Spaces:
Running
Running
File size: 1,751 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 | {
"plan_source": "Claude priority improvement plan",
"run": "runs/finetune_v2/20260716_210208",
"weights": "models/adaptformer_delhi/best_v2",
"logit_probe": {
"verdict": "CORRECT — AdaptFormer logits (N,2,H,W); change = softmax channel 1 (last)",
"why_5e-5_before": "Hub Delhi probs peak ~1e-2 with mean ~1e-4; after CE-only fix, calibration grid favored tiny thr. With focal+dice + stratified split, thr converges to ~0.5"
},
"class_imbalance": {
"train_bg_to_change": "≈11.5:1",
"note": "Previous non-stratified split had val change%≈25% vs test≈7% — inflated Val F1 and Val/Test gap"
},
"implemented": [
"1 logit→prob probe + output_probe.json",
"2 auto thr = fixed grid + score quantiles; freeze with best ckpt for test",
"3 imbalance report; Focal+Dice; positive-tile oversample×3 + WeightedRandomSampler",
"4 Precision/Recall/IoU logged each epoch",
"5 visuals/ Before|After|GT|Pred|Prob each epoch",
"6 lazy tile index (pair_i, kind, x, y)",
"7 brightness/contrast + rot + flips; ReduceLROnPlateau",
"8 stratified split (--stratify)"
],
"scores": {
"day5_broken": {"val_f1": 0.0001, "test_f1": 0.0025, "thr": 0.5},
"fix_6ep": {"val_f1": 0.3664, "test_f1": 0.1389, "thr": 5e-05},
"v2_12ep": {
"best_val_f1": 0.5926,
"test_f1": 0.3975,
"test_precision": 0.557,
"test_recall": 0.34,
"test_iou": 0.272,
"frozen_threshold": 0.5
}
},
"targets": {
"val_f1_0.5_plus": "HIT (0.59)",
"test_closer_to_val": "IMPROVED (0.14→0.40; gap 0.23→0.20). Still limited by N=24 and domain shift; next: more labels / recall-focused loss weight"
},
"load": "ADAPTFORMER_WEIGHTS=models/adaptformer_delhi/best_v2"
}
|