search

nianticlabs / simplerecon

[ECCV 2022] SimpleRecon: 3D Reconstruction Without 3D Convolutions
Other
1.3k stars 121 forks source link

No CUDA support on macOS with M1 #32

Open forEachWhileTrue opened 1 year ago

forEachWhileTrue commented 1 year ago

Hey there!

I'm going to write my bachelor thesis about solving the occlusion problem in AR with 3D reconstruction. There is really exciting stuff like NeuralRecon and this project.

Sadly I can't run SimpleRecon because my Mac doesn't support CUDA. Is there a way to switch the device to CPU or even MPS? https://pytorch.org/docs/master/notes/mps.html

Otherwise I have to get SimpleRecon running on a remote device which supports CUDA and my first thought was Google Colab. I'm really new to PyTorch and neural networks in general so maybe you can help me out here and/or give me some recommendations. This would be great!

Thanks in advance and making this stuff open source. :)

jgibson2 commented 1 year ago

Hi,

There shouldn't be anything preventing you from running on CPU (as long as you don't try to use any of the point cloud fusion utils or some of the visualization scripts, which might contain non-portable .cuda() calls, but you should probably be able to remove those without much trouble).

Re: MPS, I haven't tried it, so I can't guarantee the operators will all work. You will definitely have to play around with the map_location argument to torch.load to get the weights to the correct device. If you have success, let us know :)

alsuren commented 1 year ago

I've made a start here: https://github.com/rerun-io/simplerecon/pull/1 . Once I have something working, I will clean it up and turn it into a PR against this repo.