You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I propose this first version of working colab of Plenoxels to Readme.md for people to have form of fast realtime playing with training and rendering?
#50
Open
neurall opened this issue
Jan 31, 2022
· 0 comments
Hi guys.
Many people are struggling with trying to train this on common gpus just to findout that gpus with 16g or more are needed on many coolest provided sample datasets due to their size. PlenOctree page mentions 24g ?
Sadly gpus are currently pretty much impossible to buy plus those big GPU mem sizes are mostly domain of cloud provided ones.
Too bad current svox2 version doesnt fit 12g ram and 15g gpu of Free colab instance to train datasets of M60 size
So I created at least barely fitting to gpu mem (training and rendering of coolest m60 tank dataset ) on Colab Pro
it's 10$ bucks for month which is price of one hamburger but still better than nothing.
Too bad that algo doesn't probably yet support? training on smaller mem gpus even in parallel.
There is gazillion mining rigs that people have at homes typically with multiple 4-11g gpus. I have one 10g 3080 twice as fast as colabpro, 11g 1080 which is the same speed , and 6g of 1060 which is 27g gpu mem lying idly in this case
I noticed that free kaggle gpu is 16g and should fit but system ram is problem which is solvable faster by multiple smaller 10g capped cpu passes ? or by using manually memory mapping file to go around mem size problem I guess? https://thechief.io/c/editorial/comparison-cloud-gpu-providers/
Perhaps we could in theory split big dataset to smaller ones by segmenting images with known similar camera locations to smaller chunks from bundle adjustment colmap files that datasets already have?
It would be cool if Plenoxels github readme had link to some colab to allow others to play with technology faster.
Perhaps "download checkpoints and render-images only part" could be perhaps made working even on free colab ?
That would still provide people with an interactive access to get feel for this incredible new rendering possibility without all the finding expensive gpu and installing fuss?
I was just thinking UE5 nanite team ? ;D
Just think about the possibilities If photorealistic scene reconstruction is this good and fast thx to Plenoxels.
UE5 is able to render massive datasets with unlimited details on whim and you don't even need file you can stream vertices to it directly. I think it will not take a long until we see app that streams to real-time VR experience as a new form of live reporting concert or journalism?
I.e. we could be transported to any place on earth in VR and move around as things are happening ;D. Imagine VR news network. exiting times
The text was updated successfully, but these errors were encountered:
neurall
changed the title
I propose working colab of Plenoxels to Readme.md for people to have form of fast realtime playing with training and rendering?
I propose this first version of working colab of Plenoxels to Readme.md for people to have form of fast realtime playing with training and rendering?
Jan 31, 2022
Hi guys.
Many people are struggling with trying to train this on common gpus just to findout that gpus with 16g or more are needed on many coolest provided sample datasets due to their size. PlenOctree page mentions 24g ?
Sadly gpus are currently pretty much impossible to buy plus those big GPU mem sizes are mostly domain of cloud provided ones.
Too bad current svox2 version doesnt fit 12g ram and 15g gpu of Free colab instance to train datasets of M60 size
So I created at least barely fitting to gpu mem (training and rendering of coolest m60 tank dataset ) on Colab Pro
it's 10$ bucks for month which is price of one hamburger but still better than nothing.
https://colab.research.google.com/drive/1SODy_HiP_kkjL5E4IiBM1XZsD1HofElZ?usp=sharing
Too bad that algo doesn't probably yet support? training on smaller mem gpus even in parallel.
There is gazillion mining rigs that people have at homes typically with multiple 4-11g gpus. I have one 10g 3080 twice as fast as colabpro, 11g 1080 which is the same speed , and 6g of 1060 which is 27g gpu mem lying idly in this case
I noticed that free kaggle gpu is 16g and should fit but system ram is problem which is solvable faster by multiple smaller 10g capped cpu passes ? or by using manually memory mapping file to go around mem size problem I guess?
https://thechief.io/c/editorial/comparison-cloud-gpu-providers/
Perhaps we could in theory split big dataset to smaller ones by segmenting images with known similar camera locations to smaller chunks from bundle adjustment colmap files that datasets already have?
Anyway. I made colab editable for others to make improvements and on
https://github.com/neurall/PlenoxelsColab to make forks easily possible too
It would be cool if Plenoxels github readme had link to some colab to allow others to play with technology faster.
Perhaps "download checkpoints and render-images only part" could be perhaps made working even on free colab ?
That would still provide people with an interactive access to get feel for this incredible new rendering possibility without all the finding expensive gpu and installing fuss?
I was just thinking UE5 nanite team ? ;D
Just think about the possibilities If photorealistic scene reconstruction is this good and fast thx to Plenoxels.
UE5 is able to render massive datasets with unlimited details on whim and you don't even need file you can stream vertices to it directly. I think it will not take a long until we see app that streams to real-time VR experience as a new form of live reporting concert or journalism?
I.e. we could be transported to any place on earth in VR and move around as things are happening ;D. Imagine VR news network. exiting times
The text was updated successfully, but these errors were encountered: