Dominic Jack

I spent _a lot_ of time trying to work this out. ~As far as I can tell shuffling the points that are used as inputs to the model won't make...

@zubair1502 It's been a while since I've looked at this, but is it not just `prob_sample` (is this the IFP algorithm? I can't remember...) followed by `gather_point`? If you're using...

I don't have any fundamental disagreements there. I suppose it might be worth documenting known differences either in the layer docstrings or wherever benchmarks results are documented?

Just put in formatting w/ CI + pre-commit [PR](https://github.com/danielegrattarola/spektral/pull/152). W.r.t. typing, tensorflow has [expressed interest in typing](https://github.com/tensorflow/tensorflow/issues/12345), and other packages (e.g. [tensorflow-probability](https://github.com/tensorflow/probability), [tensorflow-datasets](https://github.com/tensorflow/datasets)) has already implemented them. I suspect it...

Pretty sure this can be resolved using `tfds.download.DownloadConfig` with `verify_ssl=False`.

@julienvalentin @amakadia this supercedes [PR 101](https://github.com/tensorflow/graphics/pull/101).

@avneesh-sud sorry about the delay on this - ECCV deadline is taking priority right now.

@avneesh-sud - sorry about the delay, rolled straight into a new role after ECCV. I've had another look and... things aren't as straight forward as I remember. I'm running experiments...

For the models from the notebook, the sparse implementation (proposed in this PR) takes ~170ms per step, vs 215ms for the original implementation on my laptop with 1050Ti. Not much...

[Truncating points](https://github.com/tensorflow/graphics/blob/master/tensorflow_graphics/projects/pointnet/train.py#L73) is along the batch dimension, not the points dimension. For some reason this changes the shape from `(?, 2048, 3)` to `(32, 2034, 3)` (which in the past...