File size: 5,037 Bytes
d775561
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
"""
Stage 03b (v8b): select the layers to KEEP.

Keep a layer iff, over the calibration sub-sweep 0.7 -> 0.3 -> 0.0:
  (a) the reflection reduction is monotonic (non-decreasing as alpha
      drops from the 1.0 baseline down to 0.0, within `slack`), AND
  (b) EVERY one of those three alpha reductions is STRICTLY > min_reduction.

i.e. starting from the alpha=1.0 baseline (reduction 0), reflection
"steps down" at every alpha and each step removes more than `min_reduction`
markers. No top-N truncation — every qualifying layer is kept.

Reads:  calibration_monitoring.json (from stage 03, read-only)
Writes: selected_layers_monitoring.json
"""
import os, sys, argparse
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))

from configs import get_config
from configs.paths import LOG_DIR, dim_paths, ensure_dirs
from src.utils import read_json, write_json, setup_logger

DEFAULT_SLACK = 0.5
SUBSWEEP = ["0.70", "0.30", "0.00"]   # 1.00 is the baseline (reduction 0)


def _red(sweep_detail, key):
    d = sweep_detail.get(key)
    return float(d["avg_reduction"]) if d else None


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--dimension", default="monitoring")
    ap.add_argument("--slack", type=float, default=DEFAULT_SLACK,
                    help="Monotonicity slack (reduction may dip by up to this).")
    ap.add_argument("--min-reduction", type=float, default=1.0,
                    help="Keep layer iff EVERY alpha reduction is STRICTLY > this.")
    args = ap.parse_args()

    ensure_dirs(args.dimension)
    cfg = get_config(args.dimension)
    p = dim_paths(args.dimension)
    out_path = p.SELECTED_LAYERS

    log = setup_logger("03b_select", os.path.join(LOG_DIR, f"03b_select_{cfg.NAME}.log"))
    log.info("=" * 72)
    log.info(f"Stage 03b [{cfg.NAME}] select KEEP layers")
    log.info(f"  criterion: monotonic over {SUBSWEEP} (slack={args.slack}) "
             f"AND EVERY alpha reduction > {args.min_reduction}")
    log.info(f"  output -> {out_path}")
    log.info("=" * 72)

    if not os.path.exists(p.CALIBRATION):
        log.error(f"missing {p.CALIBRATION} — run stage 03 first"); sys.exit(1)
    calib = read_json(p.CALIBRATION)
    per_layer = calib.get("calibration_per_layer", {})
    log.info(f"  layers in calibration: {len(per_layer)}")

    cands = []
    for L_str, d in per_layer.items():
        L = int(L_str)
        sd = d.get("sweep_detail", {})
        reds = [_red(sd, k) for k in SUBSWEEP]
        if any(r is None for r in reds):
            log.info(f"  L{L}: SKIP (missing sweep points)")
            continue

        mono = all(reds[i] >= reds[i - 1] - args.slack for i in range(1, len(reds)))
        achievable = max(reds)
        best_idx = max(range(len(reds)), key=lambda i: reds[i])
        best_alpha = float(SUBSWEEP[best_idx])

        if not mono:
            log.info(f"  L{L}: SKIP (not mono 0.7->0; reds={['%.2f' % r for r in reds]})")
            continue
        if not all(r > args.min_reduction for r in reds):
            log.info(f"  L{L}: SKIP (some alpha <= {args.min_reduction}; "
                     f"reds={['%.2f' % r for r in reds]})")
            continue

        cands.append({
            "layer": L,
            "alpha": best_alpha,
            "achieved_red": achievable,
            "reds_0p7_0p3_0p0": [round(r, 3) for r in reds],
            "fully_monotonic": bool(d.get("fully_monotonic", False)),
            "safe": bool(d.get("safe", False)),
        })

    if not cands:
        log.error("No layers passed the 0.7->0 mono + every-alpha>min filter.")
        sys.exit(2)

    cands.sort(key=lambda x: x["layer"])
    log.info(f"  KEEP {len(cands)} layers:")
    cum_r = 0.0
    for it in cands:
        cum_r += it["achieved_red"]
        ff = "" if it["fully_monotonic"] else " [not-full-mono]"
        log.info(f"    L{it['layer']:>2}  best_a={it['alpha']:.2f}  "
                 f"red={it['achieved_red']:+.2f}  "
                 f"reds(0.7/0.3/0)={it['reds_0p7_0p3_0p0']}{ff}")
    log.info(f"  cumulative achievable reduction: {cum_r:.2f}")

    alpha_per_layer = {it["layer"]: it["alpha"] for it in cands}
    out = {
        "dimension": cfg.NAME,
        "selected_layers": sorted(alpha_per_layer.keys()),
        "alpha_per_layer": {str(L): v for L, v in alpha_per_layer.items()},
        "work_alpha": 0.7,
        "n_selected": len(alpha_per_layer),
        "cumulative_reduction": cum_r,
        "policy": "every_alpha_gt_min_and_monotonic_no_topn",
        "policy_params": {
            "subsweep": SUBSWEEP,
            "slack": args.slack,
            "min_reduction": args.min_reduction,
            "rule": "EVERY alpha reduction > min AND monotonic from baseline",
        },
        "per_layer_diagnostics": cands,
    }
    write_json(out, out_path)
    log.info(f"Saved {out_path}  ({len(alpha_per_layer)} layers). Done.")
    print("SELECTED_LAYERS =", sorted(alpha_per_layer.keys()))
    print("OUTPUT =", out_path)


if __name__ == "__main__":
    main()