optim.checkgrad dimensions dependent

[bermanmaxim]

I'm cross-posting from a question on google groups because I believe there is a problem with checkgrad, although it's possible I have overlooked something. Considering this function

function f(x)
    local vect = x.new(x:size()):fill(1)
    local fx = torch.dot(x:view(-1), vect:view(-1))
    return fx, vect
end

fx is the dot product between flattened x and flattened vect, therefore the jacobian of fx should simply be vect (ones of the size of x).

Yet when checking with optim.checkgrad

th> a = torch.rand(2, 5)
th> diff, dC, dC_est = optim.checkgrad(f, a)
th> dC_est
 5.0000  5.0000  5.0000  5.0000  5.0000
 5.0000  5.0000  5.0000  5.0000  5.0000
[torch.DoubleTensor of size 2x5]

dC_est evaluates to 5.0 * (ones of the size of x). PyTorch's autograd returns the result I would expect, a tensor of ones:

>>> import torch
>>> from torch.autograd import Variable
>>> a = torch.rand(2, 5)
>>> av = Variable(a, requires_grad=True)
>>> b = Variable(torch.ones(10))
>>> c = torch.dot(av.view(-1), b)
>>> c.backward()
>>> print(av.grad)
Variable containing:
 1  1  1  1  1
 1  1  1  1  1
[torch.FloatTensor of size 2x5]