Squeeze-and-Excitation Networks _(paper)

By Jie Hu^[1], Li Shen^[2], Gang Sun^[1].

Momenta^[1] and University of Oxford^[2].

Approach

<div align="center"> <img src="https://github.com/hujie-frank/SENet/blob/master/figures/SE-pipeline.jpg"> </div> <p align="center"> Figure 1: Diagram of a Squeeze-and-Excitation building block. </p> <div align="center">  <img src="https://github.com/hujie-frank/SENet/blob/master/figures/SE-Inception-module.jpg" width="420"> <img src="https://github.com/hujie-frank/SENet/blob/master/figures/SE-ResNet-module.jpg" width="420"> </div> <p align="center"> Figure 2: Schema of SE-Inception and SE-ResNet modules. We set r=16 in all our models. </p>

Implementation

In this repository, Squeeze-and-Excitation Networks are implemented by Caffe.

Augmentation

| Method | Settings | |:-:|:-:| |Random Mirror| True | |Random Crop| 8% ~ 100% | |Aspect Ratio | 3/4 ~ 4/3 | |Random Rotation| -10° ~ 10°| |Pixel Jitter| -20 ~ 20 |

Note:

• To achieve efficient training and testing, we combine the consecutive operations channel-wise scale and element-wise summation into a single layer "Axpy" in the architectures with skip-connections, resulting in a considerable reduction in memory cost and computational burden.

• In addition, we found that the implementation for global average pooling on GPU supported by cuDNN and BVLC/caffe is less efficient. In this regard, we re-implement the operation which achieves significant acceleration.

Trained Models

Table 1. Single crop validation error on ImageNet-1k (center 224x224 crop from resized image with shorter side = 256). The SENet-154 is one of our superior models used in ILSVRC 2017 Image Classification Challenge where we won the 1st place (Team name: WMW).

| Model | Top-1 | Top-5 | Size | Caffe Model | Caffe Model |:-:|:-:|:-:|:-:|:-:|:-:| |SE-BN-Inception| 23.62 | 7.04 | 46 M| GoogleDrive | BaiduYun |SE-ResNet-50 | 22.37 | 6.36 | 107 M | GoogleDrive | BaiduYun |SE-ResNet-101 | 21.75 | 5.72 | 189 M | GoogleDrive | BaiduYun |SE-ResNet-152 | 21.34 | 5.54 | 256 M | GoogleDrive | BaiduYun |SE-ResNeXt-50 (32 x 4d) | 20.97 | 5.54 | 105 M | GoogleDrive | BaiduYun |SE-ResNeXt-101 (32 x 4d) | 19.81 | 4.96 | 187 M | GoogleDrive | BaiduYun |SENet-154 | 18.68 | 4.47 | 440 M | GoogleDrive | BaiduYun

Here we obtain better performance than those reported in the paper. We re-train the SENets described in the paper on a single GPU server with 8 NVIDIA Titan X cards, using a mini-batch of 256 and a initial learning rate of 0.1 with more epoches. In contrast, the results reported in the paper were obtained by training the networks with a larger batch size (1024) and learning rate (0.6) across 4 servers.

Third-party re-implementations

0. Caffe. SE-mudolues are integrated with a modificated ResNet-50 using a stride 2 in the 3x3 convolution instead of the first 1x1 convolution which obtains better performance: Repository.
1. TensorFlow. SE-modules are integrated with a pre-activation ResNet-50 which follows the setup in fb.resnet.torch: Repository.
2. TensorFlow. Simple Tensorflow implementation of SENets using Cifar10: Repository.
3. MatConvNet. All the released SENets are imported into MatConvNet: Repository.
4. MXNet. SE-modules are integrated with the ResNeXt and more architectures are coming soon: Repository.
5. PyTorch. Implementation of SENets by PyTorch: Repository.
6. Chainer. Implementation of SENets by Chainer: Repository.

Citation

If you use Squeeze-and-Excitation Networks in your research, please cite the paper:

@inproceedings{hu2018senet,
  title={Squeeze-and-Excitation Networks},
  author={Jie Hu and Li Shen and Gang Sun},
  journal={IEEE Conference on Computer Vision and Pattern Recognition},
  year={2018}
}