Memory usage optimization

[bloom256]

Hello Oliver! Thank you for your work.
Are there any plans to improve memory efficiency in your implementation? All Gaussians3D fields are float32 now, but float16 is probably sufficient for many of them. If at least the SH coefficients became float16, it would significantly reduce memory usage because there are 48 of them. Positions, radii, opacity and rotations are also good candidates in my opinion. Of course, this can be optional, as done with depth.

[bloom256]

I apologize if my previous message was unclear. After some tests, I've found that "rotation," "log_scaling," and "alpha_logit" can be converted to float16 without issues. However, converting "feature" or "position" to float16 causes crashes.

I believe making "feature" field float16 could greatly improve memory efficiency. Do you have any plans for this, or could you provide guidance on how to achieve it?

Thank you again for your amazing work!

[oliver-batchelor]

Hi Danil, Thanks for trying things out - I admit I haven't put any effort into optimizing for memory use as for all the things I use it for I'm not close to running out for any of the applications I use it for. It's definitely worthy to investigate, though it's quite far down on my list of things to do! I'm happy to accept contributions and I'll leave this here as a TODO - but it's not high on my priority right now! As to why it might crash I'm not too sure. I think the gradients will end up as fp32 even if the inputs are fp16, that may have something to do with it. Cheers, Oliver

…

On Thu, Jun 13, 2024 at 11:05 PM Danil ***@***.***> wrote: I apologize if my previous message was unclear. After some tests, I've found that "rotation," "log_scaling," and "alpha_logit" can be converted to float16 without issues. However, converting "feature" or "position" to float16 causes crashes. I believe making "feature" field float16 could greatly improve memory efficiency. Do you have any plans for this, or could you provide guidance on how to achieve it? Thank you again for your amazing work! — Reply to this email directly, view it on GitHub <#13 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAITRZJHULLBF47S67HBFUTZHF4GZAVCNFSM6AAAAABIZDIF2OVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCNRVGMZTCNRVGM> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>