Initial commit: caption quality test on UAV-VisLoc

Self-contained experimental track validating generated text captions
via retrieval R@1 lift on UAV-VisLoc.

Architecture: GeoRSCLIP ViT-B/32 dual encoder, 512-dim shared space.
Loss: 4-term InfoNCE (img-img + sat-cap + drone-cap + cap-cap)
      with cosine temperature decay, PALW-like curriculum.
Metric: delta R@1 (with text - without text) >= +3% => PASS.

Gin-configured (balanced / baseline_no_text / text_heavy variants).
Follows NADEZHDA code style.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

conf/balanced.gin

@@ -0,0 +1,54 @@
+# Balanced configuration — primary test setup.
+# L = 1.0 * L_img_img + 0.3 * L_sat_cap + 0.3 * L_drone_cap + 0.1 * L_cap_cap
+
+import src.datasets.visloc_with_captions
+import src.losses.multi_infonce
+import src.models.dual_encoder
+import src.training.train
+
+# ---- Dual encoder ----
+DualEncoderCaptionTest.variant = "ViT-B-32"
+DualEncoderCaptionTest.pretrained_path = "checkpoints/RS5M_ViT-B-32.pt"
+DualEncoderCaptionTest.unfreeze_mode = "last_block"
+DualEncoderCaptionTest.embed_dim = 512
+DualEncoderCaptionTest.use_mlp_heads = False
+DualEncoderCaptionTest.shared_image_head = True
+DualEncoderCaptionTest.device = "cuda"
+
+ProjectionHead.in_dim = 512
+ProjectionHead.out_dim = 512
+ProjectionHead.use_mlp = False
+
+# ---- Loss ----
+MultiTermInfoNCE.temperature_init = 0.1
+MultiTermInfoNCE.temperature_final = 0.01
+MultiTermInfoNCE.label_smoothing = 0.1
+MultiTermInfoNCE.asym_drone_to_sat = 0.6
+MultiTermInfoNCE.asym_sat_to_drone = 0.4
+MultiTermInfoNCE.warmup_epochs = 3
+MultiTermInfoNCE.text_ramp_epochs = 10
+MultiTermInfoNCE.lambda_ii = 1.0
+MultiTermInfoNCE.lambda_sc_max = 0.3
+MultiTermInfoNCE.lambda_dc_max = 0.3
+MultiTermInfoNCE.lambda_cc_max = 0.1
+
+# ---- Dataset ----
+VisLocCaptionDataset.caption_strategy = "hybrid"
+VisLocCaptionDataset.drop_caption_prob = 0.0
+VisLocCaptionDataset.seed = 42
+
+# ---- Training ----
+TrainConfig.train_manifest = "data/visloc_train.json"
+TrainConfig.val_manifest = "data/visloc_val.json"
+TrainConfig.image_root = "data/visloc/images"
+TrainConfig.output_dir = "out/caption_test/balanced"
+TrainConfig.epochs = 30
+TrainConfig.batch_size = 128
+TrainConfig.num_workers = 4
+TrainConfig.learning_rate = 1e-4
+TrainConfig.weight_decay = 1e-4
+TrainConfig.grad_clip = 1.0
+TrainConfig.use_amp = True
+TrainConfig.eval_every = 1
+TrainConfig.seed = 42
+TrainConfig.device = "cuda"

conf/baseline_no_text.gin

@@ -0,0 +1,11 @@
+# Baseline: image-image only, no captions. Reference R@1 for delta computation.
+# L = 1.0 * L_img_img
+
+include 'balanced.gin'
+
+# Disable all caption loss terms.
+MultiTermInfoNCE.lambda_sc_max = 0.0
+MultiTermInfoNCE.lambda_dc_max = 0.0
+MultiTermInfoNCE.lambda_cc_max = 0.0
+
+TrainConfig.output_dir = "out/caption_test/baseline_no_text"

conf/text_heavy.gin

@@ -0,0 +1,11 @@
+# Text-heavy configuration — stress test of caption contribution.
+# L = 0.5 * L_img_img + 0.5 * L_sat_cap + 0.5 * L_drone_cap + 0.2 * L_cap_cap
+
+include 'balanced.gin'
+
+MultiTermInfoNCE.lambda_ii = 0.5
+MultiTermInfoNCE.lambda_sc_max = 0.5
+MultiTermInfoNCE.lambda_dc_max = 0.5
+MultiTermInfoNCE.lambda_cc_max = 0.2
+
+TrainConfig.output_dir = "out/caption_test/text_heavy"