ZhangVision / bitnet

Convolutional Networks with Binary Tree Architectures

Convolutional Networks with Binary Tree Architectures

This repository is the torch implementation for the paper Truncating Wide Networks using Binary Tree Architectures.

Installation and Usage

To run our code,

• Clone and install fb.resnet.torch.
• Put bitnet.lua into the folder models/.
• Replace opts.lua of fb.resnet.torch with our opts.lua.
• (Optional) To use the same learning rate schedule for cifar classification task as in the paper, one can change line 180/182 of train.lua to decay = epoch >= 160 and 3 or epoch >= 120 and 2 or epoch >= 60 and 1 or 0.

To train on ImageNet with bitnet-26 model,

th main.lua -dataset imagenet -netType bitnet -depth 26 -nEpochs 90 -data /path/to/ILSVRC2012/.

To train on cifar,

th main.lua -dataset cifar100 -netType bitnet -width 10 -bdepth 2 -nblock 3 -LR 0.1 -nEpochs 200

where width bdepth nblock implement the d k n in the Table 1 of the paper. For other traing options e.g. batchSize and nGPU, one can check fb.resnet.torch.

Pretrained Models

Pretrained bitnet models for ImageNet are available for download (Google Drive): bitnet-26, bitnet-34

How to use pretrained models? See instructions here.

Citation

If you use our code in your research and would like to cite our paper:

@article{zhang2017bitnet,
  title={Truncating Wide Networks using Binary Tree Architectures},
  author={Yan Zhang and Mete Ozay and Shuohao Li and Takayuki Okatani},
  journal={International Conference on Computer Vision},
  year={2017}
}

If you have any further questions or suggestion, feel free to contact us at zhang at vision.is.tohoku.ac.jp