Tests against torch.Tensor API

[szagoruyko]

I think we should add some basic tests against torch tensor API, that would prevent #91 for example. Maybe some tests against torch/image too.

[shrubb]

Yeah...now it's a complete mess!

I've set bindings to take storageOffset into account while fixing #91, but I'm REALLY afraid of backwards conversion... Indeed, we need tests...

[szagoruyko]

that was quick, thanks! I will add some tests when I have time.

[szagoruyko]

oops wrong issue

[edgarriba]

hey guys! I have some contributions to OpenCV and now starting with Torch. I would like to join to the party! Is there any road map for this bindings?

[fmassa]

@edgarriba cool! It will be great to have contributions!
So, I'm not 100% sure that it's up-to-date (@shrubb can confirm), but there is a TODO list in the wiki https://github.com/VisionLabs/torch-opencv/wiki

[shrubb]

@edgarriba
That's great, you're very welcome! Indeed, the to-do list is a bit outdated, however, it's true that our primary concern is unit testing. There's currently even no testing framework in the project at all, so we need one.

[edgarriba]

nice! so let's focus on that. What are candidates? AFAIK, some packages in Torch use 'totem' to do that

[shrubb]

For those following, there's been a nice start in #131, thanks Sergey

[chaneyw]

The wrapper for Torch Tensors seems to sometimes (or all the time?) ignore changes in view. For example, using image.load from Torch Image package yields images that are CxHxW. If you call permute, Torch says the Tensor is now HxWxC (the format that OpenCV functions expect) but it does not work correctly (for example, cv.undistort()) has this behavior.

[fmassa]

@chaneyw torch-opencv expects the tensors to be contiguous in memory, and permute makes it non-contiguous. You should call :contiguous() after permute.
There was some discussion about that in #32 .