File size: 8,429 Bytes
262fa3f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 | #!/usr/bin/env python3
"""Summarize LatentASR threshold-sweep JSON/log files."""
from __future__ import annotations
import argparse
import json
import re
from datetime import datetime, timezone
from pathlib import Path
from typing import Any, Dict, Iterable, List, Optional, Tuple
DEFAULT_BASELINE_DIR = Path("eval_runs/hf_asr_showcase_full_20260503_152506")
BASELINE_JSON = {
"fleurs_en_us": "fleurs_en_us_clean.json",
"voxpopuli_en": "voxpopuli_en_clean.json",
}
DATASET_LABEL = {
"fleurs_en_us": "FLEURS en-US",
"voxpopuli_en": "VoxPopuli en",
}
THETA_VALUE = {
"full": "-2.0",
"neg0p2": "-0.2",
"zero": "0.0",
"pos0p2": "0.2",
"pos0p5": "0.5",
}
def _as_float(value: Any) -> Optional[float]:
if value is None:
return None
try:
return float(value)
except Exception:
return None
def _as_int(value: Any) -> int:
try:
return int(value or 0)
except Exception:
return 0
def _pct(value: Optional[float]) -> str:
if value is None:
return "-"
return f"{value * 100.0:.3f}"
def _pp(value: Optional[float]) -> str:
if value is None:
return "-"
return f"{value * 100.0:+.3f}"
def _fmt_float(value: Optional[float]) -> str:
if value is None:
return "-"
return f"{value:.2f}"
def _read_json(path: Path) -> Dict[str, Any]:
return json.loads(path.read_text(encoding="utf-8"))
def _sample_count(rows: Iterable[Dict[str, Any]]) -> int:
return sum(_as_int(row.get("samples_used")) for row in rows)
def _weighted(rows: List[Dict[str, Any]], key: str) -> Optional[float]:
numer = 0.0
denom = 0
for row in rows:
value = _as_float(row.get(key))
count = _as_int(row.get("samples_used"))
if value is None or count <= 0:
continue
numer += value * count
denom += count
if denom <= 0:
return None
return numer / denom
def _baseline_metrics(baseline_dir: Path, tag: str) -> Tuple[Optional[float], Optional[float]]:
name = BASELINE_JSON.get(tag)
if not name:
return None, None
path = baseline_dir / name
if not path.exists():
return None, None
payload = _read_json(path)
summary = payload.get("summary") or {}
base_wer = _as_float(summary.get("base_model_weighted_wer"))
base_cer = _as_float(summary.get("base_model_weighted_cer"))
if base_wer is None:
base_wer = _weighted(payload.get("rows") or [], "base_model_wer")
if base_cer is None:
base_cer = _weighted(payload.get("rows") or [], "base_model_cer")
return base_wer, base_cer
def _parse_step_distribution(log_path: Path) -> Dict[int, int]:
if not log_path.exists():
return {}
text = log_path.read_text(encoding="utf-8", errors="ignore").replace("\r", "\n")
dist: Dict[int, int] = {}
for step, count in re.findall(r"N=(\d+):\s+(\d+)\s+\(", text):
dist[int(step)] = int(count)
return dist
def _step_stats(dist: Dict[int, int]) -> Tuple[Optional[float], Optional[float], Optional[float]]:
total = sum(dist.values())
if total <= 0:
return None, None, None
avg_steps = sum(step * count for step, count in dist.items()) / total
skip_rate = dist.get(0, 0) / total
full_rate = dist.get(4, 0) / total
return avg_steps, skip_rate, full_rate
def _tag_theta(path: Path) -> Tuple[str, str]:
stem = path.stem
marker = "_theta_"
if marker not in stem:
return stem, "unknown"
tag, theta = stem.split(marker, 1)
return tag, theta
def load_records(out_dir: Path, baseline_dir: Path) -> List[Dict[str, Any]]:
records: List[Dict[str, Any]] = []
for json_path in sorted(out_dir.glob("*_theta_*.json")):
tag, theta = _tag_theta(json_path)
payload = _read_json(json_path)
rows = payload.get("rows") or []
summary = payload.get("summary") or {}
base_wer, base_cer = _baseline_metrics(baseline_dir, tag)
lr_wer = _as_float(summary.get("latent_reasoning_weighted_wer"))
lr_cer = _as_float(summary.get("latent_reasoning_weighted_cer"))
if lr_wer is None:
lr_wer = _weighted(rows, "latent_reasoning_wer")
if lr_cer is None:
lr_cer = _weighted(rows, "latent_reasoning_cer")
dist = _parse_step_distribution(out_dir / "logs" / f"{tag}_theta_{theta}.log")
avg_steps, skip_rate, full_rate = _step_stats(dist)
records.append(
{
"tag": tag,
"dataset": DATASET_LABEL.get(tag, tag),
"theta": theta,
"theta_value": THETA_VALUE.get(theta, theta),
"n": _sample_count(rows),
"base_wer": base_wer,
"lr_wer": lr_wer,
"dwer": None if base_wer is None or lr_wer is None else base_wer - lr_wer,
"base_cer": base_cer,
"lr_cer": lr_cer,
"dcer": None if base_cer is None or lr_cer is None else base_cer - lr_cer,
"avg_steps": avg_steps,
"skip_rate": skip_rate,
"full_rate": full_rate,
"step_dist": dist,
"json": json_path.name,
}
)
return records
def _sort_key(record: Dict[str, Any]) -> Tuple[str, int]:
order = {"full": 0, "neg0p2": 1, "zero": 2, "pos0p2": 3, "pos0p5": 4}
return record["tag"], order.get(record["theta"], 99)
def _table(records: List[Dict[str, Any]]) -> List[str]:
lines = [
"| Dataset | Theta | N | Avg steps | Skip N=0 | Full N=4 | Base WER | LR WER | dWER | Base CER | LR CER | dCER |",
"|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|",
]
for rec in sorted(records, key=_sort_key):
lines.append(
"| {dataset} | {theta} | {n} | {avg_steps} | {skip} | {full} | {base_wer} | {lr_wer} | {dwer} | {base_cer} | {lr_cer} | {dcer} |".format(
dataset=rec["dataset"],
theta=rec["theta_value"],
n=rec["n"],
avg_steps=_fmt_float(rec["avg_steps"]),
skip=_pct(rec["skip_rate"]),
full=_pct(rec["full_rate"]),
base_wer=_pct(rec["base_wer"]),
lr_wer=_pct(rec["lr_wer"]),
dwer=_pp(rec["dwer"]),
base_cer=_pct(rec["base_cer"]),
lr_cer=_pct(rec["lr_cer"]),
dcer=_pp(rec["dcer"]),
)
)
return lines
def write_report(out_dir: Path, records: List[Dict[str, Any]]) -> Path:
report_path = out_dir / "threshold_sweep_report.md"
lines: List[str] = []
lines.append("# LatentASR Threshold Sweep")
lines.append("")
lines.append(f"Generated UTC: {datetime.now(timezone.utc).isoformat(timespec='seconds')}")
lines.append("")
lines.append("`Theta=-2.0` is the no-halting control: value-head scores are bounded in [-1, 1], so no early halt is triggered.")
lines.append("`dWER = Base WER - LR WER`; positive values mean LR is better. WER/CER and rates are percentages.")
lines.append("")
lines.extend(_table(records))
lines.append("")
lines.append("## Step Distributions")
lines.append("")
lines.append("| Dataset | Theta | N=0 | N=1 | N=2 | N=3 | N=4 |")
lines.append("|---|---:|---:|---:|---:|---:|---:|")
for rec in sorted(records, key=_sort_key):
dist = rec["step_dist"]
lines.append(
"| {dataset} | {theta} | {n0} | {n1} | {n2} | {n3} | {n4} |".format(
dataset=rec["dataset"],
theta=rec["theta_value"],
n0=dist.get(0, 0),
n1=dist.get(1, 0),
n2=dist.get(2, 0),
n3=dist.get(3, 0),
n4=dist.get(4, 0),
)
)
lines.append("")
report_path.write_text("\n".join(lines), encoding="utf-8")
return report_path
def main() -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("out_dir", type=Path)
parser.add_argument("--baseline-dir", type=Path, default=DEFAULT_BASELINE_DIR)
args = parser.parse_args()
records = load_records(args.out_dir, args.baseline_dir)
report = write_report(args.out_dir, records)
print(report)
print("\n".join(_table(records)))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|