thomasbrandon

Having also noted that `transform_points` was a major performance point I rewrote it using [Numba](https://numba.pydata.org/) with good results. I get a 60-90% improvement in synthetic tests (depending on number of...

OK, getting it into shape to PR now. I've also converted the creation of the boxes for `BoxRasterizer` to Numba with a good improvement there too (~60% faster iterating the...

To update, I haven't had too much time to spend on this but have now pushed an [initial version](https://github.com/thomasbrandon/l5kit/tree/numba_acceleration). Just want to do some final checks and then will do...

No problem, thanks for the heads-up. Would that just be #150 or are there other changes in the works? Will check out that PR for conflicts/clashes.

Thanks for this rigred, the provided configs helped a lot for an ASRock B450 Pro4. I too have the negative values for temp2. Do you know what the bug/missing feature...

That colon shouldn't be there. Looks like it's part of the CUDA path detected by `torch.utils.cpp_extension` (which does the compilation). Check if `CUDA_HOME` environment variable is set and verify value....

@rafale77 Did you re-install the extension after upgrading PyTorch. I wouldn't expect binary compatibility across versions so you need to re-install to re-compile.

Without a specific error message it's hard to know. Make sure the CUDA SDK is installed and accessible through the `CUDA_HOME` environment variable. Otherwise you'd need to get a better...

CUDA 11.0 doesn't support RTX30 series cards. They were introduced in CUDA 11.1.

You should be able to just pip install the repo (i.e. `pip install git+https://github.com/thomasbrandon/mish-cuda/`). The main thing that's needed is correct environment variables for the CUDA toolkit which should be...