-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
83 lines (76 loc) · 6.1 KB
/
index.html
File metadata and controls
83 lines (76 loc) · 6.1 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
<title>DPT</title>
<link href="style.css" rel="stylesheet">
<!--<script type="text/javascript" src="./ptp_files/jquery.mlens-1.0.min.js"></script>-->
<!--<script type="text/javascript" src="./ptp_files/jquery.js"></script>-->
<!-- <script src="http://ajax.googleapis.com/ajax/libs/jquery/1.7.1/jquery.min.js" type="text/javascript"></script>-->
<script type="text/javascript" src="./ptp_files/cat_hat.js"></script>
</head>
<body>
<div class="content">
<h1>Diffusion Models and Semi-Supervised Learners Benefit Mutually with Few Labels</h1>
<p id="authors"><a href="https://yyyouy.github.io/" target="_blank">Zebin You<sup>1</sup></a> Yong Zhong<sup>1</sup> <a href="https://baofff.github.io/" target="_blank">Fan Bao<sup>2</sup> </a> Jiacheng Sun<sup>3</sup> <a href="https://zhenxuan00.github.io/" target="_blank">Chongxuan li<sup>1</sup></a> <a href="https://ml.cs.tsinghua.edu.cn/~jun/index.shtml" target="_blank">Jun Zhu<sup>2</sup></a><br>
<span style="font-size: 16px"><sup>1</sup> Gaoling School of AI, Renmin University of China; Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China <sup>2</sup> Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua-Huawei Joint Center for AI, BNRist Center, State Key Lab for Intell. Tech. & Sys., Tsinghua University <sup>3</sup> Huawei Noah's Ark Lab
</span></p>
<br>
<img src="./ptp_files/dpt-overview.png" class="teaser-gif" style="width:100%;"><br>
<font size="+2">
<p style="text-align: center;">
<a href="https://arxiv.org/abs/2302.10586" target="_blank">Paper</a>
<a href="https://github.com/ML-GSAI/DPT/" target="_blank">Code</a>
</p>
</font>
</div>
<div class="content">
<h2 style="text-align:center;">Abstract</h2>
<p>We propose a three-stage training strategy called <b>dual pseudo training (DPT)</b> for conditional image generation and classification in semi-supervised learning. First, a classifier is trained on partially labeled data and predicts pseudo labels for all data. Second, a conditional generative model is trained on all data with pseudo labels and generates pseudo images given labels. Finally, the classifier is trained on real data augmented by pseudo images with labels. We demonstrate large scale diffusion models and semi-supervised learners benefit mutually with a few labels via DPT. In particular, on the ImageNet 256×256 generation benchmark, DPT can generate realistic, diverse, and semantically correct images with very few labels. With <b>two (i.e., < 0.2%) and five (i.e., < 0.4%)</b> labels per class, DPT achieves an FID of <b>3.44 and 3.37</b> respectively, outperforming strong diffusion models with <b>full labels</b>, such as IDDPM, CDM, ADM, and LDM. Besides, DPT outperforms competitive semi-supervised baselines substantially on ImageNet classification benchmarks with <b>one, two, and five</b> labels per class, achieving state-of-the-art top-1 accuracies of <b>59.0 (+2.8), 69.5 (+3.0), and 73.6 (+1.2)</b> respectively.
</p>
</div>
<div class="content">
<h2>An overview of DPT</h2>
<img class="summary-img" src="./ptp_files/overview.png" style="width:100%;">
<p> <b>First</b>, a (self-supervised) classifier is trained on partially labeled data and used to predict pseudo labels
for all data. <b>Second</b>, a (diffusion-based) conditional generative model is trained on all data with pseudo labels and used to generate pseudo
images given random labels. <b>Finally</b>, the classifier is trained or fine-tuned on real data augmented by pseudo images with labels.</p>
<br>
<br>
</div>
<div class="content">
<h1>Generation and classification results of DPT on ImageNet with few labels</h1>
<h2>DPT with <b>< 0.4%</b> labels vs. supervised diffusion models</h2>
<img class="summary-img" src="./ptp_files/fid.jpg" style="width:80%;">
<p>
DPT with <b>< 0.4%</b> labels outperforms strong supervised diffusion models, including <a href="https://arxiv.org/abs/2106.15282" target="_blank">CDM(Ho et al., 2022)</a>, <a href="https://arxiv.org/abs/2105.05233" target="_blank">ADM(Dhariwal & Nichol, 2021)</a> and <a href="https://arxiv.org/abs/2112.10752" target="_blank">LDM(Rombach et al., 2022).</a> <b>The bubble area</b> indicates the label fraction.
</p>
<h2>DPT vs. SOTA semi-supervised classifiers</h2>
<img class="summary-img" src="./ptp_files/accuracy.jpg" style="width:80%;">
<p>
With <b>one, two, five</b> labels per class, DPT improves the state-of-the-art semi-supervised learner <a href="https://arxiv.org/abs/2204.07141" target="_blank">MSN(Assran et al., 2022)</a> consistently and substantially.
</p>
</div>
<div class="content">
<h2>Random samples by varying the number of real labels in the first stage</h2>
<p>More real labels result in smaller noise in pseudo labels and samples of better visual quality and correct semantics. Top: “Custard apple”. Middle: “Geyser”. Bottom: “Goldfish”.</p>
<p>For a given class <b>y</b>, the precision and recall w.r.t. the classifier is defined by <b>P = TP/(TP + FP)</b> and <b>R = TP/(TP + FN)</b>, where TP, FP, and FN denote the number of true positive, false positive, and false negative samples respectively.</p>
<img class="summary-img" src="./ptp_files/range_samples.png" style="width:100%;">
</div>
<div class="content">
<h2>Conclusion</h2>
<p>
This paper presents a training strategy DPT for conditional image generation and classification in semi-supervised learning. DPT is simple to implement and achieves excellent performance in both tasks on ImageNet. DPT probably inspires future work in semi-supervised learning
</p>
</div>
<div class="content">
<h4>BibTex</h4>
<p> @article{you2023diffusion,<br>
title={Diffusion Models and Semi-Supervised Learners Benefit Mutually with Few Labels},<br>
author={You, Zebin and Zhong, Yong and Bao, Fan and Sun, Jiacheng and Li, Chongxuan and Zhu, Jun},<br>
booktitle={arXiv preprint arXiv:2302.10586},<br>
year={2023}<br>
}</p>
<br>
</div>
</body>
</html>