[WIP] FIX Unit tests on Windows

[rth]

This PR aims to fix #82

It includes,

• an Appveyor CI setup
• a fix of the randint overflow issue
• an attempt to fix dtype mismatch between IntTensor and LongTensor by casting them in forward as suggested here embeddings layer with IntTensor / cuda.IntTensor inputs pytorch/pytorch#145 (comment) . I must be missing something though as I still don't understand why this not an issue on Linux but only on Windows..

The latest Appveyor output can be seen here some of the failures will go away once #83 is merged..

[maciejkula]

I suspect both have something to do with numpy integers being 64-bit on Linux but not (apparently) on Windows.

[maciejkula]

It seems we are very efficiently duplicating each other's work :)

[rth]

@maciejkula yes, I just realised this is a duplicate of your PR #84.

I can remove the redundant parts, up to you how to split this work, I wouldn't mind helping a bit :)

[maciejkula]

OK, I like your set up better, so let's roll with what's here and I'll close the other PR.

With respect to the int/long issue: I think I would prefer to track down the root issue (why do we have ints anywhere at all?) and leave the layers without casting. Is there a way you could roll back the embedding layer changes only so that I could see where it fails in Appveyor?

[maciejkula]

What surprises me is that I have explicit type conversions here, but it looks like this gets ints anyway.

That's probably a fruitful line of research if you have pdb handy.

[rth]

Great!

With respect to the int/long issue: I think I would prefer to track down the root issue (why do we have ints anywhere at all?) and leave the layers without casting. Is there a way you could roll back the embedding layer changes only so that I could see where it fails in Appveyor?

Yes, sure, done.

What surprises me is that I have explicit type conversions here, but it looks like this gets ints anyway.

I don't get it either, but I'm also not used to debugging PyTorch code. The function deeper in the traceback that raises the exception (e.g. BloomEmbedding._get_hashed_indices) also doesn't get ints all the time, it gets mostly longs and sometimes ints and then it fails. I can try to debug this some more tomorrow if needed.

[maciejkula]

Great! I'm of limited use because I don't have access to a Windows machine, so can't easily make it fail without going to Appveyor all the time.

I think the key will be dropping into a debugger where it fails and going up the stack until we hit the source of the problem.

[maciejkula]

In hindsight, so obvious. Thanks a lot for the fixes! Are you happy for me to merge?

If so, I'll push a new version out to Conda tonight.

[rth]

Yes, everything appears to be due to randint. Opened an issue in numpy/numpy#10317 .
Sure, feel free to merge. Thanks for the quick review!

[maciejkula]

🎉

(I have seen similar things with arange, and I'm not sure if there is an underlying rationale to defaults changing between version/platforms of numpy.)