| --- |
| 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` | 0.8408 | |
| | `iou` | 0.7556 | |
| | `pixel_acc` | 0.9360 | |
|
|
| ## 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). |
|
|
