Binary CNN by chainer

Binarized Neural Networks: Training Neural Networks with Weights and Activation Constrained to +1 or -1

I implement Binarized Neural Network by chainer. There are three different point from ordinary CNN.

1. Using Binarized Weight

2. Using Binarized Input

3. Using weight clip that constraine gradient to -1 < x < 1

But I don't implement these below.

• Shift Based Operation of

• Batch Normalization

• AdaMax

• XNOR Dot

• stochastic Binarization

Usage

./mnist_cnn.py

./cifar10_cnn.py

You can choose options

• gpu

• epoch

• batchsize

code explanation

link_binary_convolution.py and function_binary_convolution.py define Link of chainer's object

net.py defines network

weight_clip.py constraines gradient to -1 < x < 1 at update step

Reference

I implemented these codes hillbig/binary_net as reference