Attentional Feature Fusion
MXNet/Gluon code for "Attentional Feature Fusion" https://arxiv.org/abs/2009.14082
What's in this repo so far:
- Code, trained models, and training logs for CIFAR-100 and ImageNet
PS:
- If you are the reviewers of our submitted paper, please note that the accuracy of current implementation is a bit higher than the accuracy in the paper because it is a new implementation with a bag of tricks.
- 如果您是我的学位论文评审专家，发现论文与这个 repo 的数字有些出入，那是因为在论文提交后我又将代码重新实现了一遍，添加了 AutoAugment、Label Smooting 这些技巧，所以目前这个 repo 中的分类准确率会比论文中的数字高一些，还请见谅。
Change Logs:
- 2020-10-08: Re-implement the image classification code with a bag of tricks
- 2020-09-29: Upload the image classification codes and trained models for the submitted paper
To Do:
- Running AFF-ResNeXt-50 and AFF-ResNet-50 on ImageNet
- Update Grad-CAM results on new trained models
- Re-implement the segmentation code
- Convert to PyTorch
In Progress:
- Running iAFF-ResNeXt-50 on ImageNet
Done:
- Re-implement the image classification code with a bag of tricks
Requirements
pip install --upgrade mxnet-cu101 gluoncv
If you are going to use autoaugment:
python3 -m pip install --upgrade "mxnet_cu101<2.0.0"
python3 -m pip install autogluon
Experiments
All trained model params and training logs are in ./params
The training commands / shell scripts are in cmd_scripts.txt
CIFAR-100
| Architecture | Params | Accuracy |
|---|---|---|
| Attention-Augmented-Wide-ResNet-28-10 [3] | 36.2M | 81.6 |
| SENet-29 [4] | 35.0M | 82.2 |
| SKNet-29 [7] | 27.7M | 82.7 |
| PyramidNet-272-alpha-200 [8] | 26.0M | 83.6 |
| Neural Architecture Transfer (NAT-M4) [9] | 9.0M | 88.3 |
| AutoAugment+PyramidNet+ShakeDrop [10] | 26.0M | 89.3 |
| AFF-ResNet-32 | 7.9M | 90.8 |
| AFF-ResNeXt-38-32x4d (ours) | 7.8M | 91.2 |
| AFF-ResNeXt-47-32x4d (ours) | 9.7M | 91.8 |
ImageNet
| Architecture | Params | top-1 err. |
|---|---|---|
| ResNet-101 [1] | 42.5M | 23.2 |
| Efficient-Channel-Attention-Net-101 [2] | 42.5M | 21.4 |
| Attention-Augmented-ResNet-101 [3] | 45.4M | 21.3 |
| SENet-101 [4] | 49.4M | 20.9 |
| Gather-Excite-$\theta^{+}$-ResNet-101 [5] | 58.4M | 20.7 |
| Local-Importance-Pooling-ResNet-101 [6] | 42.9M | 20.7 |
| iAFF-ResNet-50 (ours) | 35.1M | 20.2 |
| iAFF-ResNeXt-50-32x4d (ours) | 34.7M | 19.78 |
References
[1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun: Deep Residual Learning for Image Recognition. CVPR 2016: 770-778
[2] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, Qinghua Hu: ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks. CVPR 2020: 11531-11539
[3] Irwan Bello, Barret Zoph, Quoc Le, Ashish Vaswani, Jonathon Shlens: Attention Augmented Convolutional Networks. ICCV 2019: 3285-3294
[4] Jie Hu, Li Shen, Gang Sun: Squeeze-and-Excitation Networks. CVPR 2018: 7132-7141
[5] Jie Hu, Li Shen, Samuel Albanie, Gang Sun, Andrea Vedaldi: Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks. NeurIPS 2018: 9423-9433
[6] Ziteng Gao, Limin Wang, Gangshan Wu: LIP: Local Importance-Based Pooling. ICCV 2019: 3354-3363
[7] Xiang Li, Wenhai Wang, Xiaolin Hu, Jian Yang: Selective Kernel Networks. CVPR 2019: 510-519
[8] Dongyoon Han, Jiwhan Kim, Junmo Kim: Deep Pyramidal Residual Networks. CVPR 2017: 6307-6315
[9] Zhichao Lu, Gautam Sreekumar, Erik D. Goodman, Wolfgang Banzhaf, Kalyanmoy Deb, Vishnu Naresh Boddeti: Neural Architecture Transfer. CoRR abs/2005.05859 (2020)
[10] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, Quoc V. Le: AutoAugment: Learning Augmentation Strategies From Data. CVPR 2019: 113-123