Home

Welcome to the BNN_Caffe2 wiki!

Quantized Neural Network on Caffe2

Goal: Porting QNN(BNN) from Theano version to Caffe2 version.

What I Do

1.Test some Caffe2 ops using brew,

for example, in MNIST_BNN.ipynb:

// NxCxHxW: 1x1x28x28 -> 1x1x28x28

drop1 = brew.dropout(model,data, 'drop1', ratio=0.5, is_test=0)

// NxCxHxW: 1x1x28x28 -> 20x1x28x28

fc2 = brew.fc(model, drop1, 'fc2', dim_in=1 * 28 * 28, dim_out=20)

bn2 = brew.spatial_bn(model, fc2, 'bn2',dim_in = 20)

see https://zhuanlan.zhihu.com/p/27096092

2.Key parts: build binarization functions

1. Understand and test binarization functions of Theano version (I'm here now)
2. Porting.

Refs.

1. QNN Theano version paper https://arxiv.org/abs/1609.07061
2. Theano version github https://github.com/MatthieuCourbariaux/BinaryNet