com; yidongwang37@gmail.com) for more information. We plan to add more SSL algorithms and expand TorchSSL from CV to NLP and Speech.

Statements

For ImageNet datasets: Please download the ImageNet 2014 dataset (unchanged from 2012) from the official site (link: https://image-net.org/challenges/LSVRC/2012/2012-downloads.php) Extract the train and the val set into subfolders (the test set is not used), and put them under train/ and val/ respectively. Each subfolder will represent a class. Note: the offical val set is not zipped into subfolders by classes, you may want to use: https://github.com/jiweibo/ImageNet/blob/master/valprep.sh, which is a nice script for preparing the file structure.

References

[1] Antti Rasmus, Harri Valpola, Mikko Honkala, Mathias Berglund, and Tapani Raiko. Semi-supervised learning with ladder networks. InNeurIPS, pages 3546–3554, 2015.

[2] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averagedconsistency targets improve semi-supervised deep learning results. InNeurIPS, pages 1195–1204, 2017.

[3] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning methodfor deep neural networks. InWorkshop on challenges in representation learning, ICML,volume 3, 2013.

[4] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training:a regularization method for supervised and semi-supervised learning.IEEE TPAMI, 41(8):1979–1993, 2018.

[5] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and ColinRaffel. Mixmatch: A holistic approach to semi-supervised learning.NeurIPS, page 5050–5060,2019.

[6] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmen-tation for consistency training.NeurIPS, 33, 2020.

[7] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang,and Colin Raffel. Remixmatch: Semi-supervised learning with distribution matching andaugmentation anchoring. InICLR, 2019.

[8] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raf-fel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence.NeurIPS, 33, 2020.

[9] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao wu, Jindong Wang, Okumura Manabu, and Shinozaki Takahiro. FlexMatch: Boosting Semi-Supervised Learning with Curriculum Pseudo Labeling. NeurIPS, 2021.