Charles Gaydon

> _Original comment by_ **Ionel Cristian Mărieș (Bitbucket: [ionelmc](https://bitbucket.org/ionelmc), GitHub: [ionelmc](https://github.com/ionelmc))** > > I've taken a look at this as well, my opinion is that a combination of configuration, subprocess...

Additionally, PointNet is not supposed to be affected by redundant points due to its max-pooling operation. It is therefore possible to just sample point clouds with replacement, to a unique...

Gentle bump on this issue. It would be really helpful to know to which extent the performance are different than the ones in the original paper ! :)

Is there any news on this visualization bug? I have the same issue since the new view and expected it to be corrected by now :octopus:

To me this is a similar feature request to https://github.com/comet-ml/issue-tracking/issues/218. Has this really been implemented? Right now we would need to hack into the logging in order to order experiments...

A note here: from what I understand the Sampler will be considering each combination of labels as a kind of *meta-label*. This could be a highly combinatorial setting, with **a...

I implemented RandLa-Net for segmentation as well, and made some small refactor. The model seems to learn quite well, and reaches 70% accuracy after 3 epochs. It takes ~1s to...

Thank you, this worked like a charm. I think this is ready for review. :) Right now both scripts follow the paper's architecture (in terms of hyperparameters, depth and number...

I identified some differences between my implementation and the original paper (in particular in terms of batch norms, activations, and number of channels). Will come back with fixes and modifications!

Here comes the update! I made a few changes: - Set the BatchNorm and LeakyReLU with parameters from the paper (in particular: 0.99 momentum for BN, 0.2 alpha for lrelu,...