weird kernel dispatch for gradient of 3-interpolate about votenet HOT 3 CLOSED

@AluminiumUnibody

Good catch!

I think it seems a bug/typo. It happens that the shape matches so there is no explicit error.

Unfortunately I don't have time right now to examine into this issue.
It would be very helpful if you have time to take a deeper look! e.g. use the three_interpolate_grad_kernel_wrapper and try to verify the gradients or directly re-run the training experiment.

otherwise, I may find some time during the weekend to look into it.

from votenet.

@charlesq34 Okay thanks. It's quite interesting that it didn't seem to have affected the training much.

from votenet.

I looked into the issue and found that it is a bug. It has now been fixed.

I've added pointnet2/pointnet2_test.py to gradient check the fixed three_interpolate ops.
Re-training with the fixed ops resulted in a slightly improved model. With the eval code in train.py, we can get around 58.6 [email protected] for SUN RGB-D (although evaluation with eval.py seems not show such big improvement).

Thank you @AluminiumUnibody for reporting this issue!

from votenet.