Bi-Real-net

Bi-Real-net copied to clipboard

请问你们提出的“Magnitude-aware gradient with respect to weights”和xnor中有什么区别？

1

看起来sign的系数好像都是weight的一范数的均值

pppLang

Conv Layer backward computation

3

The Readme says "The binary convolution layer is modified from https://github.com/loswensiana/BWN-XNOR-caffe, in which we modified the gradient computation method" However, looking at the Diff of https://github.com/loswensiana/BWN-XNOR-caffe/blob/master/src/caffe/ex_layers/binary_conv_layer.cpp and https://github.com/liuzechun/Bi-Real-net/blob/master/Bi-Real-net-caffe/caffe-train/src/caffe/ex_layers/binary_conv_train_layer.cpp It looks...

adrianloy

a question about BinaryActivation

1

Hello.Sorry to bother you,I am very confused about the code in class BinaryActivation, could give me some explanation? I don't know what is the meaning of out1, out2 and out3.

etoilestar

Great work! Recently, we are replicating your work (Bi-Real-net 34) Could you please provide the training log of the PyTorch version for our reference? Thank you very much!

WenHao-L

Wonderful work! Recently, we are researching on the binary object detection, and need the pre-trained binary backbones(Bi-Real Res18/34/50...). Could you provide such pre-trained models? Thanks very much!

LZHgrla

输出权重为什么不是二值化的

3

我用python输出了训练得到的model，为什么权重都是小数，不是+1 -1 啊

feifanzui

PyTorch Implementation: Forward Pass

2

First of all, thank you for sharing the PyTorch implementation, it's wonderful. I've been going over the code and found this line: binary_weights = binary_weights_no_grad.detach() - cliped_weights.detach() + cliped_weights in...

killawhale2

The downsample layer in basic block is not binarized?

2

https://github.com/liuzechun/Bi-Real-net/blob/0a4b802f94009b8b2419d53b0152a94feff81189/18-layer/Bi-Real-net-18-train.prototxt#L355 In your code, it seems that the downsample layer in basic block is not binarized. Is it standard in the quantization field?

panda1949

你好，我在复现pytorch版本的过程中输出binary_weights，但是不是二值化数值，请问是我哪里有遗漏或者没有操作对吗？

lmy418lmy

Accuracy of ResNet-34 and ResNet-50

2

Thanks for your update of pytorch implementation. I noticed that the accuracies of binary ResNet-34 and binary ResNet-50 are similar. What do you think is the cause? Just for academic...

iamhankai