Minimal working example of something that doesn't work with nn.Concat.

gistfile1.lua

#!/usr/bin/env th

require 'nn';

local cmd = torch.CmdLine()
cmd:text()
cmd:text("What's wrong with this use of nn.Concat?")
cmd:text('Options:')
cmd:option('-noConcat', false, 'do not include concat layer')

local opt = cmd:parse(arg or {})

-- Mock up a training example with regression target 3.
local target = torch.Tensor(1):fill(3)
local input = torch.Tensor({{ 1,2,4,1 }}):resize(4)

-- A lookup table layer, output size 6.
local lookupSeq = nn.Sequential()
lookupSeq:add(nn.LookupTable(5, 6))
lookupSeq:add(nn.Sum(1))

-- A concatenation layer, output size 6.
local concat = nn.Concat(1)
if opt.noConcat then
    concat = lookupSeq
else
    -- Normally we'd add more than one module to the concat layer.
    -- This is just to demonstrate the error.
    concat:add(lookupSeq)
end

-- A simple regression model.
local model = nn.Sequential()
model:add(concat)
model:add(nn.Linear(6, 1))

-- Do a forward and backward pass.
local output = model:forward(input)
local criterion = nn.MSECriterion()
local err = criterion:forward(output, target)
local df_do = criterion:backward(output, target)

--[[
The backward pass on the model results in this error.

luajit: nn/Concat.lua:73: inconsistent tensor size at torch/pkg/torch/lib/TH/generic/THTensorCopy.c:7
stack traceback:
    [C]: in function 'copy'
    nn/Concat.lua:73: in function 'backward'
    nn/Sequential.lua:73: in function 'backward'
    ./repro.lua:31: in main chunk
    [C]: in function 'dofile'
    th:131: in main chunk
    [C]: at 0x0100e112d0
--]]
model:backward(input, df_do)