Add generic_segmentation probe for dinov2_vitb14

README.md

@@ -0,0 +1,64 @@
+---
+library_name: visbench
+tags:
+- visbench
+- probing
+- generic_segmentation
+---
+
+# generic_segmentation probe for `dinov2_vitb14`
+
+A **trained probe head**, not a backbone. It is the small module VisBench fits
+on top of frozen `dinov2_vitb14` features to measure what those features carry.
+
+```python
+import visbench
+from visbench.hub import load_probe_from_hub
+
+backbone = visbench.get_backbone("dinov2_vitb14")
+probe = load_probe_from_hub("turhancan97/visbench-generic_segmentation-dinov2_vitb14", backbone=backbone)
+```
+
+## It only works with this backbone
+
+These weights were fitted on features from `dinov2_vitb14`, taken with
+`pooling=mean` and `feature_mode=dense_only`. Loading them against
+anything else is refused, because the failure is otherwise silent: a head fitted
+on one pooling and fed another has the right shapes and produces a plausible,
+wrong number.
+
+| | |
+| --- | --- |
+| backbone | `dinov2_vitb14` |
+| backbone key | `dinov2/dinov2_vitb14/224/7764ea0f912e` |
+| task | `generic_segmentation` (mid_level) |
+| pooling | `mean` (requested `mean`) |
+| feature mode | `dense_only` |
+| layers | `None` |
+
+**Reported scores**
+
+| metric | value |
+| --- | --- |
+| `f1` | 0.8422 |
+| `iou` | 0.7579 |
+| `pixel_acc` | 0.9366 |
+
+## Reproducing it
+
+Fitted with:
+
+- `batch_size`: `8`
+- `epochs`: `10`
+- `head`: `linear`
+- `hidden_dim`: `512`
+- `layers`: `None`
+- `loss`: `masked_bce`
+- `lr`: `0.0005`
+- `optimizer`: `adamw`
+- `protocol`: `visbench_binary_seg`
+- `threshold`: `0.5`
+- `warmup_epochs`: `1.5`
+- `weight_decay`: `0.0001`
+
+Generated by [VisBench](https://github.com/turhancan97/VisBench).

probe.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d0bde3b1baad6533131507494375a9eac6e3337712fead0e0e4680788be6424
+size 5653