There should be a method to reduce an edge in a graph, such that it continues to render correctly, and the graph data remains self-consistent.

Thanks for your attention
I want to share the parameters in some layers when using the nngraph to construct a net.please help me

Couldn't find a mailing list for nngraph so I'll ask here. Is there any speed hit when using nngraph to specify a model over vanilla nn? Also, do these models port to the GPU like NN models do?

If arguments have different types, the module will break. Tested with : {torch.DoubleTensor, torch.FloatTensor} and {torch.DoubleTensor, torch.CudaTensor}.

Hello,
it seems nobody get problem to install it but I've try on linux and mac and got a similar problem : nngraph can't find graphiz.
nngraph is installed with :
luarocks install nngraph
and graphiviz with :
sudo apt-get install graphviz -y
and brew command on osx

on linux :
/torch/install/share/lua/5.1/graph/graphviz.lua:145: attempt to call field 'gvContext' (a nil value)
I have this error message
on mac it's another message : it can't load libgvplugin_dot_layout.6.dylib
is the installation sequence is important ? luarocks before graphviz ?

best regards

Hi,
Is it possible to have split(n) working with n=1 ? That would help in generalising part of a model.
Thanks

There seems to be a bug in the way I'm writing my LSTM architecture.

local LSTM = {}
function LSTM.create(input_size, input2_size, output_size, rnn_size)
    local inputs = {}
    local outputs = {}

    table.insert(inputs, nn.Identity()())
    table.insert(inputs, nn.Identity()())
    for L=1,2 do
        table.insert(inputs, nn.Identity()())
        table.insert(inputs, nn.Identity()())
    end

    local x, x_p, x2
    for L=1,2 do
        if L == 1 then
            x_p = inputs[1]
            input_size_L = input_size
        else 
            x = outputs[(L-2)*2]
            x2 = nn.LookupTable(input2_size+1, rnn_size)(inputs[2])
            x_p = nn.JoinTable(2)({x, x2})
            input_size_L = rnn_size*2
        end

        prev_c = inputs[L*2+1]
        prev_h = inputs[L*2+2]

        local i2h = nn.Linear(input_size_L, 4 * rnn_size)(x_p)
        local h2h = nn.Linear(rnn_size, 4 * rnn_size)(prev_h)
        local all_input_sums = nn.CAddTable()({i2h, h2h})

        local reshaped = nn.Reshape(4, rnn_size)(all_input_sums)
        local n1, n2, n3, n4 = nn.SplitTable(2)(reshaped):split(4)

        local in_gate = nn.Sigmoid()(n1)
        local forget_gate = nn.Sigmoid()(n2)
        local out_gate = nn.Sigmoid()(n3)

        local in_transform = nn.Tanh()(n4)

        local next_c           = nn.CAddTable()({
            nn.CMulTable()({forget_gate, prev_c}),
            nn.CMulTable()({in_gate,     in_transform})
          })

        local next_h = nn.CMulTable()({out_gate, nn.Tanh()(next_c)})

            table.insert(outputs, next_c)
            table.insert(outputs, next_h)
    end

    local last_h = outputs[#outputs]
    local proj = nn.Linear(rnn_size, output_size)(last_h)
    local logsoft = nn.LogSoftMax()(proj)
    table.insert(outputs, logsoft)

    return nn.gModule(inputs, outputs)
end

return LSTM

The problem seems to be when I try to use nn.JoinTable. What I'm trying to do here is send inputs[2] through a lookuptable and concatenate that (column wise) with the hidden state output of the previous layer.

The hidden state output will be: batchSize x rnnSize. The lookuptable output will be batchSize x rnnSize. If I concatenate these tensors, the new tensor will be batchSize x 2*rnnSize.

Regardless, I'm getting a "expecting only one start" issue, which probably means I'm doing something fundamentally wrong.

Thanks!

• seems not to be installable via luarocks? I tried various variations of luarocks install nngraph, but none seemed to work for me? Searching at https://luarocks.org/search?q=nngraph seems not to provide any results?

Crazy idea: make an nngraph 'optimizer'.

I'm calling it crazy so I dont accidentally commit myself to doing something I then find wont work for reason x,y,z :-P Plus, it is a little way out there.

Thinking of creating an optimizer for nngraph, that takes a graph as input, and then culls edges where it can. eg, maybe replaces nn.Narrow with torch.Narrow, and removes the nn.Narrow node. Since I haven't actually implemented this yet, I dont know how possible this is going to be, and whether obstacles I meet are merely challenging, or are weeks of work.

I'm also thinking of something more general, and more gpu-specific, of implementing gmodule within clnn, and walking the graph within clnn, and joining some of the nodes together, during the forward/backward calls, eg replace a Linear + a Tanh by a single kernel that calls both, in one launch.

Overriding motivatoin for all of this is basically on certain gpus, I notice that the overhead of launching kernels, in char-rnn, appears to dominate the runtime. I havent used any particular diagnostic tools to confirm this, and perhaps I should, but I notice that I can increase the data size 20 times whilst the execution time only increases by aobut 10-20%, which suggests to me that it's not a calc/data issue, but plausibly linked to kernel launches.

nodes should have parent link. would make certain transformations much easier (or rather: possible. the easiest way to achieve them is precisely to add the parents, or to create an associative map of nodes to parents).

I have two outputs.
And I pass in two gradOutputs.
The nngraph complains:

#outputs   =1   
#gradients =2
nngraph/gmodule.lua:251: Number of gradients do not match my graph

Example code to reproduce the error:

local in1 = nn.Sigmoid()()
local splitTable = nn.SplitTable(1)({in1})
local module = nn.gModule({in1}, {splitTable})

local input = torch.randn(2, 3)
local output = module:forward(input)
assert(#output == 2, "we have two outputs")
module:backward(input, {torch.randn(3), torch.randn(3)})

Is there an easy way to remove the last node or two from an architecture? I feel like this would be a fairly common operation, especially when adding a different classifier to fine-tune a model.

I am working on creating an nngraph module with multiple outputs. Per https://github.com/torch/nngraph/blob/master/gmodule.lua#L31 , this is not currently supported. Thoughts on:

• reasons/obstacles for only one output allowed currently?
• things that will likely be tricky to do, need careful consideration, in order to generalize nngraph to handle multiple outputs?

hey,

i am curious to know why the case of #input == 1 is handled separately
e.g.
https://github.com/torch/nngraph/blob/master/gmodule.lua#L195
and for the backward:
https://github.com/torch/nngraph/blob/master/gmodule.lua#L266
https://github.com/torch/nngraph/blob/master/gmodule.lua#L277

if i understand correctly, this also impacts the split, which cannot be done on tables of size 1:
https://github.com/torch/nngraph/blob/master/node.lua#L50

could you give me an idea why you made this choice?

thanks!

ronan

I'd like to be able to build graphs that take no inputs during the forward pass, and build them cleanly, without unused bogus modules or having to pass placeholder bogus tensors... This example script shows what I mean...

Hi, consider the following example to illustrate this issue. It creates a gModule
and converts it to CUDA and back to a float.
After converting back to a float, I think some internal state isn't correctly updated
and I'm getting the error below.

I'm using the latest master branch of:

• torch f62a95d3184fd730acdaa4754647b338d7686301
• cutorch a7147d00e61a5e182a277995f5d1e99ec3bdf0f8
• nn bc056eeb09f83aaba354d44b985b1819b6b6ee4a
• cunn 3827fcd820d5d0d90cb37a443c403b47009cb7d4
• nngraph d0c239b

require 'cutorch'
require 'nn'
require 'cunn'
require 'nngraph'
torch.manualSeed(1)
input = nn.Identity()()
L1 = nn.ReLU()(nn.Linear(3, 1)(input))
net = nn.Sequential()
net:add(L1)
g = nn.gModule({input}, {L1})
x = torch.randn(3)
g:forward(x)
g:cuda()
g:forward(x:cuda())
g:float()
g:forward(x)

Output

th> g:forward(x)
 0.1432
[torch.DoubleTensor of size 1]

                                                                      [0.0001s]
th> g:cuda()
nn.gModule
                                                                      [0.0596s]
th> g:forward(x:cuda())
 0.1432
[torch.CudaTensor of size 1]

                                                                      [0.0003s]
th> g:float()
nn.gModule
                                                                      [0.0004s]
th> g:forward(x)
/home/bamos/torch/install/share/lua/5.1/nn/Linear.lua:51: expected arguments: *FloatTensor~1D* [FloatTensor~1D] [float] FloatTensor~2D FloatTensor~1D | *FloatTensor~1D* float [FloatTensor~1D] float FloatTensor~2D FloatTensor~1D
stack traceback:
        [C]: in function 'addmv'
        /home/bamos/torch/install/share/lua/5.1/nn/Linear.lua:51: in function 'func'
        /home/bamos/torch/install/share/lua/5.1/nngraph/gmodule.lua:311: in function 'neteval'
        /home/bamos/torch/install/share/lua/5.1/nngraph/gmodule.lua:346: in function 'forward'
        [string "_RESULT={g:forward(x)}"]:1: in main chunk
        [C]: in function 'xpcall'
        /home/bamos/torch/install/share/lua/5.1/trepl/init.lua:630: in function 'repl'
        ...amos/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:185: in main chunk
        [C]: at 0x00406670

For example, I could get a 2-layer MLP with Container or nngraph, will it be a little bit faster when I use nngraph?

Running this (from the example)

require 'nngraph'

h1 = nn.Linear(20, 10)()
h2 = nn.Linear(10, 1)(nn.Tanh()(nn.Linear(10, 10)(nn.Tanh()(h1))))
mlp = nn.gModule({h1}, {h2})

x = torch.rand(20)
dx = torch.rand(1)
mlp:updateOutput(x)
mlp:updateGradInput(x, dx)
mlp:accGradParameters(x, dx)

-- draw graph (the forward graph, '.fg')
graph.dot(mlp.fg, 'MLP', '/home/yongfei')

Got this error.

.. /lua/5.1/graph/graphviz.lua:156: wrong number of arguments for function call

stack traceback:
    [C]: in function 'gvRender'
    ...e/yongfei/torch/install/share/lua/5.1/graph/graphviz.lua:156: in function 'graphvizFile'
    ...e/yongfei/torch/install/share/lua/5.1/graph/graphviz.lua:181: in function 'dot'
    train.lua:154: in main chunk
    [C]: in function 'dofile'
    ...gfei/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:133: in main chunk

What have I missed?

Is there any neat way to debug nngraph? Thanks!

Currently I write a Debug layer which is almost the same as Identity layer except that I can print whatever I want inside updateOutput() and updateGradInput()

I'm trying to install the nngraph rock on my university high performance computing cluster, on which I don't have root access. I tried to install it in the local tree with:

luarockt --local install nngraph

but the installation fails with the following error:

Install the project...
-- Install configuration: "Release"
-- Old export file "/software/torch/share/cmake/torch/TorchExports.cmake" will be replaced. Removing files [/software/torch/share/cmake/torch/TorchExports-release.cmake].
-- Installing: /software/torch/share/cmake/torch/TorchExports.cmake
CMake Error at cmake_install.cmake:48 (file):
file INSTALL cannot copy file
"/scratch/tmp/luarocks_torch-scm-1-5519/torch7/build/CMakeFiles/Export/share/cmake/torch/TorchExports.cmake"
to "/software/torch/share/cmake/torch/TorchExports.cmake".

Any suggestion?

Following script crashes, with error nesting.lua:36: bad argument #1 to 'resizeAs':

require 'nngraph'

x = nn.Identity()()
m1 = nn.Linear(5, 2)(x)
m2 = nn.Linear(5, 2)(x)
m = nn.JoinTable(1)({m1, m2})

g = nn.gModule({x}, {m})

input = torch.DoubleTensor(5):uniform()
gradOutput = torch.DoubleTensor(4):uniform()

g:forward(input)
g:updateGradInput(input, gradOutput)

g:float()
g:forward(input:float())
g:updateGradInput(input:float(), gradOutput:float())

Result:

$ th /norep/dev/issue-61.lua
/home/user/torch/install/bin/luajit: /home/user/torch/install/share/lua/5.1/nngraph/nesting.lua:36: bad argument #1 to 'resizeAs' (torch.DoubleTensor expected, got torch.FloatTensor)
stack traceback:
    [C]: in function 'resizeAs'
    /home/user/torch/install/share/lua/5.1/nngraph/nesting.lua:36: in function 'resizeNestedAs'
    /home/user/torch/install/share/lua/5.1/nngraph/gmodule.lua:14: in function 'getTotalGradOutput'
    /home/user/torch/install/share/lua/5.1/nngraph/gmodule.lua:300: in function 'neteval'
    /home/user/torch/install/share/lua/5.1/nngraph/gmodule.lua:346: in function 'updateGradInput'
    /norep/dev/issue-61.lua:18: in main chunk
    [C]: in function 'dofile'
    ...user/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:131: in main chunk
    [C]: at 0x00406670

Would be nice to have an example for how to use multiple criteria on a network graph (e.g. MSE on the last layer, and L1 on the activations of the first hidden layer).

Sorry I didn't find related documentations. I would also apologize if github issues is not a good place to ask such questions (e.g if there is a google group for such discussions)

What I'd like to do is something similar to:

m = nn.Sequential()
m:add(nn.Linear(5,5))
m:get(1)

I have a rnn.GRU module in my gModule, and I have to manually call forget() function and change self.userPrevOutput sometimes. So it would be great if I can get access to intermediate node. Otherwise, I have to move rnn.GRU out of gModule :(

Thanks in advance!

How do you get graphviz to display gradOutput?

It does'nt display automatically on my backward graphs?

I've done something really stupid, and I can't backprop ??? 👎

Hi

I can't manage to get this package to work. the error comes up when require 'qtsvg' is being used.
the error is as follows:

require 'qtsvg'
/usr/local/share/lua/5.1/qtsvg/init.lua:2: module 'qt' not found:No LuaRocks module found for qt
no field package.preload['qt']
no file '/usr/local/share/lua/5.1/qt.lua'
no file '/usr/local/share/lua/5.1/qt/init.lua'
no file '/home/siavash/.luarocks/share/lua/5.1/qt.lua'
no file '/home/siavash/.luarocks/share/lua/5.1/qt/init.lua'
no file '/usr/local/share/lua/5.1/qt.lua'
no file '/usr/local/share/lua/5.1/qt/init.lua'
no file '/home/siavash/.luarocks/share/lua/5.1/qt.lua'
no file '/home/siavash/.luarocks/share/lua/5.1/qt/init.lua'
no file '/usr/local/share/lua/5.1/qt.lua'
no file '/usr/local/share/lua/5.1/qt/init.lua'
no file './qt.lua'
no file '/usr/local/share/luajit-2.0.2/qt.lua'
no file '/usr/local/share/lua/5.1/qt.lua'
no file '/usr/local/share/lua/5.1/qt/init.lua'
no file '/usr/local/lib/lua/5.1/qt.so'
no file '/home/siavash/.luarocks/lib/lua/5.1/qt.so'
no file '/usr/local/lib/lua/5.1/qt.so'
no file '/home/siavash/.luarocks/lib/lua/5.1/qt.so'
no file './qt.so'
no file '/usr/local/lib/lua/5.1/qt.so'
no file '/usr/local/lib/lua/5.1/loadall.so'
stack traceback:
[C]: in function 'require'
/usr/local/share/lua/5.1/qtsvg/init.lua:2: in main chunk
[C]: in function 'f'
[string "local f = function() return require 'qtsvg' e..."]:1: in main chunk
[C]: in function 'xpcall'
/usr/local/share/lua/5.1/trepl/init.lua:553: in function </usr/local/share/lua/5.1/trepl/init.lua:489>

what can be the problem?

I'm writing an nngraph module with multiple non-identical outputs. Do we have any existing examples of modules which do this? As I remember, it is a principle of nngraph in general that a module can only produce one unique output?

Update: line 31 of gmodule.lua asserts each output can only have one output https://github.com/torch/nngraph/blob/master/gmodule.lua#L31 , so I guess the answer is: no, would need modification of gmodule, and socialization of why one might want do do that

=> closing, for now

Hi,

I am trying to install nngraph locally using $ luarocks install --local nngraph. I was surprised to see that it said that I had missing dependencies for nngraph, even though I could seemly import them with require.

Missing dependencies for nngraph:
torch >= 7.0
nn
graph

After trying to build I end up with the following message.

Install the project...
-- Install configuration: "Release"
-- Installing: /usr/share/torch7-dist/install/share/cmake/torch/TorchExports.cmake
CMake Error at cmake_install.cmake:48 (FILE):
 file INSTALL cannot copy file
 "/tmp/luarocks_torch-scm-1-446/torch7/build/CMakeFiles/Export/share/cmake/torch/TorchExports.cmake"
 to "/usr/share/torch7-dist/install/share/cmake/torch/TorchExports.cmake".
make: *** [install] Error 1

Error: Failed installing dependency: https://raw.githubusercontent.com/torch/rocks/master/torch-scm-1.rockspec - Build error: Failed installing.

Would appreciate any help.

when I run the code in terminal under 'th' with the following codes, I got the Segmentation fault error.

nn.CAddTable, when used with nngraph, throws a strange error when given a table of size 1.

Here's a minimal test case:

require 'nn'
require 'nngraph'

N = 1
input = nn.Identity()()

hiddens = {}
for i = 1, N do
    table.insert(hiddens, nn.Identity()(input))
end
output = nn.CAddTable()(hiddens)
net = nn.gModule({input}, {output})
net:forward(torch.Tensor(1))

If you change N to be greater than 1, this works fine. For N = 1, it throws the error

...lua/5.1/nn/CAddTable.lua:10: bad argument #1 to 'resizeAs' (torch.DoubleTensor expected, got userdata)
stack traceback:
        [C]: in function 'resizeAs'
        ...lua/5.1/nn/CAddTable.lua:10: in function 'func'
        ...lua/5.1/nngraph/gmodule.lua:253: in function 'neteval'
        ...lua/5.1/nngraph/gmodule.lua:288: in function 'forward'

This seems like undesirable / unexpected behaviour so I thought I'd bring it up, though maybe it is an intentional consequence of #46.

Hi guys,

This crashes :

require 'nngraph';

n1 = 3
n2 = 4
n3 = 3

x1 = nn.Identity()()
x23 = nn.Identity()()
x2,x3 = x23:split(2)
z  = nn.JoinTable(1)({x1,x2,x3})
y1 = nn.Linear(n1+n2+n3,n2)(z)
y2 = nn.Linear(n1+n2+n3,n3)(z)
m = nn.gModule({x1,x23},{y1,y2})

input = {torch.randn(n1), {torch.randn(n2), torch.randn(n3)}}
output = m:forward(input)
print(output)
print(input)
input[2] = output
print(input)
m:forward(input)

The error:

/usr/local/bin/luajit: /usr/local/share/lua/5.1/nngraph/gmodule.lua:314: split(2) cannot split 0 outputs
stack traceback:
    [C]: in function 'error'
    /usr/local/share/lua/5.1/nngraph/gmodule.lua:314: in function 'neteval'
    /usr/local/share/lua/5.1/nngraph/gmodule.lua:346: in function 'forward'
    issues/issue172.lua:22: in main chunk
    [C]: in function 'dofile'
    /usr/local/lib/luarocks/rocks/trepl/scm-1/bin/th:131: in main chunk
    [C]: at 0x00405e60

Basically, the issue happens when feeding back outputs of a previous forward as inputs to the next forward to a gModule using split.

Torch is installed. I was installing everything by this rnn guide. Would appreciate any help.
`sudo luarocks install nngraph

Error: No results matching query were found`

Just wanted to point out that gModules do not configure their output by default (before the call to forward), unlike nn.Module which set "self.output = torch.Tensor()". This can be an unfortunate cause of bugs when nn.Modules are swapped out for gModules. See code below for example.

require 'nn'
require 'nngraph'

local mod = nn.Module()
print(mod.output)

local input = nn.Identity()()
local output = nn.Linear(10,20)(input)
local gmod = nn.gModule({input},{output})
print(gmod.output)

which outputs

[torch.DoubleTensor with no dimension]

nil

The documentation referenced at https://www.cs.ox.ac.uk/people/nando.defreitas/machinelearning/practicals/practical5.pdf suggests installing nngraph using LuaRocks

$ luarocks install nngraph

However, when I run this I get

[~]$ luarocks install nngraph
Warning: falling back to curl - install luasec to get native HTTPS support

Error: No results matching query were found.

Additionally, if you try searching for the package at https://luarocks.org/ it's not found.

What if I build a model named my_module by using nngraph submodules, that is, In the model, other than naive node, there is also a nngraph module type node, however my_module.forwardnodes will not return the the node directly in the nngraph submodule for me to access. I try to solved it recursively tranverse the table of nodes. Is there some clean way to do that? hope someone give me tips

I noticed that feeding tables into nngraph nodes has inconsistent behavior. If I feed a node a table with multiple elements it treats the table as a table, however, if I feed it a table with a single element, it pops that element. Is this intentional? Here's a minimal example showing how it can cause trouble:

The module:

id1 = nn.Identity()()
id2 = nn.Identity()()
j = nn.JoinTable(1)({id1, id2})
net = nn.gModule({id1, id2}, {j})

behaves as expect when the input is a pair of tensors, producing their join. However:

id = nn.Identity()()
j = nn.JoinTable(1)({id})
net = nn.gModule({id}, {j})

produces an error in JoinTable (the input is not a table) because it pops the the value of id from the table before feeding it to JoinTable. I would like it to feed {id.output} into JoinTable in the latter case, just as it's feeding {id1.output, id2.output} in the former.

node ids should be on the data rather than on the node, so that they are invariant.

eg, currently node ids in fg and node ids in bg are different, I think?

Hi,

I'm trying to plot the graph for a NN, in debug mode, but I get this error:

.../torch/install/share/lua/5.1/graph/graphviz.lua:167: dlsym(RTLD_DEFAULT, _fopen): symbol not found stack traceback: [C]: in function '__index' .../torch/install/share/lua/5.1/graph/graphviz.lua:167: in function 'graphvizFile' ...e/torch/install/share/lua/5.1/graph/graphviz.lua:195: in function 'dot' ./Source/NN.lua:125: in function 'opfunc'

The nn is this:
question = nn.Identity()() local target = nn.Identity()(question) self.model = nn.gModule({question},{target})

The graph does plot when I use more complicated NNs (with two inputs, for example)

I installed graphviz through brew, nngraph through luarocks.
I am on OS X El Capitan v 10.11.4

How do you guys do type casting? The current :type() method seems very partial; running the following code prints out lots of DoubleTensors.

setprintlevel(20)

require 'nngraph'

local i1 = nn.Identity()()
local i2 = nn.Identity()()

local o1 = nn.Tanh()( nn.Linear(10,10)(i1) )
local o2 = nn.Tanh()( nn.Linear(10,10)(i2) )

local m = nn.gModule({i1,i2}, {o1,o2})

m:forward({torch.randn(10), torch.randn(10)})
m:backward({torch.randn(10), torch.randn(10)}, {torch.randn(10), torch.randn(10)})

m:float()

print{m}

In the following example, the gradInput layout should be:

{
  1 : DoubleTensor - size: 3
  2 : 
    {
      1 : DoubleTensor - size: 3
      2 : DoubleTensor - size: 3
    }
}

The obtained wrong gradInput layout is:

{
  1 : DoubleTensor - size: 3
  2 : 
    {
      1 : 
        {
          1 : DoubleTensor - size: 3
          2 : DoubleTensor - size: 3
        }
      2 : DoubleTensor - size: 3
    }
}

It happens when:

1. An input is obtained by split().
2. And the output of the split is used multiple times in the graph.

Code to reproduce the problem:

local xInput = torch.randn(3)
local h = torch.randn(3)

local x = nn.Identity()()
local prevRnnState = nngraph.Node({input={}})
local prevH1, prevCell = prevRnnState:split(2)
local prevH = prevH1
-- Wrapping the input by Identity() avoids the problem.
--local prevH = nn.Identity()(prevH1)

local cellOut = nn.CAddTable()({
        nn.CMulTable()({x, prevH}),
        nn.CMulTable()({prevH, prevCell})})
local module = nn.gModule({x, prevRnnState}, {cellOut})

local c = torch.randn(h:size())
local prevRnnState = {h, c}
local output = module:forward({xInput, prevRnnState})

local gradOutput = torch.randn(h:size())
local gradInput = module:backward({xInput, prevRnnState}, gradOutput)
local gradX, gradPrevState = unpack(gradInput)
local gradPrevH, gradPrevCell = unpack(gradPrevState)
assert(type(gradPrevH) == type(h), "wrong gradPrevH type")

Is there a way to pass a subset of outputs to the next module?
For example, nn.SplitTable() can have 2 outputs.
And I want to send output[1] to nn.Linear and output[2] to nn.Tanh.

I see that the gradInput returned from a gModule
has wrong elements.
The expected gradInput is a table.
The shapes of the elements in the table seem to be switched.
The shapes should be (2) and (5).
But the obtained shapes are (5) and (2).

Am I using the nngraph incorrectly?

Code to reproduce the problem:

require 'nn'
require 'nngraph'

local getInput1 = nn.Identity()()
local getInput2 = nn.Identity()()
local mlp = nn.Tanh()(getInput1)
local net = nn.gModule({getInput1, getInput2}, {mlp, getInput2})


local input1 = torch.randn(2)
local input2 = torch.randn(5)

net:forward({input1, input2})
local gradInput = net:backward({input1, input2},
    {torch.randn(input1:size()), torch.randn(input2:size())})
print("gradInput[1]:", gradInput[1])
print("gradInput[2]:", gradInput[2])
assert(gradInput[1]:nElement() == input1:nElement(), "size mismatch")

Hi, in many situations we don't need to count dervative wrt the input, so for optimiztion it is good idea to write
....updateGradInput = function(input) return end
In container nn.Sequential it works.
In gModule it doesn't.
But if you run once your model with standart updateGradInput function and than write updateGradInput = function(input) return end, everything will be fine.

From my current experimentation it appears that nesting graphs withing graphs (using the same __ call__ syntax) is not yet possible. This is a nice-to-have, but not urgent.

When I try to use "graph.dot" on mac, it gives following error:

Warning: Could not load "/usr/local/Cellar/graphviz/2.38.0/lib/graphviz/libgvplugin_dot_layout.6.dylib" - file not found
Error: Layout type: "dot" not recognized. Use one of: circo dot fdp neato nop nop1 nop2 osage patchwork sfdp twopi
/Users/luo123n/torch/install/bin/luajit: ...s/luo123n/torch/install/share/lua/5.1/graph/graphviz.lua:135: graphviz layout failed
stack traceback:
    [C]: in function 'assert'
    ...s/luo123n/torch/install/share/lua/5.1/graph/graphviz.lua:135: in function 'graphvizFile'
    ...s/luo123n/torch/install/share/lua/5.1/graph/graphviz.lua:162: in function 'dot'
    test.lua:23: in main chunk
    [C]: in function 'dofile'
    ...123n/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:131: in main chunk
    [C]: at 0x010e387330

I can find similar issues using google, but there seems no solution yet.

On ubuntu 14.04, to get nngraph graphviz t owork, I need to changes lines 29 and 30 of graphviz.lua to:

local graphvizOk, graphviz = pcall(function() return ffi.load('libgvc.so.6') end)
local cgraphOk, cgraph = pcall(function() return ffi.load('libcgraph.so.6') end)

(ie adding postfix '.6' to each of hte filenames)

(I'm not sure how to do this portably though, hence putting this as an issue, rather than a PR.

Hey guys,

suppose I create a gmodule with two input A and B.
Now A will be the training example, and B will be some random sampled negative example
Now during forward path I want the full calculation only be done once, and from the second time only B will be changing and A will be fixed, so the calculation related with A is not necessary to be conducted again.

So is there some example show that how to skip some nodes in the nngraph during forward pass?

I'm following the README.md file and when I get around to running:

graph.dot(mlp.fg, 'MLP')

from the first example (two hidden layers MLP), I get the following error:

...l/Cellar/torch/HEAD/share/lua/5.1/graph/graphviz.lua:162: attempt to index local 'g' (a nil value) stack traceback: ...local/Cellar/torch/HEAD/share/lua/5.1/trepl/init.lua:501: in function <...local/Cellar/torch/HEAD/share/lua/5.1/trepl/init.lua:494> ...l/Cellar/torch/HEAD/share/lua/5.1/graph/graphviz.lua:162: in function 'graphvizFile' ...l/Cellar/torch/HEAD/share/lua/5.1/graph/graphviz.lua:195: in function 'dot' [string "_RESULT={graph.dot(mlp.fg, 'MLP')}"]:1: in main chunk [C]: in function 'xpcall' ...local/Cellar/torch/HEAD/share/lua/5.1/trepl/init.lua:651: in function 'repl' ...lar/torch/HEAD/lib/luarocks/rocks/trepl/scm-1/bin/th:199: in main chunk [C]: ?

I'm running this from trepl on a mac. I have already run brew install graphviz (it installed version 2.38.0) and luarocks install graph (as suggested at the end of this thread) successfully. I haven't been able to find anything on this error online either. How do I get around this error?

For example the 'mapindex' field is now more consistent, and the explicit table-splitting nodes are new.

Please see the issue posted in torch/torch7#711. I had accidentally submitted an issue for torch/nngraph there, and I'm not sure how to remove it. Thank you!

Is is intended to use nngraph.Node(input={}) instead of nn.Identity()?

The following example works, only if using nngraph.Node instead of nn.Identity.

local in1 = nn.Identity()()
local in2 = nn.Identity()()
-- The nngraph.Node() is needed for the example to work.
--local in2 = nngraph.Node({input={}})
local prevH, prevCell = in2:split(2)

local out1 = nn.CMulTable()({in1, prevH, prevCell})
local module = nn.gModule({in1, in2}, {out1})

local input = {torch.randn(3), {torch.randn(3), torch.randn(3)}}
module:forward(input)
local gradInput = module:backward(input, torch.randn(3))
assert(type(gradInput[2]) == "table", "wrong gradInput[2] type")