Styjun

I'm trying to implement a DeconvNet for visualizing my CNN in order to see what are the features that the different layers are looking for in my network, and for that I need the reverse functions used in the network (like Relu, BatchNormalization).

You can check this paper to understand what I'm trying to do : https://arxiv.org/abs/1311.2901

This is the Deconvolution code that I found on the internet:

Class DConvolution2D(object):

 def __init__(self, layer):

 self.layer = layer

 weights = layer.get_weights()
 W = weights[0]
 b = weights[1]


 filters = W.shape[3]
 up_row = W.shape[0]
 up_col = W.shape[1]
 input_img = keras.layers.Input(shape = layer.input_shape[1:])

 output=keras.layers.Conv2D(filters,(up_row,up_col),kernel_initializer=tf.constant_initializer(W),
 bias_initializer=tf.constant_initializer(b),padding='same')(input_img)
 self.up_func = K.function([input_img, K.learning_phase()], [output])
 # Deconv filter (exchange no of filters and depth of each filter)
 W = np.transpose(W, (0,1,3,2))
 # Reverse columns and rows
 W = W[::-1, ::-1,:,:]
 down_filters = W.shape[3]
 down_row = W.shape[0]
 down_col = W.shape[1]
 b = np.zeros(down_filters)
 input_d = keras.layers.Input(shape = layer.output_shape[1:])

 output=keras.layers.Conv2D(down_filters,(down_row,down_col),kernel_initializer=tf.constant_initializer(W),
 bias_initializer=tf.constant_initializer(b),padding='same')(input_d)
 self.down_func = K.function([input_d, K.learning_phase()], [output])

 def up(self, data, learning_phase = 0):
 #Forward pass
 self.up_data = self.up_func([data, learning_phase])
 self.up_data=np.squeeze(self.up_data,axis=0)
 self.up_data=numpy.expand_dims(self.up_data,axis=0)
 #print(self.up_data.shape)
 return self.up_data

 def down(self, data, learning_phase = 0):
 # Backward pass
 self.down_data= self.down_func([data, learning_phase])
 self.down_data=np.squeeze(self.down_data,axis=0)
 self.down_data=numpy.expand_dims(self.down_data,axis=0)
 #print(self.down_data.shape)
 return self.down_data

So I'm looking to do the same with the other function on the architecture of YOLO.

Thanx for helping, and sorry for my english if I wasn't clear