Model loading for Aegis-River-Segmentation (SegFormer-B2, water body mask).
The underlying model was trained as a 7-class semantic segmentation model
(background, water, sky, vegetation, building, vehicle, person), but only
the 'water' clas
s (id=1) is reliable in practice. This Space collapses the
output to binary: water vs everything else.
IMPORTANT: transformers must be pinned to ==4.45.2. A later transformers
version renamed an internal decode_head submodule
(decode_head.linear_projections -> decode_head.linear_c), which silently
breaks the correspondence between this checkpoint's keys and the model
architecture on a mismatched version.
