Needs a test!

@ezyang, @fmassa, some remarks?

@fmassa, PR is ready for review.
It's usual for target to be 2 or 3, and when summing loss for a large sample the error of approximation grows, so I wanted to add optional ability to compute exact log(target!) for small target values. This could be implemented as:

mask = target <= 10
loss[mask] += target[mask].apply_(lambda x: math.log(math.factorial(x)))
loss[~mask] += ...  # Stirling approximation

But autograd.Variable has no apply member. This would not have impact on automatic differentiation because we do not need to compute derivative w.r.t target. Are there some workarounds?

@apaszke, could you comment on what is written above?

@vlasenkov using an apply_ function would be extremely slow on the GPU, so I wouldn't really advise on that. I'll have a look at your latest changes.
On a quick look, I think it looks ok. But I wonder if you don't have problems with autograd in loss[mask] += ... which complains that you can't use an in-place operation there (I had sometimes to do things like x[1:10] = x[1:10] + y instead of x[1:10] += y for it to work).

loss[mask] += ... works without problem. I've tested on OS X (cpu) and Ubuntu (cpu & gpu)

ping @fmassa

Can anyone merge this loss? Does it still need any fixes?

I think this looks good, thanks!
@vlasenkov could you add link to the docs in https://github.com/pytorch/pytorch/blob/master/docs/source/nn.rst? If not they won't show in the website.

thanks @vlasenkov !