2d-weight-space / checkpoints /checkpoints.md
h-d-h's picture
Add SIREN bbox baseline checkpoints
4d600cb verified
# 2D Weight-space Checkpoints
Bounding-box regression checkpoints for SIREN/ImageNet-Loc experiments.
| Checkpoint | Best epoch | Best val mean IoU | Best val IoU >= 0.5 |
|---|---:|---:|---:|
| `dinov2_cls_cached-linear.pt` | 9 | 0.4316 | 0.4518 |
| `dinov2_cls_cached-two_layer_64.pt` | 13 | 0.4576 | 0.5005 |
| `dinov2_cls_cached-two_layer_512.pt` | 3 | 0.4563 | 0.4959 |
| `dinov2_cls-two_layer_64.pt` | 4 | 0.4559 | 0.4973 |
| `identity-linear.pt` | 13 | 0.4013 | 0.3909 |
| `identity-two_layer_64.pt` | 1 | 0.4022 | 0.3921 |
| `identity-two_layer_512.pt` | 2 | 0.4037 | 0.3935 |
The `dinov2_cls_cached-*` models use frozen `dinov2_vits14` CLS features cached from 224x224 SIREN renders. The `identity-*` models use flattened SIREN parameter stacks directly.