Three fixes for working Landsat detector training + inference pipeline
Fix missing matches between Landsat scenes and small windows. This is because scenes actually go beyond the WRS2 polygon so we need to buffer the polygons.
There was a broken test.
Fix training with FreezeUnfreeze + plateau scheduler running into error about plateau scheduler having incorrect number of min_lrs (compared to number of parameter groups).
I updated to add unfreeze_lr_factor setting (which controls how much lower the learning rate of the now unfrozen component should be vs the other parameters) along with a test for it.
Three fixes for working Landsat detector training + inference pipeline