python - Custom padding for convolutions in TensorFlow

Question

In tensorflow function tf.nn.conv2d, the padding option just has 'SAME' and 'VALID'.

But in the conv layer of Caffe, there is pad option can define the number of pixels to (implicitly) add to each side of the input.

How to achieve that in Tensorflow?

Thank you very much.

Answer 1

You can use tf.pad() (see the doc) to pad the Tensor before applying tf.nn.conv2d(..., padding="VALID") (valid padding means no padding).

---

For instance, if you want to pad the image with 2 pixels in height, and 1 pixel in width, and then apply a convolution with a 5x5 kernel:

input = tf.placeholder(tf.float32, [None, 28, 28, 3])
padded_input = tf.pad(input, [[0, 0], [2, 2], [1, 1], [0, 0]], "CONSTANT")

filter = tf.placeholder(tf.float32, [5, 5, 3, 16])
output = tf.nn.conv2d(padded_input, filter, strides=[1, 1, 1, 1], padding="VALID")

output will have shape [None, 28, 26, 16], because you have only a padding of 1 in width.