rog93 / caffe-mobilenetv2-relu6

Caffe implementation of ReLU6 Layer

Caffe-MobileNetV2-ReLU6

Caffe Implementation of ReLU6 Layer, Required By MobileNetV2.

ReLU6 Utilize

You should move .c* files to /path/to/caffe/src/caffe/layers/ and .hpp files to /path/to/caffe/include/caffe/layers/

Then add these lines to your caffe.proto file:

optional ReLU6Parameter relu6_param = 100000;

message ReLU6Parameter {
  optional float negative_slope = 1 [default = 0];
}

Experiments On ImageNet

I got top-1 error of 28.208% on imagenet which is slightly better than the performance claimed in paper(https://arxiv.org/abs/1801.04381).

I trained my model using farmingyard's MobileNetV2 config(https://github.com/farmingyard/caffe-mobilenet_v2) but replacing all ReLU Layers by ReLU6 Layers.

Unfortunately my inference is really time-consuming mainly caused by the ineffient caffe implementation of depthwise conv.

caffe-mobilenetv2-relu6's Issues

How to modify caffe-gpu in anaconda with given addition files?

I installed caffe-gpu via Anaconda in my linux machine.

I was just wondering how to add relu6_layer.cpp and relu6_layer.cu to /path/to/caffe/src/caffe/layers/ and relu6_layer.hpp files to /path/to/caffe/include/caffe/layers/ and recompile to be able to use Relu6 ?

pretrained model

Can you share the pretraind model of imagenet?That will be a lot helpful for me! Thanks!