Caffe Add a new neural layer

first step:
Creating a new defined header file include / caffe / layers / my_neuron_layer.hpp

You can refer to the frame header file of other layers, etc., inherit the neural layer

Retreat method: virtual inline const char * type () const {return "myneuron";

If you just need a CPU method, you can comment on forward_gpu () and backward_gpu ().

Step 2:

Create source files for corresponding src / caffe / src / my_neuron_layer.cpp

Can refer to the frame header file of other layers, etc.

Rewriting method Layersetup, implementing the parameters from prototxt, if you don't need to read parameters, you can not rewrite

Rewriting the method RESHAPE, if you don't modify the inheritance class, you can use it without rewriting

Rewriting method forward_cpu; Caffe has a function of some mathematical operations. CAFFE calculation is a matrix operation

Override method backward_cpu (no must)

If you want GPU support, you need to create src / caffe / src / my_neuron_layer.cu. Similar rewriting method

Forward_gpu / backward_gpu (non-must)

third step:

Register a new Layer in Proto / Caffe.Proto

message LayerParameter{

...

++optional MyNeuronParameter

mysquare_param = 150; Cannot repeat the flag number

...

}

...

++message MyNeuronParameter{

++ Optional float power = 1 {default = 2}; parameters in Layer

++

}

...

message V1LayerParameter{

...

++ MYNEURON = 40; add a parameter setting

...

}

Fourth Step MYNEURON_LAYER.CPP Add registration macro definition

INSTANTIATE_CLASS(MyNeuronLayer);

REGISTER_LAYER_CLASS(MyNeuron);

If there is MY_NEURON_LAYER.CU, add

INSTANTIATE_LAYER_GPU_FUNCS(MyNeuronLayer);

the fifth step

Re-build Caffe

Sixth step test

You can write a prototxt that only the input layer and custom layer

Then test it

The test procedure is as follows:

import numpy as np
import matplotlib.pyplot as plt
import os 
import sys
import caffe

deploy_file="./deploy.prototxt"
test_data="5,jpg"
if __name__ ='__main__':
    net = caffe.Net(deploy_file,caffe.TEST)
    transformer = caffe.io.Transformer({'data':net.blobs['data'].data.shape})
    transformer.set_transpose('data',(2,0,3))
    img= caffe.io.load_image(test_data,color=False)
    net.blobs['data'].data[...]=transformer.preprocess('data',img)
    print net.blobs['data'].data
    out = net.forward()
    print out['data_out']

 

other:

Custom data layer

1. Create a new defined header file include / caffe / layers / my_data_Layer.hpp

Retreat method: Virtual inline const char * type () const {return "mydata";

2, rewrite Layersetup, shuffleimage, load_batch

3, proto / caffe.proto registration new Layer

//my new data parameter

optional MyDataParameter my_data_param = 50

//my new data layer for infer image data

optional MyDataPatameter my_data_param=150

...

message MyDataParameter{

 

}