the bp of binary_conv layer

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

the bp of binary_conv layer #2

Open NHZlX opened 7 years ago

NHZlX commented 7 years ago

I take a cursory look at the implementation of the binary_conv_layer , and i find that the bp process of the weight of the layer is inconsistent with the paper. This is for simplicity or mistake, i will try the effect.

HongxingGao commented 7 years ago

Yes, I have noticed that. I have implemented the gradient calculation as specified in the paper. But the final result did not show any improvement on ImageNet. Hence, i didn't update that code.

NHZlX commented 7 years ago

@HongxingGao I am trying to modify it, is it convenient for you to give me your revised code? zlx_hg@163.com

caseyanya commented 7 years ago

I'm also suffering from accuracy on ImageNet classification (darknet19 in binary weight from scratch). What is your Top-1 accuracy? I cannot even pass 20% Top-1 accuracy using binary weight only.