Spaces:
Sleeping
Sleeping
File size: 2,748 Bytes
d70361b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 | {
"v3_frozen": {
"path": "models/adaptformer_delhi/v3_frozen",
"val_f1": 0.6771,
"test_f1": 0.5809,
"test_precision": 0.6782,
"test_recall": 0.5351,
"test_iou": 0.4117,
"threshold": 0.2,
"loss": "tversky",
"epochs": 7,
"n_train_pairs": 16,
"status": "BASELINE — do not overwrite"
},
"thr_sweep_0.10_0.40": {
"artifact": "runs/v3_baseline_analysis/analysis.json",
"best_val_thr": 0.35,
"best_val_f1": 0.6917,
"test_at_0.35": {"f1": 0.5855, "P": 0.725, "R": 0.516, "IoU": 0.418},
"test_at_0.20": {"f1": 0.5954},
"decision": "Keep thr=0.20 for recall. Raising to 0.35 helps Val F1/P but drops Test R. Prefer post-filter low-conf Other on GeoTIFF reports."
},
"pdf_report_vulnerabilities": {
"file": "DDA_Report_37_Grid_54_tif_vs_H43X2E1_tif.pdf",
"findings": [
"60 regions / 3.08% change — many Other at 26-50% conf (FP risk from thr=0.2)",
"Frequent Vegetation Change — seasonal/spectral FPs on 10m data",
"Largest New Construction 107k px — check edge completeness (FN)",
"All reviews pending — no human FP feedback into training"
],
"ops_fix_applied": "DDA_OTHER_MIN_CONF=0.5 filter in detect_service (drops weak Other)"
},
"error_analysis": {
"panels": "runs/v3_baseline_analysis/panels (24)",
"fp_dir": "runs/v3_baseline_analysis/fp",
"fn_dir": "runs/v3_baseline_analysis/fn",
"hardest_fn": ["delhi_0030", "delhi_0020", "delhi_0023", "delhi_0005", "delhi_0017"],
"note": "Large settlement polygons nearly total miss — main recall bottleneck"
},
"experiments": [
{
"id": "v3_baseline",
"variable": "none",
"status": "frozen",
"val_f1": 0.6771,
"test_f1": 0.5809,
"test_P": 0.6782,
"test_R": 0.5351,
"test_iou": 0.4117,
"threshold": 0.2,
"loss": "tversky",
"epochs": 7,
"n_train": 16
},
{
"id": "v4_pos_sampling",
"variable": "pos_only + oversample×6 + min_tile_change 0.02 (loss=tversky unchanged)",
"status": "done_no_improve",
"warm_start": "models/adaptformer_delhi/v3_frozen",
"run": "runs/finetune_v4/20260717_211854",
"val_f1": 0.6682,
"test_f1": 0.5759,
"test_P": 0.6689,
"test_R": 0.5341,
"test_iou": 0.4064,
"threshold": 0.1,
"loss": "tversky",
"epochs": 9,
"n_train": 16,
"verdict": "REJECT — slightly worse than v3; thr collapsed to 0.1; keep v3"
}
],
"next": [
"Loss ablation only (warm-start v3_frozen): bce_dice vs tversky, sampling unchanged",
"More labeled large-settlement pairs (0030/0020 failure mode)",
"Ops: keep thr=0.2 + Other conf filter; optional veg suppress"
]
}
|