binary weight neural network implementation on tensorflow
This is an implementation code for reproducing BNN
python mnist.py
python cifar10.py
| DataSet | accuracy |
|---|---|
| MNIST | 99.04% |
| Cifar10 | 86.18% |
layer-wise learning rate, paper is layer_lr = 1./sqrt(1.5 / (num_inputs + num_units)), my implement is layer_lr / 4
[BinaryNet: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1"](http://arxiv.org/abs/1602.02830),
Matthieu Courbariaux, Yoshua Bengio