Vadim Kantorov

here is a limited example of such bindings: https://github.com/vadimkantorov/dotnetdlpack

I guess for proper ref-counting like semantics (so that NumPy doesn't call the deleter too early in presence of other array views) something like weakref would be needed: https://stackoverflow.com/questions/37988849/safer-way-to-expose-a-c-allocated-memory-buffer-using-numpy-ctypes, but...

Yeah. It shouldn't release the buffer and shouldn't call deleter either if there're some other existing arrays (it should also ideally work when torch.from_numpy is called on such a NumPy...

What are valid integer codes for color and materials? non-zero? Colors and materials are encoded as uint8 in CLEVRwithmasks and with strings in CLEVR.

An example of dump that I get: ```js { 'color': [0, 1, 2, 3, 1, 1, 4, 5, 0, 0, 0], 'material': [0, 1, 1, 2, 2, 1, 2, 2,...

Thanks! It would help adding these to README! I've got another question: how were train/test splits done? (for both CLEVR6 and CLEVR10) Could you provide the file lists?

Do I understand correctly that `sum(visibility) 7` - test? Or did test also contain some (or all?) of `sum(visibility)

Does test split ensure it doesn't intersect too much with train? E.g. are train images excluded? Are there any filtering wrt object properties? Do you have somewhere still sizes of...

Is it true that: - first 70k examples are used for train (and further filtered to contain

Seems it's 100K, it wouldn't hurt to put this number in README :)