Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

3d高斯的tune.py运行速度稍微有些慢 #16

Closed
chenqi13814529300 opened this issue Jul 10, 2024 · 12 comments
Closed

3d高斯的tune.py运行速度稍微有些慢 #16

chenqi13814529300 opened this issue Jul 10, 2024 · 12 comments

Comments

@chenqi13814529300
Copy link

train.py比tune.py运行速度快好几倍?为什么?

@Abyssaledge
Copy link
Collaborator

In tune.py, we need to render depth maps and calculate other losses. We cannot help you if you do not specify the adopted data and commands.

@YuxueYang1204
Copy link
Owner

I suggest trying Trim2DGS instead. Trim2DGS inherently handles Normal Regularization within the rasterization pipeline, resulting in faster fine-tuning speeds. Moreover, Trim2DGS often offers more stable and better reconstruction performance.

@Abyssaledge
Copy link
Collaborator

Feel free to reopen this issue if you would like to offer more details.

@chenqi13814529300
Copy link
Author

chenqi13814529300 commented Jul 10, 2024

我没看到论文关于速度的评价。以你们默认参数的情况下,原版高斯和你们修改的高斯,速度差距多少(30K的情况下)。原版2D和你们修改的2D高斯,速度差距是多少(30K的情况下)。

@Abyssaledge
Copy link
Collaborator

这个文章并不是为了加速,所以并没有专门讨论速度。你说的速度应该是指训练速度? 渲染速度是不变的。 训练速度3D情况下因为多一次深度图渲染,所以会慢2倍多。为了方便开发,我们没对这个做优化,你想加速的话,可以把rgb和depth的渲染写到一个cuda kernel里,这样就快很多了。 2D情况下相比2DGS基本不变,因为没有额外的渲染操作。

@chenqi13814529300
Copy link
Author

好的,我知道了,谢谢您的回答。

@chenqi13814529300 chenqi13814529300 changed the title 3d高斯的tune.py运行速度怎么这么慢? 3d高斯的tune.py运行速度稍微有些慢 Jul 11, 2024
@chenqi13814529300
Copy link
Author

你们的2d高斯tune.py我测试过了,貌似失去了2d高斯多视角不变性的特点。例如下面的图片,而原始的2d高斯却不会产生这种qin情况。
image
image

@chenqi13814529300
Copy link
Author

image
还有这些虚影是多出来了,不清楚是为什么

@Abyssaledge
Copy link
Collaborator

  1. 应该是trimming得比较多导致的空洞,你可以试试调大trimming的间隔,或者调小每次trimming的比例。
  2. 虚影可能是不合理的scale split导致的,可参见 Handling Spurious Depth Points #3

另外你可以把你的具体训练参数发一下,我们看看有没有啥问题。 @chenqi13814529300

@chenqi13814529300
Copy link
Author

chenqi13814529300 commented Jul 11, 2024

训练参数就默认参数,30K,就self.lambda_dist = 0.0 ,原始2D高斯我也是这么干的。而且你们的2D高斯跑出来的速度比较慢,原始的论文的2D是27分钟,在里面的工程中你们需要50多分钟(2d的train.py和tune.py都如此)

@YuxueYang1204
Copy link
Owner

YuxueYang1204 commented Jul 11, 2024

速度比2DGS慢是因为他们进行了优化,可见 hbb1/diff-surfel-rasterization#7
我们后续也会在Trim2DGS跟进

训练参数就默认参数,30K,就self.lambda_dist = 0.0 ,原始2D高斯我也是这么干的。而且你们的2D高斯跑出来的速度比较慢,原始的论文的2D是27分钟,在里面的工程中你们需要50多分钟(2d的train.py和tune.py都如此)

@chenqi13814529300
Copy link
Author

好的,谢谢您的回复,我的问题已经解决

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants