PyTorch implementation of : Exploring Randomly Wired Neural Networks for Image Recognition.
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.
Download the ImageNet dataset and put them into the {repo_root}/data/imagenet
.
./train.sh configs/config_regular_c109_n32.yaml
All materials in this repository are released under the Apache License 2.0.