Small example on how to create a binary tree in Torch7 using NN containers and nngraph.

binary_tree_torch.lua

--[[
    Create a binary tree using two ways: NN containers and nn.gModule containers. 
    This example is fairly simple, and the default fully-connected layers are all
    of size 100. However, this should also be simple to modify to allow different 
    fc layers with varying inputs/outputs if desired (for example: input a table
    storing input+output configuration values for each of the sub-branch's level).
]]

require 'nn'
require 'nngraph'

-- (1) Example of a N-dimensional binary tree using containers

-- add tree branches into a container recursively
local function recursiveAddSubtrees(network, n_sub_trees, ninputs, noutputs)
  -- creates a pair of linear layers inside a container
  local function createPairFC(ninputs, noutputs)
    local fc1 = nn.Linear(ninputs,noutputs)
    local fc2 = nn.Linear(ninputs,noutputs)
    local pair = nn.Sequential()
    pair:add(nn.ConcatTable():add(nn.Sequential():add(fc1)):add(nn.Sequential():add(fc2)))
    return pair
  end --local function
  --
  
  if n_sub_trees > 1 then
    local pair_tree = createPairFC(ninputs, noutputs)
    network:add(pair_tree)
    recursiveAddSubtrees(pair_tree.modules[1].modules[1], n_sub_trees-1, ninputs, noutputs)
    recursiveAddSubtrees(pair_tree.modules[1].modules[2], n_sub_trees-1, ninputs, noutputs)
  else
    network:add(createPairFC(ninputs, noutputs))
  end
end --local function

-- create a tree of NN containers
local n_sub_trees = 3 -- number of branching trees
local branches = nn.Sequential() -- Define the container to add sub-trees
recursiveAddSubtrees(branches, 3, 100, 100) -- recursively adds trees of fc layers of 100,100
local bin_tree_model = nn.Sequential() -- main model container
bin_tree_model:add(nn.Linear(10,100)) -- add 'root' fully-connected layer
bin_tree_model:add(branches) -- add branches
print(bin_tree_model) -- print the binary tree


-- (2) Example of a N-dimensional binary tree using nngraph. Warning: requires qlua to display the graph

local function recursiveAddSubtreesGraph(networkGraphTable, root_fc,  n_sub_trees, ninputs, noutputs)
  -- creates a pair of linear layers
  local function createPairFCGraph(ninputs, noutputs)
    local fc1 = nn.Linear(ninputs,noutputs)
    local fc2 = nn.Linear(ninputs,noutputs)
    return fc1, fc2
  end --local function
  --

   if n_sub_trees > 1 then
    local fc1, fc2 = createPairFCGraph(ninputs, noutputs)
    recursiveAddSubtreesGraph(networkGraphTable, fc1({root_fc}), n_sub_trees-1, ninputs, noutputs)
    recursiveAddSubtreesGraph(networkGraphTable, fc2({root_fc}), n_sub_trees-1, ninputs, noutputs)
  else
    local fc1, fc2 = createPairFCGraph(ninputs, noutputs)
    table.insert(networkGraphTable, fc1({root_fc}))
    table.insert(networkGraphTable, fc2({root_fc}))
  end
end --local function

-- create a tree of nngraph.Node
local networkGraphTable = {} -- this table will sotre all the outputs necessary to define the nn.gModule
local root_fc = nn.Linear(10,100)() -- root fc layer. The () at the end is to register the module as a "nngraph.Node"
recursiveAddSubtreesGraph(networkGraphTable, root_fc, 3, 100, 100) -- recursively keep adding pairs of fc layers
-- define the nn.gModule (nngraph)
local bin_tree_modelGraph = nn.gModule(
                                        {root_fc}, -- define the input to the model
                                        networkGraphTable -- define the outputs of the model
                                      )
local ok = pcall(require,'qt')
if ok then
    -- to plot this graph you should use qlua to start this script
  graph.dot(bin_tree_modelGraph.fg, 'binary tree') --display the forward node graph of the binary tree
end