Chinese Journal of Liquid Crystals and Displays, Volume. 39, Issue 9, 1223(2024)
Hourglass attention and progressive hybrid Transformer for image classification
Figures & Tables(9)
Fig. 1. Overall architecture of HAPHFormer.（a）ConvStem module；（b）Embed module；（c）PatchEmbed module.
Fig. 2. General architecture of Transformer（a）and structure of P-LocalMLP module（b）
Table 1. Ablation experiment on CIFAR100 dataset
View tableView in ArticleTable 1. Ablation experiment on CIFAR100 dataset
Model Top-1 accuracy/% ViT 73.81 ViT+τ 74.35 ViT+M 74.34 ViT+τ+M 74.87 ViT+τ+M+Stem 78.24 ViT+τ+M+Stem+DTSA 83.45 ViT+τ+M+Stem+DTSA+PLM 83.76
Table 2. Top-1 accuracy comparison of downsampling module on CIFAR100 dataset
View tableView in ArticleTable 2. Top-1 accuracy comparison of downsampling module on CIFAR100 dataset
Downsampling module Top-1 accuracy/% Patch Emb 81.71 Patch Stem 78.45 ConvStem 83.76
Table 3. Top-1 accuracy comparison of P-LocalMLP module %
View tableView in ArticleTable 3. Top-1 accuracy comparison of P-LocalMLP module %
Model T-ImageNet CIFAR10 BN+FN 69.39 94.64 DW+LN 71.51 96.97 DW+LN+ skip connect 72.09 97.32
Table 4. Top-1 accuracy comparison of attention module on CIFAR100 dataset
View tableView in ArticleTable 4. Top-1 accuracy comparison of attention module on CIFAR100 dataset
Attention module M+τ Top-1 accuracy/% DTSA × 83.18 √ 83.76 EMSAv2 × 80.59 √ 82.29 Linear SRA × 79.58 √ 82.52
Table 5. Top-1 accuracy comparison of overall architecture
View tableView in ArticleTable 5. Top-1 accuracy comparison of overall architecture
Architecture Top-1 accuracy/% Attention，Attention，Attention，Attention 73.69 Pool，Pool，Pool，Pool 62.09 SpatialMLP，SpatialMLP，SpatialMLP，SpatialMLP 61.73 Pool，Pool，SpatialMLP，SpatialMLP 77.83 Pool，Pool，Attention，Attention 83.76
Table 6. Top-1 accuracy comparison of different models on small-size datasets
View tableView in ArticleTable 6. Top-1 accuracy comparison of different models on small-size datasets
Model Params/M FLOPs/G T-ImageNet/% CIFAR10/% CIFAR100/% SVHN/% ConvNeXt-T 29 4.46 67.51 94.77 76.87 96.17 SwinV2-T 28 4.35 50.41 78.22 68.97 82.19 PVTv2-B2 26 4.04 63.93 93.58 73.29 97.01 MViTv2-T 24 3.99 53.94 74.44 73.16 85.99 ResTv2-T 30 4.10 68.59 95.57 80.57 97.17 HAPHFormer-T 25 3.41 72.09 97.32 83.76 97.42 ConvNeXt-S 50 8.69 67.88 95.82 81.33 96.21 SwinV2-S 50 8.45 58.97 80.06 70.32 85.03 PVTv2-B3 45 6.92 66.85 94.27 73.44 97.41 MViTv2-S 35 6.08 56.92 76.76 74.13 86.34 ResTv2-S 40 5.97 69.70 95.91 80.89 97.23 HAPHFormer-S 35 5.11 72.29 97.37 83.88 97.77 ConvNeXt-B 88 15.36 68.69 96.04 83.30 96.33 SwinV2-B 87 14.99 60.14 80.38 70.51 85.51 PVTv2-B4 62 10.14 67.96 94.94 77.25 97.65 MViTv2-B 52 8.88 62.43 77.18 74.51 86.68 ResTv2-B 55 7.88 70.10 96.21 81.04 97.26 HAPHFormer-B 49 6.84 72.55 97.38 84.12 97.89 PVTv2-B5 81 11.75 68.37 95.14 78.42 97.88 ResTv2-L 86 13.83 72.34 96.37 81.74 97.34 HAPHFormer-L 72 11.27 72.80 97.46 84.71 98.16
Tools
Get Citation
Copy Citation Text
Yanfei PENG, Yun CUI, Kun CHEN, Yongxin LI. Hourglass attention and progressive hybrid Transformer for image classification[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(9): 1223
Paper Information
Category:
Received: Oct. 25, 2023
Accepted: --
Published Online: Nov. 13, 2024
The Author Email: Yun CUI (1727015916@qq.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20