search

Chaphlagical / Deblur-GS

[I3D 2024] Deblur-GS: 3D Gaussian Splatting from Camera Motion Blurred Images
https://chaphlagical.icu/Deblur-GS/
Other
354 stars 36 forks source link

Training takes so long #4

Open pdxmusic opened 6 months ago

pdxmusic commented 6 months ago

Thanks for your work! After a series of problems compiling Pytorch3d i was able to make everything work.

I just started a training using the bench dataset but i see it takes about 9/10 hours to accomplish. Is this the expected timing or is there something wrong on my side?

I'm on ubuntu 22.04 with a Nvidia 4090 Driver Version: 535.171.04 CUDA Driver Version: 12.2 Conda environment NVCC is 11.6

Chaphlagical commented 6 months ago

You can decrease the blur_sample_num to make it faster if necessary. Most of the scene might not need 31.

pdxmusic commented 6 months ago

What would be a valid value? 15?

Chaphlagical commented 6 months ago

15 samples with 5-6 order (if you use bezier curve) should be enough.

pdxmusic commented 6 months ago

Thanks for the suggestion, now it's way better!

NOTE: I encountered an error while running python train.py -s /home/rdover2/deblur-datasets/bench/ --eval --deblur --blur_sample_num 15 --mode Bezier

AttributeError: module 'scipy' has no attribute 'special'

I had to add from scipy import special in utils/pose_utils.py

giich commented 6 months ago

Thanks for your work! After a series of problems compiling Pytorch3d i was able to make everything work.

I just started a training using the bench dataset but i see it takes about 9/10 hours to accomplish. Is this the expected timing or is there something wrong on my side?

I'm on ubuntu 22.04 with a Nvidia 4090 Driver Version: 535.171.04 CUDA Driver Version: 12.2 Conda environment NVCC is 11.6

Pytorch problem is a mean and lean in nvcc args? Edit: Oh you are on Ubuntu, sorry ))

GopiRajuMatta commented 3 weeks ago

Why in general training time is longer than vanilla Gaussian Splatting? Its almost like training NeRF.. Can you please clarify..