how to control kernel as we need in mxnet

Question

eg: in tensorflow we can do it as followed,how can we do the same thing to control kernel in mxnet,(weights = weights * mask),thank you very much.

if mask_type is not None:
      #C
      mask[:center_h, :, :, :] = 1
      if mask_type == 'A':
        mask[center_h, :center_w, :, :] = 1

      if mask_type == 'B':
        mask[center_h, :center_w+1, :, :] = 1

    else:
      mask[:, :, :, :] = 1

    weights_shape = [kernel_h, kernel_w, in_channel, num_outputs]
    weights = tf.get_variable("weights", weights_shape,
      tf.float32, tf.truncated_normal_initializer(stddev=0.1))
    weights = weights * mask
    biases = tf.get_variable("biases", [num_outputs],
          tf.float32, tf.constant_initializer(0.0))

    outputs = tf.nn.conv2d(inputs, weights, [1, stride_h, stride_w, 1], padding="SAME")
    outputs = tf.nn.bias_add(outputs, biases)

# #########this is the full code by tensorflow

def conv2d(inputs, num_outputs, kernel_shape, strides=[1, 1], mask_type=None, scope="conv2d"):
  with tf.variable_scope(scope) as scope:
    kernel_h, kernel_w = kernel_shape
    stride_h, stride_w = strides
    batch_size, height, width, in_channel = inputs.get_shape().as_list()

    center_h = kernel_h // 2
    center_w = kernel_w // 2

    assert kernel_h % 2 == 1 and kernel_w % 2 == 1, "kernel height and width must be odd number"
    mask = np.zeros((kernel_h, kernel_w, in_channel, num_outputs), dtype=np.float32)
    if mask_type is not None:
      #C
      mask[:center_h, :, :, :] = 1
      if mask_type == 'A':
        mask[center_h, :center_w, :, :] = 1

      if mask_type == 'B':
        mask[center_h, :center_w+1, :, :] = 1

    else:
      mask[:, :, :, :] = 1

    weights_shape = [kernel_h, kernel_w, in_channel, num_outputs]
    weights = tf.get_variable("weights", weights_shape,
      tf.float32, tf.truncated_normal_initializer(stddev=0.1))
    weights = weights * mask
    biases = tf.get_variable("biases", [num_outputs],
          tf.float32, tf.constant_initializer(0.0))

    outputs = tf.nn.conv2d(inputs, weights, [1, stride_h, stride_w, 1], padding="SAME")
    outputs = tf.nn.bias_add(outputs, biases)

    return outputs

score 0 · Answer 1 · answered Mar 14 '18 at 08:34

In MXNet, you can use your own variable for weight by first creating a weight variable and then using that variable for the convolution operator. Example:

weight = mx.sym.Variable('weights', init=mx.initializer.Xavier())
conv1 = mx.sym.Convolution(data=data, weight=weight, kernel=(5,5), num_filter=20)

If you want to apply a mask to the weight like you have done in your code:

mask = #create the mask you want
weight = weight * mask

how to control kernel as we need in mxnet

1 Answers1