Squeeze-and-Excitation Networks (arXiv)
By Jie Hu[1], Li Shen[2], Gang Sun[1].
Momenta[1] and University of Oxford[2].
Figure 1: Diagram of a Squeeze-and-Excitation building block.
Figure 2: Schema of SE-Inception and SE-ResNet modules. We set r=16 in all our models.
In this repository, Squeeze-and-Excitation Networks are implemented by Caffe.
| Method | Settings |
|---|---|
| Random Mirror | True |
| Random Crop | 8% ~ 100% |
| Aspect Ratio | 3/4 ~ 4/3 |
| Random Rotation | -10° ~ 10° |
| Pixel Jitter | -20 ~ 20 |
-
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.
Table 1. Single crop validation error on ImageNet-1k (center 224x224 crop from resized image with shorter side = 256). The SENet* 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 |
|---|---|---|---|---|
| SE-BN-Inception | 23.62 | 7.04 | 46 M | GoogleDrive |
| SE-ResNet-50 | 22.37 | 6.36 | 107 M | GoogleDrive |
| SE-ResNet-101 | 21.75 | 5.72 | 189 M | GoogleDrive |
| SE-ResNet-152 | 21.34 | 5.54 | 256 M | GoogleDrive |
| SE-ResNeXt-50 (32 x 4d) | 20.97 | 5.54 | 105 M | GoogleDrive |
| SE-ResNeXt-101 (32 x 4d) | 19.81 | 4.96 | 187 M | GoogleDrive |
| SENet* | 18.68 | 4.47 | 440 M | GoogleDrive |
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.
- 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.
- TensorFlow. SE-modules are integrated with a pre-activation ResNet-50 which follows the setup in fb.resnet.torch: Repository.
- MatConvNet. All the released SENets are imported into MatConvNet: Repository.
If you use Squeeze-and-Excitation Networks in your research, please cite the paper:
@article{hu2017,
title={Squeeze-and-Excitation Networks},
author={Jie Hu and Li Shen and Gang Sun},
journal={arXiv preprint arXiv:1709.01507},
year={2017}
}