Togo 2020 #157
Togo 2020 #157
Conversation
|
Performance is quite a bit worse than LP on the same labels in the Helios dataset.... |
| int(dst_geometry.shp.y), | ||
| ) | ||
| else: | ||
| bounds = ( |
There was a problem hiding this comment.
This looks like it is copied from Yawen's previous code that had bug; dst_geometry.shp.x/y is the center so the window is offset, and after cropping to center 8x8 depending on the configured window size it may no longer contain the labeled point at all.
There was a problem hiding this comment.
Yawen put fixed function here that you could use:
https://github.com/allenai/rslearn_projects/blob/master/rslp/utils/windows.py
I don't think the docstring is too informative, maybe can fix it to make it clear that it is calculating bounds for a certain window size centered at the centroid of the given geometry (and then use shp.centroid.x/y instead of shp.x/y which won't work if shp is not point).
https://wandb.ai/eai-ai2/2025_06_11_helios_finetuning/runs/d40ftfir