using optimfn with wrapped nn modules

[sebastiangonsal]

Is there anything wrong with using wrapped ``nnmodules together withoptim`.

I have code like following:. The problem is when predict is called. I keep getting the following error:

...install/share/lua/5.1/autograd/runtime/codegen/Graph.lua:40: bad argument #2 to 'fn' (expecting number or torch.DoubleTensor or torch.DoubleStorage at /tmp/luarocks_torch-scm-1-9261/torch7/generic/Tensor.c:1125)

[C]: in function 'fn'
...install/share/lua/5.1/autograd/runtime/codegen/Graph.lua:40: in function 'set'
 /home/user/torch/install/share/lua/5.1/torch/Tensor.lua:458: in function 'fn'
...install/share/lua/5.1/autograd/runtime/codegen/Graph.lua:40: in function 'view'
...user/torch/install/share/lua/5.1/autograd/nnwrapper.lua:206: in function 'fn'
 .../install/share/lua/5.1/autograd/runtime/codegen/Node.lua:72: in function 'evaluateForward'
...install/share/lua/5.1/autograd/runtime/codegen/Graph.lua:25: in function 'conv1'
(my stacktrace below this. first line of predict function here)

local conv1, params.conv1
conv1, params.conv1 = grad.nn.SpatialConvolutionMM(64, 128, 5, 5, 1, 1, 2, 2)

...

params = autograd.util.cast(params, "float")

function predict(params, input)
   local h1 = pool1(acts1(conv1(params.conv1, input)))
   ....
  return output

function feval(params, input, target)
   local prediction = predict(params, input)
   ...
   return loss, prediction

local df = autograd(feval, {optimize = true})
state = {
learningRate = learningRate,
momentum = momentum,
weightDecay = weightDecay
}

for i = 1, numepochs do
   local optimfn, states = grad.optim.sgd(df, state, params)
   for j = 1, datasize, batchSIze do
       local X, target = loadInputTensor(batchSize)
       local grads, loss = optimfn(X, target)
   end
end
...