How to create on-the-fly python layer for calculating class weights in caffe?

Question

I want to add a python layer that calculates probability matrix H for INfoGainLoss Layer on-the-fly. I have already wrote the program for creating this matrix by calculating probabilities of each class of one image and saving it into .binaryproto file. I really appreciate if you give me some hints that how can I write a python layer that creates this matrix and send it as third parameter to InfoGainLoss layer? I drew a schematic here, Is it correct? If yes, how to write the python layer for this? I already read some codes online but still have confusion on setup,reshape,forward functions.

7a.png

Answer 1

Your Python layer is quite similar to an input layer: It has no backward propagation, which makes it easy to implement. See this thread for more details.

Your layer expects "label" bottom, and produced "H" matrix as top:

layer { name: "classWeightH" bottom: "label" top: "H" type: "Python" python_param { module: # file name where python code is layer: "classWeightHLayer" } }

The python code should look something like:

import sys, os, numpy as np
sys.path.insert(0, os.environ['CAFFE_ROOT']+'/python')
import caffe
class classWeightHLayer(caffe.Layer):
  def setup(self,bottom,top):
    assert len(bottom)==1, "expecting exactly one input"
    assert len(top)==1, "producing exactly one output"
    # you might want to get L - the number of labels as a parameter...

  def reshape(self,bottom,top):
    top[0].reshape(1,1,L,L) # reshape the output to the size of H

  def forward(self,bottom,top):         
    labels = bottom[0].data
    H = np.zeros((1,1,L,L), dtype='f4') 
    # do your magic here...
    top[0].data[...] = H    

  def backward(self, top, propagate_down, bottom):
    # no back-prop for input layers
    pass