Skip to content
PyTorch implementation of binary neural networks
Branch: master
Clone or download
Latest commit d48f5e8 Mar 20, 2019

Binary neural networks

Implementation of some architectures from Structured Binary Neural Networks for Accurate Image Classification and Semantic Segmentation in Pytorch


All architectures are based on ResNet18 now.

There are two groups of models:

  • torchvision ResNet compatible. The only difference is BasicBlock that is used inside.
  • deep-person-reid compatible. Models with small change in the forward method to be easily integrated with deep-person-reid project.

Note about binary NN training

When we train binary neural networks we usually use quantized weights and activations for forward and backward passes and full-precision weights for update. That's why usual backward pass and weights update


should be changed with:

for p in list(model.parameters()):
    if hasattr(p, 'original'):
for p in list(model.parameters()):
    if hasattr(p, 'original'):
        p.original.copy_(, 1))

ONNX compatibility:

Some changes were made into models with fusion gate to make them ONNX-compatible. Models for training use modules with custom backward function, that can't be converted with ONNX, that's why they are changed with simple sign function for inference. To create inference model you should pass freeze=True flag.


Proper initialization for inference models:

  • Mean of weights should be merged into batchnorms
  • Weights binarization should be done
You can’t perform that action at this time.