Johan Edstedt
Johan Edstedt
Hi, then it should be possible to train the descriptor part. The detector relies on SfM, so can be a bit tricky to retrain. You will need to make a...
Ubuntu 18 is still stuck at glibc 2.28 (2.27?), can we get this on pip? Don't like conda installs.
@WANG-KX I'm also interested in the training data :)
> the dropbox train link seems to be empty. test works. could you please reupload? @WANG-KX A fellow dataset enjoyer I see :)
Want to bump specifically selecting EyeMode a'la https://github.com/rerun-io/rerun/blob/9fe16ba57c2f83f0644f9d7624384b047d7fe282/crates/viewer/re_space_view_spatial/src/ui.rs#L163-L167 blueprintly. I basically always want the first person viewer.
Haven't run into this myself, possibly due to torch.float16 overflows or underflows. It should (I think), be fine to run dinov2 in torch.bfloat16, could you try that?
I think dinov2 dtype is hardcoded
Also, what image are you sending in to the model? What are the pixel values?
We match with softmax over scaled cosine similarly. We call the scaling inverse temperature. People often refer to temperature when scaling logits, probably it originally comes from statistical mechanics if...
This must be due to alpha channel, Ill check it out.