Simplify model: shared DINOv3 WEB + MONA in last 12/24 blocks

Three related architecture changes, driven by a cost/simplicity trade-off:

1. **Shared encoder**: one DINOv3 LVD-1689M (WEB) processes both drone
   and satellite images. Previously asymmetric — separate WEB (drone) and
   SAT-493M (satellite) encoders. Saves ~303M frozen params and halves
   VRAM for the image tower. Expected to lose some satellite-domain
   inductive bias; MONA adapters pick up the slack.

2. **MONA in last 12/24 blocks**: adapters injected only in the top half
   of the ViT. The lowest 12 blocks keep their pretrained features
   untouched. Trainable MONA count drops from 14.0M (48 adapters × 2
   encoders) to 3.5M (24 adapters × 1 encoder).

3. **No DINO_SAT**: `nn_models/DINO_SAT` is no longer loaded by the
   default config. It stays on disk and the path param is kept for
   backward compat with asymmetric checkpoints.

Parameter counts (with text fusion + LoRA + gates):
  Before: 17.6M trainable / 733M total (2.35%)
  After:   7.06M trainable / 434M total (1.63%)

Also fixes a pre-existing resume bug: checkpoints now record
`shared_encoder`, `baseline_mode`, `mona_bottleneck`, `mona_last_n_blocks`
so `AsymmetricEncoder.load_checkpoint` can rebuild the right architecture.
Old checkpoints still load (missing keys fall back to asymmetric defaults
via `ckpt.get(..., <default>)`).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

CLAUDE.md

@@ -1,10 +1,13 @@
 # Caption Quality Test for Cross-View Geo-Localization
 
-## Архитектура системы (v3, 2026-04-21) — GTA-UAV эксперимент
+## Архитектура системы (v3, 2026-04-24) — GTA-UAV эксперимент
 
 ```
+Shared DINOv3 ViT-L/16 (LVD-1689M, frozen + MONA in last 12/24 blocks)
+для обеих веток — drone и satellite кодируются одним encoder.
+
 QUERY BRANCH (drone + L1/L2/L3 captions):
-  drone_img [B,3,256,256] --> DINOv3 ViT-L/16 LVD-1689M (frozen) --> d_img [B,1024]
+  drone_img [B,3,256,256] --> DINOv3 ViT-L/16 (shared) --> d_img [B,1024]
                                                                         |
   L1 --> DGTRS-CLIP (248 tok) --> z₁ [768] --\                         |
   L2 --> DGTRS-CLIP (248 tok) --> z₂ [768] ---+-- cat --> MLP(2304→1024→1024) --> d_txt [B,1024]
@@ -15,7 +18,7 @@ QUERY BRANCH (drone + L1/L2/L3 captions):
                                   q̂ = q/‖q‖₂ --> query [B,1024]
 
 GALLERY BRANCH (satellite + satellite captions):
-  sat_img [B,3,256,256] --> DINOv3 ViT-L/16 SAT-493M (frozen) --> s_img [B,1024]
+  sat_img [B,3,256,256] --> DINOv3 ViT-L/16 (shared, same weights) --> s_img [B,1024]
                                                                       |
   sat_L1 --> DGTRS-CLIP --> z₁ --\                                    |
   sat_L2 --> DGTRS-CLIP --> z₂ ---+-- cat --> MLP (shared) --> s_txt [B,1024]
@@ -54,15 +57,16 @@ BASELINE: σ(α_q)=σ(α_g)=1.0, text disabled, DGTRS not loaded
 - Transformer: sequence-first (LND), nn.MultiheadAttention, 12 layers
 - Tokenizer: BPE SimpleTokenizer (248 tokens, vocab 49408)
 
-### Trainable parameters: 17.6M из 748M (2.35%)
+### Trainable parameters: 7.06M из 434M (1.63%)
-- **MONA adapters** (2×DINOv3): 14.0M (2 per block × 24 × 2 encoders, bottleneck=64)
+- **MONA adapters** (shared DINOv3): 3.5M (2 per block × 12 last blocks, bottleneck=64)
 - **LoRA** (DGTRS-CLIP): 147K (Q+V, rank=4, 12 blocks)
 - TextFusionMLP (shared): Linear(2304,1024)+GELU+Linear(1024,1024) = ~3.4M
 - gate α_q + α_g: 2 scalars
 - logit_scale: 1 scalar (learnable temperature)
-- DINOv3 x2 + DGTRS: frozen backbone weights
+- DINOv3 (1 encoder) + DGTRS: frozen backbone weights
 - **Без projection layers** — retrieval space = DINOv3 native 1024-dim
 - **AMP:** frozen layers fp16, adapters + loss fp32
+- **Примечание:** ранее была asymmetric setup (2×DINOv3 WEB+SAT, MONA во всех 24 блоках) с 17.6M trainable / 733M total. Упростили до shared + last-12 MONA.
 
 ### Optimizer & Scheduler
 - **AdamW** с per-group LR: projections lr=1e-4, text encoder lr=1e-5
@@ -123,20 +127,14 @@ Eval: Resize(256) + CenterCrop(256) + ImageNet normalization.
 
 ## Backbones (v3)
 
-### DINOv3 ViT-L/16 — Drone (web pretrained)
+### DINOv3 ViT-L/16 — Shared (web pretrained)
 - **Checkpoint:** `nn_models/DINO_WEB/dinov3-vitl16-pretrain-lvd1689m.pth`
 - **Arch:** ViT-L/16, 24 layers, 16 heads, hidden=1024, MLP=4096, 303M params
 - **Input:** 256x256, ImageNet normalization, patch=16 → 256 patches
 - **Register tokens:** 4, RoPE theta=100.0
-- **Status:** frozen
+- **MONA:** 24 адаптера в последних 12 блоках (blocks 12-23), bottleneck=64, 3.5M trainable
-
+- **Status:** frozen кроме MONA
-### DINOv3 ViT-L/16 — Satellite (sat pretrained)
+- **Примечание:** ранее asymmetric — использовался отдельно `nn_models/DINO_SAT/model.safetensors` (sat493m pretrain) для satellite ветки. Упростили до shared WEB-энкодера.
-- **Checkpoint:** `nn_models/DINO_SAT/model.safetensors`
-- **HuggingFace:** `facebook/dinov3-vitl16-pretrain-sat493m`
-- **Arch:** идентична DINO_WEB (ViT-L/16, hidden=1024, 303M params)
-- **Input:** 256x256
-- **Config:** `nn_models/DINO_SAT/config.json` — BROKEN (auth error), используем конфиг от DINO_WEB
-- **Status:** frozen
 
 ### DGTRS-CLIP ViT-L-14 (LRSCLIP) — Text encoder
 - **Checkpoint:** `nn_models/LRSCLIP/DGTRS-CLIP-ViT-L-14.pt`
@@ -306,10 +304,10 @@ python -m scripts.compare_runs \
 ### V3 (GTA-UAV, DINOv3 ViT-L/16, 256x256)
 | Фаза | Оценка |
 |------|--------|
-| VRAM: 2x DINOv3-L + LRSCLIP + batch 64 | ~18-22 GB |
+| VRAM: DINOv3-L (shared) + LRSCLIP + batch 64 | ~10-14 GB (было ~18-22 с 2× DINOv3) |
 | GPU mem (smoke test, batch 4) | 3.1 GB |
 | Batch size | 64 (default) |
-| Total params | 733M (10.9M trainable, 1.49%) |
+| Total params | 434M (7.06M trainable, 1.63%) — shared encoder + MONA в last 12/24 blocks |
 
 ### V2 (UAV-GeoLoc, GeoRSCLIP)
 | Фаза | Время |

conf/gtauav_balanced.gin

@@ -22,7 +22,9 @@ TrainConfigGTAUAV.device = "cuda"
 # ---- Model ----
 TrainConfigGTAUAV.init_gate = 0.7
 TrainConfigGTAUAV.baseline_mode = False
-TrainConfigGTAUAV.shared_encoder = False
+TrainConfigGTAUAV.shared_encoder = True  # single DINOv3 WEB for both branches
+TrainConfigGTAUAV.mona_bottleneck = 64
+TrainConfigGTAUAV.mona_last_n_blocks = 12  # inject MONA only in last 12/24 ViT blocks
 TrainConfigGTAUAV.gradient_checkpointing = True
 
 # ---- Loss ----

src/models/asymmetric_encoder.py

@@ -540,6 +540,8 @@ class AsymmetricEncoder(nn.Module):
             lrsclip_path=lrsclip_path,
             baseline_mode=ckpt.get("baseline_mode", False),
             shared_encoder=ckpt.get("shared_encoder", False),
+            mona_bottleneck=ckpt.get("mona_bottleneck", 64),
+            mona_last_n_blocks=ckpt.get("mona_last_n_blocks", 24),
             device=device,
         )
         model.load_state_dict(ckpt["model_state"], strict=False)

src/training/train_gtauav.py

@@ -86,7 +86,9 @@ class TrainConfigGTAUAV:
     lrsclip_path: str = _LRSCLIP
     init_gate: float = 0.7
     baseline_mode: bool = False
-    shared_encoder: bool = False  # asymmetric: WEB (drone) + SAT (satellite)
+    shared_encoder: bool = True  # single DINOv3 WEB for both branches (simpler, half the params)
+    mona_bottleneck: int = 64
+    mona_last_n_blocks: int = 12  # inject adapters only in last 12 of 24 ViT blocks
     gradient_checkpointing: bool = True  # trade compute for VRAM (allows larger batch)
 
     # Training.
@@ -537,6 +539,8 @@ def train(cfg: TrainConfigGTAUAV) -> None:
             init_gate=cfg.init_gate,
             baseline_mode=cfg.baseline_mode,
             shared_encoder=cfg.shared_encoder,
+            mona_bottleneck=cfg.mona_bottleneck,
+            mona_last_n_blocks=cfg.mona_last_n_blocks,
             device=cfg.device,
         ).to(cfg.device)
         LOGGER.info("embed_dim=%d", model.embed_dim)
@@ -1006,13 +1010,18 @@ def train(cfg: TrainConfigGTAUAV) -> None:
 
         history.append(epoch_record)
 
-        # Save checkpoint.
+        # Save checkpoint. Model architecture flags go into the ckpt so
+        # `AsymmetricEncoder.load_checkpoint` can rebuild the right shape.
         _atomic_save(
             obj={
                 "epoch": epoch,
                 "model_state": model.state_dict(),
                 "optimizer_state": optimizer.state_dict(),
                 "loss_state": loss_fn.state_dict(),
+                "baseline_mode": cfg.baseline_mode,
+                "shared_encoder": cfg.shared_encoder,
+                "mona_bottleneck": cfg.mona_bottleneck,
+                "mona_last_n_blocks": cfg.mona_last_n_blocks,
             },
             path=output_dir / f"ckpt_epoch{epoch:03d}.pt",
         )