mwe depthconcat

mwe.lua

require 'cutorch'
require 'cunn'
require 'cudnn'

torch.setnumthreads(1)
cutorch.setDevice(1)
torch.setdefaulttensortype('torch.FloatTensor')

-- model
c1 = nn.Sequential()
c1:add( cudnn.SpatialConvolution(1,10,3,3) )
c2 = nn.Sequential()
c2:add( cudnn.SpatialConvolution(1,10,5,5) )
join = nn.DepthConcat(1)
join:add( c1 )
join:add( c2 )
model = nn.Sequential()
model:add( join )
model:add( nn.SpatialConvolution(20,2,1,1) )
model:add( nn.View( 2,-1) )
model:add( nn.Transpose( {2,1} ) )
model:add( nn.LogSoftMax() )
model:cuda()
weights, gradients = model:getParameters()
loss_criterion = nn.ClassNLLCriterion():cuda()

-- loop
for epoch=1,1e6 do
   gradients:zero()
   data = torch.rand(1,25,25):cuda()
   pred = model:forward( data )
   gt_lin = torch.Tensor( pred:size(1) ):random(2):cuda()
   err = loss_criterion:forward( pred, gt_lin )
   grad_crit = loss_criterion:backward( pred, gt_lin )
   model:backward( data, grad_crit )

   print( string.format('Epoch %d, loss %.3f', epoch, err) )
   epoch = epoch + 1
   collectgarbage()
end