Using ConcatTable in Recurrent

[ethanabrooks]

Recurrent appears to mutate ConcatTable.modules into {}. As a result the following code:

require 'rnn'
require 'nn'

-- hyper-parameters
local batchSize = 3
local hiddenSize = 2
local x = torch.ones(batchSize, hiddenSize)

local feedback = nn.Sequential()
local concat = nn.ConcatTable()
feedback:add(concat)
concat:add(nn.Identity)
concat:add(nn.Identity)
feedback:add(nn.CAddTable())

local r = nn.Recurrent(
    nn.Identity(), -- start
    nn.Identity(), -- input
    feedback,
    nn.Identity() --transfer
)
print(r)
r:updateOutput(x)
r:updateOutput(x)
r:updateOutput(x)

throws this error:

tput: No value for $TERM and no -T specified
/Users/Ethan/torch/install/bin/luajit: /Users/Ethan/torch/install/share/lua/5.1/nn/Container.lua:67: 
In 1 module of nn.Sequential:
In 2 module of nn.ParallelTable:
In 1 module of nn.Sequential:
In 1 module of nn.ConcatTable:
attempt to call a nil value -- this, I believe refers to the modules of ConcatTable
stack traceback:
    [C]: in function 'xpcall'
    /Users/Ethan/torch/install/share/lua/5.1/nn/Container.lua:63: in function 'rethrowErrors'
    /Users/Ethan/torch/install/share/lua/5.1/nn/ConcatTable.lua:11: in function </Users/Ethan/torch/install/share/lua/5.1/nn/ConcatTable.lua:9>
    [C]: in function 'xpcall'
    /Users/Ethan/torch/install/share/lua/5.1/nn/Container.lua:63: in function 'rethrowErrors'
    /Users/Ethan/torch/install/share/lua/5.1/nn/Sequential.lua:44: in function </Users/Ethan/torch/install/share/lua/5.1/nn/Sequential.lua:41>
    [C]: in function 'xpcall'
    /Users/Ethan/torch/install/share/lua/5.1/nn/Container.lua:63: in function 'rethrowErrors'
    ...s/Ethan/torch/install/share/lua/5.1/nn/ParallelTable.lua:12: in function <...s/Ethan/torch/install/share/lua/5.1/nn/ParallelTable.lua:10>
    [C]: in function 'xpcall'
    /Users/Ethan/torch/install/share/lua/5.1/nn/Container.lua:63: in function 'rethrowErrors'
    /Users/Ethan/torch/install/share/lua/5.1/nn/Sequential.lua:44: in function 'updateOutput'
    /Users/Ethan/torch/install/share/lua/5.1/rnn/Recurrent.lua:75: in function 'updateOutput'
    issue.lua:32: in main chunk
    [C]: at 0x010923fbd0

nn.Recurrent {
  [{input(t), output(t-1)} -> (1) -> (2) -> (3) -> output(t)]
  (1):  {
    input(t)
      |`-> (t==0): nn.Identity
      |`-> (t~=0): nn.Identity
    output(t-1)
      |`-> nn.CAddTable
  }
  (2): nn.CAddTable
  (3): nn.ConcatTable {
    input
      |`-> (1): table: 0x0bffa7c0
      |`-> (2): table: 0x0bffa7c0
       ... -> output
  }
}

A few observations. Recurrent only throws the error on the third call. Also, feedback does not throw the error outside Recurrent. Thanks for our help!

[nicholas-leonard]

Use nn.Copy to force a copy:

local feedback = nn.Sequential()
local concat = nn.ConcatTable()
feedback:add(concat)
concat:add(nn.Copy(nil, nil, true))
concat:add(nn.Copy(nil, nil, true))
feedback:add(nn.CAddTable())

[ethanabrooks]

Thanks! That fixed it. Could you help me understand what's going on here? nn.Recurrent is a module that I plan to use extensively.

[nicholas-leonard]

The problem is that nn.Recurrent assumes that the output and input are different tensors. That is why you must explicitly copy input to output if the tensor is the same.

[ethanabrooks]

Thanks. That's very clarifying.