How to train Caffe with only G and B channels

问题

Is there anyway to use only G and B channels for training Caffe using "ImageData" input layer?

回答1:

You can add a convolution layer on top of your input that will select G and B:

layer {
  name: "select_B_G"
  type: "Convolution"
  bottom: "data"
  top: "select_B_G"
  convolution_param { kernel_size: 1 num_output: 2 bias_term: false }
  param { lr_mult: 0 } # do not learn parameters for this layer
}

You'll need to do some net surgery prior to training to set the weights for this layer to be

net.params['select_B_G'][0].data[...] = np.array( [[1,0,0],[0,1,0]], dtype='f4')

Note: sometimes images loaded to caffe are going through channel-swap transformation, i.e., RGB -> BGR, therefore you need to be careful what channels you pick.

回答2:

I wrote a simple python layer to do this, by the way, I did't test this code.

import caffe

class ExtractGBChannelLayer(caffe.Layer):
  def setup(self,bottom,top):
    pass
  def reshape(self,bottom,top):
    bottom_shape=bottom[0].data.shape
    top_shape=[bottom_shape[0],2,bottom_shape[2],bottom_shape[3]] #because we only want G and B channels.
    top[0].reshape(*top_shape)
  def forward(self,bottom,top):
    #copy G and B channel to top, note caffe BGR order!
    top[0].data[:,0,...]=bottom[0].data[:,1,...]
    top[0].data[:, 1, ...] = bottom[0].data[:, 0, ...]
  def backward(self,top,propagate_down,bottom):
    pass

You can save this file as MyPythonLayer.py

In you prototxt you can insert a layer after ImageDataLayer like this

layer {
  name: "GB"
  type: "Python"
  bottom: "data"
  top: "GB"
  python_param {
    module: "MyPythonLayer"
    layer: "ExtractGBChannelLayer"
  }
}

Hope this works fine.

回答3:

This is the Matlab code I used and it works.

 caffe.reset_all(); % reset caffe
 caffe.set_mode_gpu();  
 gpu_id = 0;  % we will use the first gpu in this demo
 caffe.set_device(gpu_id);
 net_model = ['net_images.prototxt'];
 net = caffe.Net(net_model, 'train')
 a = zeros(1,1,3,2);
 a(1,1,:,:) = [[1,0,0];[0,1,0]]'; % caffe uses BGR color channel order
 net.layers('select_B_G').params(1).set_data(a);
 solver = caffe.Solver(solverFN);
 solver.solve();
 net.save(fullfile(model_dir, 'my_net.caffemodel'));

来源：https://stackoverflow.com/questions/36551284/how-to-train-caffe-with-only-g-and-b-channels