search

HongxingGao / XNOR-Net

XNOR-Net, CUDNN5 supported version of XNOR-Net-caffe: https://github.com/loswensiana/BWN-XNOR-caffe
31 stars 15 forks source link

Model size #1

Closed chenfsjz closed 7 years ago

chenfsjz commented 7 years ago

Hi,I have a problem with your version of xnor net. after training the cifar10_xnor net,the memory and the final size of the caffemodel just are about the same with the cifar10_full_net, could you give me some suggestions.

Ivan-Zhao commented 7 years ago

me too! How to use it?

NHZlX commented 7 years ago

@chenfsjz That's because the parameters of trained network in the caffemodel are not the final binary storage.

HongxingGao commented 7 years ago

@NHZlX is right. the parameters are stored as float for simiplicity. But you can easily binarize them by sign(x). But please bear in mind, you will also need to compute and store the filter-wise \alpha = |w|/n as stated in the paper.