complex_net.lua

require 'nn'

--[[
This file contains the code definition of our complex example network in raw Torch
using nn.Table modules.  Since nngraph is not used, this network is defined in
reverse order in terms of sequential segments and branches.
]]--


-- the final layer of our base network
net3 = nn.Sequential()
net3:add(nn.SpatialConvolution(128,256,3,3,1,1,1,1))
net3:add(nn.View(-1))

-- split the last branch
branch2b = nn.ParallelTable()
branch2b:add(nn.SpatialConvolution(128,256,3,3,1,1,1,1))
branch2b:add(nn.SpatialConvolution(128,256,3,3,1,1,1,1))

-- here we add max pooling and use nn.Replicate() --> nn.SplitTable
-- to allow for a ParallelTable to split the branch
branch2a = nn.Sequential()
branch2a:add(nn.SpatialMaxPooling(2,2,2,2))
branch2a:add(nn.Replicate(4))
branch2a:add(nn.SplitTable(1))
branch2a:add(branch2b)

-- second (final) branch from base of network
branch2 = nn.ParallelTable()
branch2:add(branch2a)
branch2:add(net3)

-- using nn.Sequential() for part 2 of base network
net2 = nn.Sequential()
net2:add(nn.SpatialConvolution(64,128,3,3,1,1,1,1))
net2:add(nn.ReLU())
net2:add(nn.SpatialConvolution(128,128,3,3,1,1,1,1))
net2:add(nn.ReLU())
net2:add(nn.Replicate(4))
net2:add(nn.SplitTable(1))
net2:add(branch2)

-- split from first branch
branch1b = nn.ParallelTable()
branch1b:add(nn.SpatialConvolution(64,64,3,3,1,1,1,1))
branch1b:add(nn.SpatialConvolution(64,64,3,3,1,1,1,1))

-- add a max pooling layer to branch and then prepare to split
branch1a = nn.Sequential()
branch1a:add(nn.SpatialMaxPooling(2,2,2,2))
branch1a:add(nn.Replicate(4))
branch1a:add(nn.SplitTable(1))
branch1a:add(branch1b)

-- first branch from the base of the network
branch1 = nn.ParallelTable()
branch1:add(branch1a)
branch1:add(net2)

-- After we flatten all the nested tables, we will 'vectorize' all the tensors
prl = nn.ParallelTable()
prl:add(nn.View(-1))
prl:add(nn.View(-1))
prl:add(nn.View(-1))
prl:add(nn.View(-1))
prl:add(nn.View(-1))


-- the first layers of the network
net = nn.Sequential()
net:add(nn.SpatialConvolution(3,64,3,3,1,1,1,1))
net:add(nn.ReLU())
net:add(nn.SpatialConvolution(64,64,3,3,1,1,1,1))
net:add(nn.ReLU())
net:add(nn.Replicate(4))
net:add(nn.SplitTable(1))
net:add(branch1)

-- now need to join output from branch 1, branch 2, and net3
net:add(nn.FlattenTable())
net:add(prl)
net:add(nn.ConcatTable())