pikaliov 916854f124 MONA only on last 12 blocks (skip early low-level blocks) ...

- Add last_n_blocks parameter to inject_mona_into_dinov3 (default=12)
- Blocks 0-11: pure frozen DINOv3 (low-level features, domain-agnostic)
- Blocks 12-23: MONA adapted (high-level semantic features)
- MONA params: 3.5M (was 6.85M, -49%)
- Total trainable: ~5.7M with text (was ~9.0M)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-21 21:51:52 +03:00

..

datasets

Remove projections (1024 native), add satellite text, dual GatedFusion

2026-04-21 19:01:30 +03:00

eval

Rewrite: GatedFusion architecture + UAV-GeoLoc dataset

2026-04-17 17:13:00 +03:00

losses

Remove projections (1024 native), add satellite text, dual GatedFusion

2026-04-21 19:01:30 +03:00

models

MONA only on last 12 blocks (skip early low-level blocks)

2026-04-21 21:51:52 +03:00

training

Add gradient checkpointing for DINOv3 and DGTRS-CLIP (bs 8→24)

2026-04-21 21:34:31 +03:00

__init__.py

Initial commit: caption quality test on UAV-VisLoc

2026-04-17 00:04:46 +03:00