mit-han-lab/torchsparse

✨ New Features

• Supported mixed-precision training and inference with `torch.cuda.amp` (https://github.com/mit-han-lab/torchsparse/pull/69, https://github.com/mit-han-lab/torchsparse/pull/75).
• Added generalized sparse convolution (https://github.com/mit-han-lab/torchsparse/pull/77).
• Added group normalization (https://github.com/mit-han-lab/torchsparse/pull/63).

📋 API Changes

• `sparse_{collate,quantize}` now needs to be imported from `torchsparse.utils.{collate,quantize}`.
• `sparse_quantize` now takes in `coords`, `voxel_size` (defaults to `1.0`), `return_index` (defaults to `False`) and `return_inverse` (defaults to `False`) as input, and returns the quantized `coords` as well as `indices` and `inverse_indices` (if requested).
• `transpose` is renamed as `transposed` for `torchsparse.nn.Conv3d` and `torchsparse.nn.functional.conv3d`.
• `KernelRegion` is removed and replaced with `get_kernel_offsets` (in `torchsparse.nn.utils`).