i tried to install detectron2 on a windows machine and it is not supported
Facebook research say so in the install instructions:
but when i saw that their Circle-Ci job didn’t successfully built for a year i decided to search for a solution elsewhere
https://app.circleci.com/pipelines/github/facebookresearch/detectron2
am i correct ?
I see that github’s 5 month old automated test shows a failure…
But there are also reports that it worked, so I don’t know which is true.
https://detectron2.readthedocs.io/en/latest/tutorials/install.html
not 5 month, the last successful job was over a year ago (windows and linux),
i tried so many options to launch a docker that builds from source using ubuntu:24.10, CUDA 12.4, torch 2.4.1, python 3.12.4, and when that didn’t work i started lowering more and more the versions, the only (partial) success i got was with the original Dockerfile from the git (ubuntu18.04, CUDA 11.1, torch 1.10, python 3.7) - it is very old and not updatedm and even then 40% of the unittest fails
where did you see reports of success ? can you share ? maybe someone had better luck
where did you see reports of success ? can you share ? maybe someone had better luck
Since the page is in Japanese, I used a Google Translate link.
thanks for the share, but they install detectron from git (meaning they are using the old versions), also they don’t state what CUDA, python, torch or Linux version are they using
(since Ubuntu 24.04 you cannot install pip packages globally and i think detectron needs the torch to be installed globally and not in a venv, so a workaround should be in place until Facebook will update their code and build)
so this is also not an updated versions install,
i haven’t given up yet, i still have some ideas and trying them out, but i haven’t seen anyone share a successful install in the last year with newer versions, its such a shame since it looks like a promising and advanced package,
Okay. That was an example of installing an older version…
There don’t seem to be a lot of resources related to HF, but there are a few people who seem to know a lot about it. Maybe you could send them a mentions from here and ask them. (@+username)
thanks for the advice, after allot of trial and error i finally got it to working with the newest compatible versions i could:
i shared the Dockerfile and Commands here:
Good.
Still, the fact that it works without modification, even in a limited operating environment, means that the github dependency settings and manual are somehow incomplete.
Maybe the branch is broken if CUDA is missing or old?
Anyway, I’m glad it works.
This topic was automatically closed 12 hours after the last reply. New replies are no longer allowed.
