Isaac Robinson

feel free to submit a pr! :) this looks like a great addition to the repo, but we don't have bandwidth to make it happen for the foreseeable future

@probicheaux tagging in case you want to think about head reinitialization a bit more. This argument makes sense to me.

@probicheaux thoughts?

Is your goal to get segmentation working? When you say training from scratch, do you also mean no dinov2 weights?

got it. we'll have a segmentation head out soon, actively working on it. yolo series heads aren't directly applicable to detr series models so we've built our own

Instance seg is live! :)

hi! pulling out the precomputed interpolated positional embeddings will make the onnx graph slower for any given image size, which is why we precomputed them in the first place. have...

You don't need the exact hardware to do a latency test, although of course it won't be comparable to the official result without it. You can use other hardware to...

Nothing comes to mind! Can you do a pip freeze and provide versions?

interesting, no obvious issues. would you be comfortable sharing your checkpoint? or a minimal reproduction checkpoint that shows the issue without violating your model / data privacy