RandWireNN(Randomly Wired Neural Network)

PyTorch implementation of : Exploring Randomly Wired Neural Networks for Image Recognition.

Update

• 2019/4/10: Release a result of regular computation(C=109) RandWird-WS(4,0.75). It has Top-1 accuracy of 77.07% on Imagenet dataset.
• 2019/4/7: The code of RandWireNN are released.

Reproduced results

| Model | Paper's Top-1 | Mine Top-1 | Epochs |LR Scheduler| Weight Decay | | :----:| :--: | :--: | :--: | :--: | :--: | |RandWire-WS(4, 0.75), C=109| 79% | 77% | 100 | cosine lr | 5e-5 | |RandWire-WS(4, 0.75), C=78| 74.7% | 73.97% | 250 | cosine lr | 5e-5 |

*This result does not take advantage of dropout, droppath and label smoothing techniques. I will use these tricks to retrain the model.

Requirements

• python packages
  • pytorch = 0.4.1
  • torchvision>=0.2.1
  • tensorboardX
  • pyyaml
  • CVdevKit
  • networkx

Data Preparation

Download the ImageNet dataset and put them into the {repo_root}/data/imagenet.

Training a model from scratch

./train.sh configs/config_regular_c109_n32.yaml

License

All materials in this repository are released under the Apache License 2.0.