improved windows installation steps

[martinResearch]

trafficstars

Problem The current documentation regarding the installation on windows is a bit scarce regarding the CUDA installation and which version of the CUDA dependencies should be used as well as numpy version, making the installation process on windows somewhat error prone. Also using an installation of the CUDA toolkit at the system level can create undesired coupling with other python environments used for other projects.

Solution Instead of suggesting to install the CUDA toolkit manually, we can suggest using conda to install all the CUDA dependencies in an isolated manner. We can provide an example of a complete conda environment.yml file with pinned dependency versions that is sufficient to get the tests running, which will hopefully provide a more robust procedure to get gsplat running on windows.

[martinResearch]

proposed improved windows installation https://github.com/nerfstudio-project/gsplat/pull/350

[softyoda]

Is this for the prebuilt wheel? I'm completely, i still have either struct.error: unpack requires a buffer of 4 bytes or gsplat-1.4.0+pt24cu124-cp310-cp310-win_amd64.whl is not a supported wheel on this platform. . I'm lost and nobody is on the discord, look like we still have to rely on wls2++ to run on windows.

[PetterGs]

look like we still have to rely on wls2++ to run on windows.

Same

[martinResearch]

@softyoda , which version of pytorch and cuda are you using? how have you installed cuda ? what error message do you get when using pip install gsplat --index-url https://docs.gsplat.studio/whl/pt20cu118

[martinResearch]

@softyoda the precompiled wheels are published only for python 3.10 at the moment. Which version of python are you using?

[n2k3]

I'm also trying to get gsplat running on Windows 11 Pro (23H2), without to having install Visual Studio and dependent build tools. My goal is to run the simple_trainer.py as stated in the docs.

I have two issues when running on Windows:

1. fused-ssim does not provide pre-built wheels (for any platform afaik): which speeds up the training and inference step by a 5-8x 🚀 - current work around: fallback to pytorch-msssim
2. the pycolmap does provide wheels, but not with CUDA support, therefor not using GPU acceleration for the SfM step (if I understand correctly) related question: wouldn't be better to move to the now available python bindings from the official repo, they also publish the package on pypi (still without CUDA support in their wheels)

---

~~_{Unrelated notes:}~~ ~~_{yesterday i ran into this and today I ran into this issue not being to download the dataset. (also nothing was mentioned about download the dataset yourself, or how to do this)}~~ ~~_{Now stuck on running the simple_trainer.py command, ModuleNotFoundError: No module named 'datasets.colmap' from line 18 in simple_trainer.py am I missing steps? should I not clone the respo and run scripts from the .\examples folder? should I run setup.py in the root? hehe it's been a rough first couple of days of my python journey}~~ _{EDIT: I'm a noob, don't clone this repo and try to run scripts from the .\examples\ folder python will get confused resolving packages}