Chris Choy

I don't understand the question, but shapes are normalized. https://github.com/StanfordVL/MinkowskiEngine/blob/master/examples/modelnet40.py#L302

I see. This is the original ModelNet40 dataset and I fixed some headers in the .off files to load them correctly. If there's another version of this dataset that cleaned...

You can use the function `ME.utils.batched_coordinates` to recreate batched coordinates. For features, you can simply concatenate along the 0th dim.

Please post the compilation log. [All these misconfigurations](https://github.com/NVIDIA/MinkowskiEngine/issues?q=is%3Aissue+Error+compiling+objects+for+extension) during installation result in the same error.

`nvcc fatal : Unsupported gpu architecture 'compute_86'` It seems like you are using a different nvcc version for your installation. Please make sure that `nvcc --version` matches the CUDA version...

Try using CUDA 11.1. https://forums.developer.nvidia.com/t/nvcc-fatal-unsupported-gpu-architecture-compute-86/161424 Or you could only use the compute architecture up to 8.0 but not 8.6 by ``` cd MinkowskiEngine TORCH_CUDA_ARCH_LIST="5.2 6.0 6.1 7.0 7.5 8.0" python...

The TensorRT mostly supports standard neural network layers that are already in cuDNN. Custom libraries like MinkowskiEngine won't get a significant speed-up from converting the entire library to a TensorRT...

This will get the gradient but it would only work with batch size 1. For other batch sizes. You will be computing attention over all points in all batch.

It would be really good to have MathJax in github readme!

Hmm I think the definition is a bit different in a sparse tensor, but definitely a useful feature to have! Let me get back to this after some deadlines.