dense_net_parts.py

def composite_function(self, _input, out_features, kernel_size=3):
    """Function from paper H_l that performs:
    - batch normalization
    - ReLU nonlinearity
    - convolution with required kernel
    - dropout, if required
    """
    with tf.variable_scope("composite_function"):
        # BN
        output = self.batch_norm(_input)
        # ReLU
        output = tf.nn.relu(output)
        # convolution
        output = self.conv2d(
            output, out_features=out_features, kernel_size=kernel_size)
        # dropout(in case of training and in case it is no 1.0)
        output = self.dropout(output)
    return output

def conv2d(self, _input, out_features, kernel_size,
         strides=[1, 1, 1, 1], padding='SAME'):
  in_features = int(_input.get_shape()[-1])
  kernel = self.weight_variable_msra(
      [kernel_size, kernel_size, in_features, out_features],
      name='kernel')
  output = tf.nn.conv2d(_input, kernel, strides, padding)
  return output

 def batch_norm(self, _input):
    output = tf.contrib.layers.batch_norm(
        _input, scale=True, is_training=self.is_training,
        updates_collections=None)
    return output

  
_input = 'some_3_dims'
layers = 3
for layer in range(layers):
  l_output = composite_function(_input, layer_out_features)
  _input = tf.concat(3, (_input, l_output))