This repository contains the code and trained models for the work:

Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, Joseph E. Gonzalez
SkipNet: Learning Dynamic Routing in Convolutional Networks 
Published at ECCV 2018

SkipNet learns to route images through a subset of layers on a per-input basis. Challenging images are routed through more layers than easy images. We talk about two model designs with both feedforward gates and reccurent gates which enable different levels of parameter sharing in the paper. The model illustrations are as follows.

For other details, please refer to the paper.

If you find the code useful for your work, please consider citing

@InProceedings{Wang_2018_ECCV,
author = {Wang, Xin and Yu, Fisher and Dou, Zi-Yi and Darrell, Trevor and Gonzalez, Joseph E.},
title = {SkipNet: Learning Dynamic Routing in Convolutional Networks},
booktitle = {The European Conference on Computer Vision (ECCV)},
month = {September},
year = {2018}
}

• [April 9, 2018] add ray tune for auto hyperparameter search on CIFAR. Thanks @richardliaw
• [March 2, 2018] add trained model checkpoints on the CIFAR-10 dataset
• [March 1, 2018] upgrade ImageNet code to support Pytorch 3.0; add trained model checkpoints for ResNet-101 and ResNet-50

All the model configuration and training/evaluation code are available under ./cifar. If you want to train your own SkipNet, you can follow the document under the same folder.

Similarly, the model configuration and train/evaluation code are available under ./imagenet. Detailed instructions can be found in the document under the same folder.