New User Question: Bad argument #2 to 'SpatialConvolutionMM_updateOutput'

[Eddie]

Hello,

I am a new user to Torch, and I am trying to implement a NN that will work with images. When attempting to train my network I get the following error:

/nn/SpatialConvolution.lua:88: bad argument #2 to 'SpatialConvolutionMM_updateOutput' (3D or 4D(batch mode) tensor expected)
stack traceback:
    [C]: at 0x7f30eee04930
    [C]: in function 'SpatialConvolutionMM_updateOutput'
    /lua/5.1/nn/SpatialConvolution.lua:88: in function 'updateOutput'
    /lua/5.1/nn/Sequential.lua:25: in function 'forward'
    MyCNN.lua:41: in main chunk

My understanding of the error is that nn.Sequential is expecting to be trained on 3D Tensors, but has received some other torch datatype.
When I look at my inputs (loaded images) I find a 3D tensor. This issue has stumped me for a few weeks now.

Any help is much appreciated. Thank you.

Here the code I use to build and train my network for reference:

require 'nn'
require 'image'
require 'lfs'

-- ACTIVATION FUNCTION
ReLU = nn.ReLU

-- NETWORK TOPOLOGY
-- SpatialConvolution(nInputPlane, nOutputPlane, kW, kH, [dW], [dH], [padding])
-- SpatialMaxPooling(kW, kH [, dW, dH])
local model = nn.Sequential()
model:add(nn.SpatialConvolution(3, 3, 240, 240))
model:add(nn.SpatialConvolution(3, 3, 11, 11, 4, 4)):add(ReLU(true))
model:add(nn.SpatialConvolution(3, 48, 5, 5)):add(ReLU(true))
model:add(nn.SpatialMaxPooling(5, 5, 3, 3))
model:add(nn.SpatialConvolution(48, 256, 3, 3)):add(ReLU(true))
model:add(nn.SpatialConvolution(256, 192, 3, 3)):add(ReLU(true))
model:add(nn.SpatialConvolution(192, 192, 3, 3)):add(ReLU(true))
model:add(nn.Linear(4096,4096)):add(nn.ReLU(true))
model:add(nn.Linear(4096,2)):add(nn.ReLU(true))

-- TRAINING THE NETWORK --
for file in lfs.dir(lfs.currentdir().."/FinalData") do
    if (file ~= ".") and (file ~= "..") then
        local input = image.load("FinalData/"..file, 3); --torch.DoubleTensor
       
        local output= torch.Tensor(2);
        local c = file:sub(1,1)
        local outpustStorage = output:storage()
        if c == "A" then
            outpustStorage[1]=0;
            outpustStorage[2]=1;
        elseif c == "G" then
            outpustStorage[1]=1;
            outpustStorage[2]=0;
        else
            error("Invalid Input Image Filename")
        end
       
        criterion = nn.MSECriterion();
        criterion:forward(model:forward(input), output);
        model:zeroGradParameters();
        model:backward(input, criterion:backward(model.output, output));
        model:updateParameters(0.01);
    end
end

[Jonghoon Jin]

What is the dimensions of input? Your model seems to require a large image as input.

It would be good to check dimensions of output in each convolution layer.

--
You received this message because you are subscribed to the Google Groups "torch7" group.
To unsubscribe from this group and stop receiving emails from it, send an email to torch7+un...@googlegroups.com.
To post to this group, send email to tor...@googlegroups.com.
Visit this group at http://groups.google.com/group/torch7.
For more options, visit https://groups.google.com/d/optout.

[Eddie]

The input of the network is a color image of 240px by 240px.
My hope was that I could achieve this by having the first layer of the network be 3 planes (one for red values, one for green, one for blue). With 1 kernal of size 240x240.

I am starting to think I am not understanding torch's convolution nn layers correctly. Am I too far off from the right idea?
Thank you.

On Saturday, April 25, 2015 at 11:43:49 AM UTC-4, Jonghoon Jin wrote:

What is the dimensions of input? Your model seems to require a large image as input.

It would be good to check dimensions of output in each convolution layer.

[Eddie]

Also, I just tried to look at the dimensions of each convolution layer using this simple loop:

conv_nodes = model:findModules('nn.SpatialConvolution')
for i = 1, #conv_nodes do
  print(conv_nodes[i].output:size())
end

This caused a new error, which has made me more sure that I am not understanding nn.SpatialConvolution
This is the error:

Tensor.lua:23: calling 'min' on bad self (tensor must have one dimension)

On Saturday, April 25, 2015 at 11:43:49 AM UTC-4, Jonghoon Jin wrote:

What is the dimensions of input? Your model seems to require a large image as input.

It would be good to check dimensions of output in each convolution layer.

[Jonghoon Jin]

The first module model:add(nn.SpatialConvolution(3, 3, 240, 240)) for the given input image of 3x240x240 produces the output tensor of 3x1x1, which is too small for the following module.

You may want to look at this https://github.com/eladhoffer/ImageNet-Training/blob/master/Models/AlexNet.lua if you plan to do sth simliar to alex's model.

On Wed, Apr 29, 2015 at 9:44 AM, Eddie via torch7 <torch7+APn2wQdEe1WPGRWqGMUpZUHAh...@googlegroups.com> wrote:

The input of the network is a color image of 240px by 240px.
My hope was that I could achieve this by having the first layer of the network be 3 planes (one for red values, one for green, one for blue). With 1 kernal of size 240x240.

I am starting to think I am not understanding torch's convolution nn layers correctly. Am I too far off from the right idea?
Thank you.



On Saturday, April 25, 2015 at 11:43:49 AM UTC-4, Jonghoon Jin wrote:

What is the dimensions of input? Your model seems to require a large image as input.

It would be good to check dimensions of output in each convolution layer.