Build network with shortcut using torch - deep-learning

Question

I now have a network with 2 inputs X and Y.
X concatenates Y and then pass to network to get result1. And at the same time X will concat result1 as a shortcut.
It's easy if there is only one input.
branch = nn.Sequential()
branch:add(....) --some layers
net = nn.Sequential()
net:add(nn.ConcatTable():add(nn.Identity()):add(branch))
net:add(...)
But when it comes to two inputs I don't actually know how to do it? Besides, nngraph is not allowed.Does any one know how to do it?

Answer 1

You can use the table modules, have a look at this page: https://github.com/torch/nn/blob/master/doc/table.md
net = nn.Sequential()
triple = nn.ParallelTable()
duplicate = nn.ConcatTable()
duplicate:add(nn.Identity())
duplicate:add(nn.Identity())
triple:add(duplicate)
triple:add(nn.Identity())
net:add(triple)
net:add(nn.FlattenTable())
-- at this point the network transforms {X,Y} into {X,X,Y}
separate = nn.ConcatTable()
separate:add(nn.SelectTable(1))
separate:add(nn.NarrowTable(2,2))
net:add(separate)
-- now you get {X,{X,Y}}
parallel_XY = nn.ParallelTable()
parallel_XY:add(nn.Identity()) -- preserves X
parallel_XY:add(...) -- whatever you want to do from {X,Y}
net:add(parallel)
parallel_Xresult = nn.ParallelTable()
parallel_Xresult:add(...) -- whatever you want to do from {X,result}
net:add(parallel_Xresult)
output = net:forward({X,Y})
The idea is to start with {X,Y}, to duplicate X and to do your operations. This is clearly a bit complicated, nngraph is supposed to be here to do that.