/GENet

Primary LanguageC++

Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks (paper)

By Jie Hu[1], Li Shen[2], Samuel Albanie[2], Gang Sun[1], Andrea Vedaldi[2].

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

Approach

Figure 1: The Diagram of a Gather-Excite module.

Figure 2: The Schema of Gather-Excite modules. Top-left: $GE-\theta^-\,(E8)$. Top-right: $GE-\theta^-$. Bottom-left: $GE-\theta\,(E8)$. Bottom-right: $GE-\theta$.

Implementation

In this repository, all the models are implemented by Caffe.

We use the data augmentation strategies with our previous work SENet.

There are three new layers introduced for efficient training and inference, these are Axpy, DepthwiseConvolution and CuDNNBatchNorm layers.

  • The Axpy layer is already implemented in SENet.
  • As for DepthwiseConvolution layer, it is publicly avaliable at the repo, or you can directly replace it with the standard Convolution layer if you don't care about the operational efficiency.

Trained Models

Coming soon!

Citation

If you use Gather-Excite in your research, please cite the paper:

@inproceedings{hu2018genet,
  title={Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks},
  author={Jie Hu and Li Shen and Samuel Albanie and Gang Sun and Andrea Vedaldi},
  journal={NeurIPS},
  year={2018}
}