Search code examples
imageimage-processingluatransformtorch

Pad a Tensor in Torch

I have a 100x100 pixel image in a Torch Tensor and I want to implement a "zoom out" transformation. How can I achieve this using the Torch Image toolbox (or other)?

I have already implemented "zoom in" by simply using image.crop followed by image.resize.

In Matlab, I would calculate the mean grayscale of the image, pad the array n pixels with that colour (keeping the original image centred), and then resize to 100x100 pixels. In there a "pad Tensor" function for Torch?

Thanks!

Solution

  • Is there a "pad Tensor" function for Torch?

    One possibility is to use the nn.Padding module from torch/nn, e.g.:

    require 'image'
require 'nn'

local x = image.lena()

local pad  = 64
local pix  = 0
local ndim = x:dim()

local s = nn.Sequential()
  :add(nn.Padding(ndim-1,  pad, ndim, pix))
  :add(nn.Padding(ndim-1, -pad, ndim, pix))
  :add(nn.Padding(ndim,    pad, ndim, pix))
  :add(nn.Padding(ndim,   -pad, ndim, pix))

local y = s:forward(x)

image.display(y) -- this requires qlua

    UPDATE

    As can be seen in the implementation padding is obtained by:

    1. allocating an output tensor of the expected size post-padding filled with the padding color,
    2. filling the area corresponding of the input tensor with the original values thanks to narrow.

    Toy example:

    require 'torch'

local input  = torch.zeros(2, 5)
local dim    = 2 -- target dimension for padding
local pad    = 3 -- amount of padding
local pix    = 1 -- pixel value (color)

-- (1) compute the expected size post-padding, allocate a large enough tensor
--     and fill with expected color
local size   = input:size()
size[dim]    = size[dim] + pad
local output = input.new():resize(size):fill(pix)

-- (2) fill the original area with original values
local area   = output:narrow(dim, 1, input:size(dim)):copy(input)

    This gives as output:

    0  0  0  0  0  1  1  1
0  0  0  0  0  1  1  1
[torch.DoubleTensor of size 2x8]

    For specific zero-padding there are other convenient possibilities like: