deconv.lua

require 'nn'

torch.setdefaulttensortype("torch.FloatTensor")

upconv = nn.SpatialFullConvolution(1, 1, 2, 2, 2, 2, 0, 0)
input = torch.Tensor({{1, 10},{100, 1000}}):reshape(1, 1, 2, 2) -- (batch_size, input_dim, height, width)
weight = torch.Tensor({{1, 2}, {3, 4}}) -- 1x1x2x2 filter
bias = torch.Tensor({0.5})
upconv.weight:copy(weight)
upconv.bias:copy(bias)
print(upconv:forward(input))

--[[ output
   (1,1,.,.) = 
     1.5000     2.5000    10.5000    20.5000
     3.5000     4.5000    30.5000    40.5000
   100.5000   200.5000  1000.5000  2000.5000
   300.5000   400.5000  3000.5000  4000.5000
   [torch.FloatTensor of size 1x1x4x4]
--]]