• install the latest PyTorch (1.1.0)
• download this code (pretrained model included)
• get the ImageNet validation set ready (you may need the script valprep.sh to pre-process the val set)

• run python weight_summary.py to check whether the weights are binary
• check the code in the module BinConv2d to see whether the input for convolution is binary

Modifications in the script evaluate_imagenet.sh:

• modify the PATH to your ImageNet dataset
• modify the batchsize (default: 256) according to your hardware (at least one GPU is requried)

Run the script evaluate_imagenet.sh and the accuracy on validation set is around 57.30.

@inproceedings{gu2019projection,
  title={Projection Convolutional Neural Networks for 1-bit CNNs via Discrete Back Propagation},
  author={Gu, Jiaxin and Li, Ce and Zhang, Baochang and Han, Jungong and Cao, Xianbin and Liu, Jianzhuang and Doermann, David},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={33},
  pages={8344--8351},
  year={2019}
}