fcn16_code_sample.py

fcn32model = fcn32_blank()
        
fcn32shape = fcn32model.layers[-1].output_shape
    
fcn32size = fcn32shape[2] # INFO: =32 when images are 512x512
    
sp4 = Convolution2D(21, 1, 1,
                    border_mode='same', # INFO : border_mode does not matter for 1x1
                    activation=None,
                    name='score_pool4')

# INFO : Keras sum is a special case of the merge function
summed = merge([sp4(fcn32model.layers[14].output), fcn32model.layers[-1].output], mode='sum')

# INFO :
# deconv setting is valid if (528-32)/16 + 1 = deconv_input_dim (= fcn32size)
deconv_output_size = (fcn32size-1)*16+32 # INFO: =528 when images are 512x512
upnew = Deconvolution2D(21, 32, 32,
                            output_shape=(None, 21, deconv_output_size, deconv_output_size),
                            border_mode='valid', # WARNING : valid, same or full ?
                            subsample=(16, 16),
                            activation=None,
                            name = 'upsample_new')

# INFO : cropping as deconv gained pixels
extra_margin = deconv_output_size - fcn32size*16 # INFO: =16 when images are 512x512
crop_margin = Cropping2D(cropping=((extra_margin/2, extra_margin/2),
                                   (extra_margin/2, extra_margin/2))) 

fcn16model = Model(fcn32model.input, crop_margin(upnew(summed)))