iBelieveCJM/Tricks-of-Semi-supervisedDeepLeanring-Pytorch：PseudoLabel 2013、增值税（VAT）、PI模型

项目介绍

iBelieveCJM/Tricks-of-Semi-supervisedDeepLeanring-Pytorch 是一个托管在 GitHub 上的开源项目，主要使用 Python 编写，采用 MIT License 许可证，目前获得 450 个 Star、72 次 Fork。

仓库全名: iBelieveCJM/Tricks-of-Semi-supervisedDeepLeanring-Pytorch

README 摘要

Tricks of Semi-supervised Deep Leanring --Pytorch The repository implements following semi-supervised deep learning methods: PseudoLabel 2013: The Simple and Efficient Semi-Supervised Learning Method for Deep Neural Networks (ICMLW 2013) PI&Tempens: Temporal Ensembling for Semi-Supervised Learning (ICLR 2017) MeanTeacher: Mean Teachers are better Role Models (NIPS 2017) VAT: Virtual Adversarial Training: A…

完整内容见 GitHub 原始 README 或本页。

常见问题

Tricks-of-Semi-supervisedDeepLeanring-Pytorch 用什么语言开发？

主要使用 Python 编写，采用 MIT License 许可证，目前有 450 个 Star。

Tricks-of-Semi-supervisedDeepLeanring-Pytorch 可以免费商用吗？

如何获取 Tricks-of-Semi-supervisedDeepLeanring-Pytorch 的源码？

可以直接克隆仓库：git clone https://github.com/iBelieveCJM/Tricks-of-Semi-supervisedDeepLeanring-Pytorch.git，默认分支为 master；安装与运行步骤见仓库 README。

Tricks-of-Semi-supervisedDeepLeanring-Pytorch 最近还在维护吗？

仓库最近一次更新为 2023年3月13日，当前有 5 个未关闭 Issue；Star 数为 450，可作为社区活跃度的参考。

关于

GitHub 原始描述（英文）

PseudoLabel 2013, VAT, PI model, Tempens, MeanTeacher, ICT, MixMatch, FixMatch

fixmatch ict meanteacher mixmatch pi pseudo-label pytorch semi-supervised-learning ssl tempens vat

MIT License
450 Stars
72 Forks
6 Watching

描述取自 GitHub 官方 API，保持仓库作者书写的原文。

简要分析

热度分

600

价值分

291

活跃状态

活跃

GitHub

语言占比

Python

Shell

贡献者

iBelieveCJM

Tricks of Semi-supervised Deep Leanring --Pytorch

The repository implements following semi-supervised deep learning methods:

PseudoLabel 2013: The Simple and Efficient Semi-Supervised Learning Method for Deep Neural Networks (ICMLW 2013)
PI&Tempens: Temporal Ensembling for Semi-Supervised Learning (ICLR 2017)
MeanTeacher: Mean Teachers are better Role Models (NIPS 2017)
VAT: Virtual Adversarial Training: A Regularization Method for Supervised and Semi-supervised Learning (TPAMI 2018)
ICT: Interpolation Consistency Training for Semi-supervised Learning (IJCAI 2019)
MixMatch: A Holistic Approach to Semi-Supervised Learning (NIPS 2019)
FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (2020)

This repository was created for my blog 半监督深度学习训练和实现小Tricks. Therefore the hyper-parameters are set for fair comparision, rather than performance.

The environment:

Ubuntu 16.04 + CUDA 9.0
Python 3.6.5:: Anaconda
PyTorch 0.4.1 and torchvision 0.2.1

To run the code:

The script run.sh includes some examples. You can try it as follow:

bash run.sh [gpu_id]

Some experimental results:

I haven’t run all models in this repository. Some results of this repo. are shown in results directory. And the following results came from this repository and the old codes which this repo. built on.

The following table shows the error rates of the CIFAR10 experiment with 4000 labeled training samples. The parameter settings are the same with the examples in run.sh.

	iPseudoLabel2013	ePseudoLabel2013	MeanTeacher	MixMatch	iFixMatch
orginal			12.31	6.24	4.26
v1	20.03	12.03	10.59	6.70	6.63
v2	15.82	10.82	9.46	6.89	6.44
	iTempens	eTempens	PI	ICT*	VAT
orginal		12.16	13.20	7.29	11.36
v1	10.98	10.74	14.11	7.12	13.84
v2	13.53	10.24	12.89	6.74	12.67

	eMixPseudoLabelv1	eMixPseudoLabelv2
soft	7.30	7.08
hard	7.33	7.20

Notes:

The MeanTeacher is the first model of the repository. So the hyper-parameters actually have been tuned for MeanTeacher.
My ICT is different from original one. The main difference is the unsupervised loss for unlabeled data.
eMixPseudoLabel is ePseudoLabel2013 with MixUp.
Instead of KL Divergence, my VAT uses MSE which bring more performance improvement.

暂无评论

Tricks-of-Semi-supervisedDeepLeanring-Pytorch