search

google-research / jax3d

Apache License 2.0
729 stars 94 forks source link

Running MobileNeRF on non-GPU server #186

Open smtabatabaie opened 1 year ago

smtabatabaie commented 1 year ago

Hi, I wanted to ask if mobileNerf can be run on a non-GPU normal server. I mean training it on a local GPU machine and then serving it on a non-GPU server for viewers to view the mobileNerf on their browser. Thanks

DWhettam commented 1 year ago

If WebGL is supported, then yes. Go here to check: https://webglreport.com/?v=2 You can also just test out the provided demos on the machine you want to use. WebGL runs on the client side, so the models will be ran on your local machine.

smtabatabaie commented 1 year ago

Thanks very much @DWhettam , Is it possible to train mobilenerf on an Nvidia 3090 ?

DWhettam commented 1 year ago

I'm not the authors, and I've not tried, but I don't see why not. Give it a go! You may need to comment out the line checking for 8 GPUs.

smtabatabaie commented 1 year ago

Thanks very much @DWhettam , I could train the chair scene with my 3090 successfully (took about a day!). Can you also help me on how can I train my own custom dataset with mobilenerf? I mean I've got NeRFs trained with Nerfstudio, can I use those scenes and convert and train them with mobilenerf? Thanks again

DWhettam commented 1 year ago

@smtabatabaie No worries. Again, I'm not the authors so I'm not an expert on this at all. I don't know about Nerfstudio, but I'm having a go at using my own dataset at the moment so I can share some insight. You will need to extract camera poses and resize your data. I have been trying to follow the process used in multinerf (https://github.com/google-research/multinerf/tree/main) as that is the source of the real360 dataset. I haven't finished testing this yet, but my understanding is you will need to create camera poses using LLFF (https://github.com/Fyusion/LLFF#general-step-by-step-usage), and then resize your images using the multinerf script (https://github.com/google-research/multinerf/blob/main/scripts/local_colmap_and_resize.sh). I think once you've done those you should have all the necessary files for training, although I will report back once I've confirmed myself

EDIT: I have got this working following the steps as described. With camera poses and resized images it should work out the box.