recurrent-multilayer.lua

-- Eugenio Culurciello
-- August 2016
-- a test to learn to code PredNet-like nets in Torch7

require 'nn'
require 'nngraph'

torch.setdefaulttensortype('torch.FloatTensor')
nngraph.setDebug(true)

local nlayers = 4

-- local input = nn.Identity()()
-- local pOut = nn.Identity()()
local inputs = {}
local outputs = {}
table.insert(inputs, nn.Identity()()) -- input image x
for L = 1, nlayers do
   table.insert(inputs, nn.Identity()()) -- previous output D
end 

for L = 1, nlayers do
   print('Creating layer-test:', L)

   -- define layer functions:
   local cD = nn.MulConstant(2)
   local cG = nn.MulConstant(0.5)
   local E = nn.CSubTable(1)
   local D
   if L == 1 then 
      D = {inputs[1]} - cD -- output
   else
      D = {outputs[2*L-3]} - cD
   end
   D:annotate{graphAttributes = {color = 'green', fontcolor = 'green'}}
   
   local G = {inputs[L+1]} - cG
   local Df 
   if L == 1 then 
      Df = {inputs[L], G} - E -- output difference
   else
      Df = {outputs[2*L-3], G} - E
   end
   Df:annotate{graphAttributes = {color = 'blue', fontcolor = 'blue'}}
   table.insert(outputs, D)
   table.insert(outputs, Df)
end
-- create graph
print('Creating model-test:')
nngraph.annotateNodes()
local model = nn.gModule(inputs, outputs)

-- test:
print('Testing model-test:')

local c = require 'trepl.colorize'

local nT = 3 -- time sequence length
--local inTable = {}
local outTable = {}
for L = 1, nlayers * 2 do
   table.insert(outTable, torch.zeros(2, 2))
end
local x = {} -- size nT
for t = 1, nT do table.insert(x, torch.ones(2, 2)) end
local tmp
for i = 1, nT do
   
   inTable = {x[i]} -- size (nlayers + 1)
   for j = 1, nlayers do
      table.insert(inTable, outTable[2*j - 1])
   end

   print(c.red('Input of iteration '.. i ..' is:'))

   tmp = inTable[1]
   for j = 2, #inTable do tmp = torch.cat(tmp, inTable[j], 2) end
   print(tmp)
   outTable = model:forward(inTable) -- size 2*nlayers
   print(c.cyan('Output of iteration '.. i ..' is: '))
   tmp = outTable[1]
   for j = 2, #outTable do tmp = torch.cat(tmp, outTable[j], 2) end
   print(tmp)
end
graph.dot(model.fg, 'test','Model-test') -- graph the model!