MaxUnpool2d segfaults for some configurations

[pytorch]

MaxUnpool2d segfaults for some configurations

[apaszke]

Probably fixed in #207. I can't reproduce it anymore, so I'm closing the issue.

[soumith]

ok, so to MaxUnpool* we also need to give as input the target size....

[fmassa]

I think so, that's how it's done in nn.

[apaszke]

I had this segfault like 2 times for me, but it works fine most of the time. I could never reproduce it with gdb attached 😕 I can't really find any reason for segfaults since we're checking the bounds at every access. Maybe calling THError/jongjmp/exception inside omp sections is UB?

[fmassa]

I think the core of this issue was solved with the keyword argument output_size from #113.

[apaszke]

It's not, it still segfaults if you don't specify it

[fmassa]

But in any case, the user is supposed to pass the output_size argument if he wants the module to compute the result properly for most of the cases, as the locations of the max can be off, even if there is no segfault or runtime error...

[apaszke]

Yes, they are supposed to pass it, but even if they don't we should never segfault.