Watching Meta’s Segment Anything Model (SAM) exploding online and wondering anyone is working on a port of the Python version to Julia
or if a similar model could be trained using the Julia ecosystem as the training data has been made public (SA-1B Dataset)
SAM seems like a fundamental tool for many AI/image processing tasks that would be a same not to have access to.
Not your question, but i got it to run pretty quickly with pycall and im currently writing a few helpers and an interactive tool to segment interactively with makie 
Will soon publish something …
That does seem useful & cool. It would be great to have a Julia port since that would free us of python package management dependency hell. Since the license is Apache, a direct port seems like it might be the most straightforward way to go. I’m in the middle of other things now, but if anyone wants to give this a stab (manually, or using their favorite AI coding assistant to build the first draft) I’d be happy to review a package prototype.
Do you mean the whole training cycle or just recreating the inference process with their weights?
I think the latter. Is there an up to date way of reading pth files in Julia that contain the trained model info? I assume some manual work would be required to build out the right neural network components to mirror the pytorch setup and then initialize it with the right data from the pth checkpoint.
I’ve seen Pickle.jl recommended. It includes special support for pth files.
Theres a bit more work still until its registered, but it works fine already.