Origin story
FPS-UDA began as a small question about whether a frozen encoder's feature space already contained enough structure, and whether the harder problem was placing a better decision plane on top of it.
Before this prototype, our experiments mostly followed the usual route of adapting the image encoder itself. We tried objectives related to contrastive and consistency-style learning, aiming to make source and target feature distributions more compatible.
During early exploration, we asked a simpler question: if a
pretrained encoder is frozen, can the existing feature space still
support a better target decision boundary? The earliest test saved
ResNet 2048-dimensional features into a small H5 file and trained
only a lightweight linear DY decision module over
those fixed features.
That prototype was not a polished benchmark. It was a direct feature-space experiment, kept here because it records the first concrete version of the idea.
The included H5 uses the early flat schema:
src_feature, src_label,
tgt_feature, and tgt_label. It is not
the dataset-level feature-bank schema used by current configs.
The notebooks are preserved without cleaning outputs, paths, device choices, or exploratory cells. Some cells may correspond to nearby historical variants rather than the exact included H5.
Target labels in the prototype were used for diagnostic or oracle-style reporting while exploring the idea. They are not a supervised signal in the current benchmark workflow.