copy_weights.py

# Done for VGG, may be different for an other architecture
# Convert VGG into a fully convolutional network (by copying the fc6 and fc7 fully convolutional layers)
# Not sure of what it does exactly


import os
import numpy as np
import caffe

def make_fully_conv(path_model_src,
                    path_weights_src,
                    path_model_full_conv,
                    param_pairs,
                    path_weights_dst
                    ):
    
    # PART A
    # Load the original network and extract the fully connected layers' parameters.
    net = caffe.Net(path_model_src, path_weights_src, caffe.TEST)
    
    params = [src for src, _ in param_pairs]
    
    # fc_params = {name: (weights, biases)}
    fc_params = {pr: (net.params[pr][0].data, net.params[pr][1].data) for pr in params}
    
    print "Original dimensions:"
    for fc in params:
        print '{} weights are {} dimensional and biases are {} dimensional'.format(fc, fc_params[fc][0].shape, fc_params[fc][1].shape)
        
    # PART B
    # Load the fully convolutional network to transplant the parameters.
    net_full_conv = caffe.Net(path_model_full_conv, 
                              path_weights_src,
                              caffe.TEST)
    params_full_conv = [dst for _, dst in param_pairs]
    # conv_params = {name: (weights, biases)}
    conv_params = {pr: (net_full_conv.params[pr][0].data, net_full_conv.params[pr][1].data) for pr in params_full_conv}
    
    for conv in params_full_conv:
        print '{} weights are {} dimensional and biases are {} dimensional'.format(conv, conv_params[conv][0].shape, conv_params[conv][1].shape)
        
    #Let's transplant!
    print 'Let\'s transplant !'
    for pr, pr_conv in zip(params, params_full_conv):
        conv_params[pr_conv][0].flat = fc_params[pr][0].flat  # flat unrolls the arrays
        conv_params[pr_conv][1][...] = fc_params[pr][1]
        
    # save new weights
    net_full_conv.save(path_weights_dst)
    
    return 0


def main(args):
    
    caffe.set_mode_gpu()
    
    param_pairs = [('fc6', 'fc6-conv'),
                   ('fc7', 'fc7-conv'),
                   ('fc8', 'fc8-conv')]
    make_fully_conv('/home/pierre/tmpModels/VGG/VGG_ILSVRC_16_layers_deploy.prototxt',
                    '/home/pierre/tmpModels/VGG/VGG_ILSVRC_16_layers.caffemodel',
                    '/home/pierre/tmpModels/VGG/VGG_ILSVRC_16_layers_fcn_deploy.prototxt',
                    param_pairs,
                    '/home/pierre/tmpModels/VGG/VGG_ILSVRC_16_layers_conv.caffemodel',
                    )
    
    
    return 0

if __name__ == '__main__':
    
    main(None)
    
    pass