VolumetricAveragePooling gives error when Kernel Size is equal to the Input Feature Map(s)'s size

[rudrapoudel]

Hi there,

When I am using Kernel Size is equal to the Input Feature Map(s)'s size, I am getting following error,

...ch/install/share/lua/5.1/nn/VolumetricAveragePooling.lua:20: bad argument #2 to 'VolumetricAveragePooling_updateOutput' (input image smaller than kernel size)

I think we need to check > rather than >= in the following kind of argcheck ???

luaL_argcheck(L, input->size[dimw] >= kW && input->size[dimh] >= kH &&
input->size[dimt] >= kT, 2,
"input image smaller than kernel size");

Here is the code to check the things quickly,

require 'torch'
require 'nn'
torch.setdefaulttensortype('torch.FloatTensor')

local nn_backend = nn
if false then
    require 'cudnn'
    nn_backend = cudnn
    torch.setdefaulttensortype('torch.CudaTensor')
end

if false then
    vap = nn_backend.VolumetricAveragePooling(2, 2, 2,
                                              2, 2, 2)
else
    vap = nn_backend.VolumetricAveragePooling(5, 7, 5,
                                              5, 7, 5)
end

inputs = torch.rand(8, 16, 5, 7, 5)  -- batch size, no. of feature maps, t, h, w
outputs = vap:forward(inputs)
print(outputs:size())

[kmul00]

I think the problem lies in the different order the dimenisons are parsed by the lua and the underlying c code, respectively.

In VolumetricAveragePooling.lua the input is parsed as kT, kW, kH (here).

Whereas the VolumetricAveragePooling.c parses the kernel indices to be dimt = 1, dimh = 2, dimw = 3 (here). The order of height and width are getting exchanged, and hence the mismatch.

In fact, on further inspection, I found that this is the case with all the _Volumetric_ modules.

@soumith and @fmassa , can either of you please confirm regarding this.
If true, then it needs to be changed. I can send fix for it.

[kmul00]

The easiest fix could be to change the input format of the _Volumetric_ modules (the lua files) from (kT, kW, kH) to (kT, kH, kW).

Then the underlying c codes as well as the cuda codes (here) for all the modules need not be changed.

[soumith]

@kmul00 i dont think that's the issue. They are correctly parsed here: https://github.com/torch/nn/blob/master/generic/VolumetricAveragePooling.c#L46-L51

[kmul00]

@soumith Yeah, that part is correct. The input format is (kT, kW, kH), and they are parsed correctly as well, like you pointed out. But then it assigns,

int dimt = 1;
int dimh = 2;
int dimw = 3;

So, when it performs input->size[dimw] >= kW, it is basically comparing the third diemension of the input (since, dimw = 3) with the second dimension of the kernel (kW). I think that's where the discrepency lies.

And similarly for input->size[dimh] >= kH.

[soumith]

dimw = 3 is correct. The ordering in the constructors is a bit legacy, and is ordered kT, kW, kH. But the inputs / outputs are ordered iT x iH x iW.

[kmul00]

I think that's why it is creating a problem. Most of the times we keep the height and width dimenion same, hence it doesn't cause any problem.

For example, if I use

input = torch.Tensor(8, 16, 5, 7, 5)
model = nn.Sequential()
model:add(nn.VolumetricAveragePooling(5, 7, 5)

it throws an error as pointed out in the issue. But instead, if I use

model:add(nn.VolumetricAveragePooling(5, 5, 7)

it executes without any error. Of course the results are erroneous, but the dimesnions doesn't cause any problem.

[soumith]

The example itself is wrong. Inputs should be:

inputs = torch.rand(8, 16, 5, 5, 7) -- t, h, w

because the constructor is:
(T, W, H, ...)

[soumith]

As given here: https://github.com/torch/nn/blob/master/doc/convolution.md#volumetricfullconvolution

[kmul00]

Oh, now I see. This made it clear. The orderings of the kernel and input are altogether different. Sorry for the confusion.