File size: 9,775 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 | #!/usr/bin/env python3
"""Summarize paper TBD ablation evaluations into Markdown/LaTeX-ready tables."""
from __future__ import annotations
import argparse
import json
import re
from pathlib import Path
from typing import Any, Dict, Iterable, List, Optional, Tuple
BASELINE_JSON = {
"fleurs": "fleurs_en_us_clean.json",
"voxpopuli": "voxpopuli_en_clean.json",
}
def as_float(value: Any) -> Optional[float]:
if value is None:
return None
try:
return float(value)
except Exception:
return None
def pct(value: Optional[float], digits: int = 2) -> str:
if value is None:
return "-"
return f"{value * 100.0:.{digits}f}"
def pp(value: Optional[float], digits: int = 2) -> str:
if value is None:
return "-"
return f"{value * 100.0:+.{digits}f}"
def read_json(path: Path) -> Dict[str, Any]:
return json.loads(path.read_text(encoding="utf-8"))
def weighted(rows: Iterable[Dict[str, Any]], key: str) -> Optional[float]:
numer = 0.0
denom = 0
for row in rows:
value = as_float(row.get(key))
count = int(row.get("samples_used") or 0)
if value is None or count <= 0:
continue
numer += value * count
denom += count
if denom <= 0:
return None
return numer / denom
def baseline_metric(baseline_dir: Path, dataset_tag: str, metric: str) -> Optional[float]:
path = baseline_dir / BASELINE_JSON[dataset_tag]
payload = read_json(path)
summary = payload.get("summary") or {}
value = as_float(summary.get(f"base_model_weighted_{metric}"))
if value is not None:
return value
return weighted(payload.get("rows") or [], f"base_model_{metric}")
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]]:
total = sum(dist.values())
if total <= 0:
return None, None
avg_steps = sum(step * count for step, count in dist.items()) / total
skip = dist.get(0, 0) / total
return avg_steps, skip
def parse_eval_name(path: Path) -> Optional[Tuple[str, str, str]]:
marker = "_theta_"
stem = path.stem
if marker not in stem:
return None
prefix, theta = stem.split(marker, 1)
for dataset_tag in ("voxpopuli", "fleurs"):
suffix = f"_{dataset_tag}"
if prefix.endswith(suffix):
return prefix[: -len(suffix)], dataset_tag, theta
return None
def load_records(out_dir: Path, baseline_dir: Path) -> Dict[Tuple[str, str, str], Dict[str, Any]]:
records: Dict[Tuple[str, str, str], Dict[str, Any]] = {}
for path in sorted(out_dir.glob("*_theta_*.json")):
parsed = parse_eval_name(path)
if not parsed:
continue
variant, dataset_tag, theta = parsed
payload = read_json(path)
rows = payload.get("rows") or []
summary = payload.get("summary") or {}
wer = as_float(summary.get("latent_reasoning_weighted_wer"))
cer = as_float(summary.get("latent_reasoning_weighted_cer"))
if wer is None:
wer = weighted(rows, "latent_reasoning_wer")
if cer is None:
cer = weighted(rows, "latent_reasoning_cer")
base_wer = baseline_metric(baseline_dir, dataset_tag, "wer")
base_cer = baseline_metric(baseline_dir, dataset_tag, "cer")
dist = parse_step_distribution(out_dir / "logs" / f"{path.stem}.log")
avg_steps, skip = step_stats(dist)
key = (variant, dataset_tag, theta)
records[key] = {
"variant": variant,
"dataset": dataset_tag,
"theta": theta,
"wer": wer,
"cer": cer,
"base_wer": base_wer,
"base_cer": base_cer,
"dwer_pp": None if wer is None or base_wer is None else wer - base_wer,
"dcer_pp": None if cer is None or base_cer is None else cer - base_cer,
"avg_steps": avg_steps,
"skip": skip,
"dist": dist,
"json": path.name,
}
return records
def rec(
records: Dict[Tuple[str, str, str], Dict[str, Any]],
variant: str,
dataset: str = "fleurs",
theta: str = "zero",
) -> Optional[Dict[str, Any]]:
return records.get((variant, dataset, theta))
def table_component(records: Dict[Tuple[str, str, str], Dict[str, Any]]) -> List[str]:
labels = [
("n4", "Full \\method{} ($N{=}4$, $\\theta{=}0.0$)"),
("component_no_bounded", "\\quad $-$ bounded delta ($L_2$ + scale $s_k$)"),
("component_no_gate", "\\quad $-$ sigmoid gate ($g_k$ fixed at $1$)"),
("component_no_anchor", "\\quad $-$ fixed-embedding anchor ($\\mathbf{e}_{\\texttt{LT}}$ removed)"),
]
lines = ["### Component Ablation", "", "| Variant | WER (%) | ΔWER (pp) |", "|---|---:|---:|"]
for variant, label in labels:
r = rec(records, variant)
lines.append(f"| {label} | {pct(r['wer']) if r else '-'} | {pp(r['dwer_pp']) if r else '-'} |")
return lines
def table_n_sweep(records: Dict[Tuple[str, str, str], Dict[str, Any]]) -> List[str]:
variants = [("n1", "1"), ("n2", "2"), ("n4", "\\textbf{4}"), ("n8", "8")]
lines = [
"### N Sweep",
"",
"| N | FLEURS WER (%) | ΔWER (pp) | VoxPopuli WER (%) | ΔWER (pp) |",
"|---:|---:|---:|---:|---:|",
]
for variant, label in variants:
f = rec(records, variant, "fleurs")
v = rec(records, variant, "voxpopuli")
lines.append(
f"| {label} | {pct(f['wer']) if f else '-'} | {pp(f['dwer_pp']) if f else '-'} | "
f"{pct(v['wer']) if v else '-'} | {pp(v['dwer_pp']) if v else '-'} |"
)
return lines
def table_pneg(records: Dict[Tuple[str, str, str], Dict[str, Any]]) -> List[str]:
full = rec(records, "n4")
p0 = rec(records, "pneg0")
skips = []
for theta in ("full", "neg0p2", "zero", "pos0p2", "pos0p5"):
row = rec(records, "pneg0", "fleurs", theta)
if row and row["skip"] is not None:
skips.append(row["skip"])
pos = rec(records, "pneg0", "fleurs", "pos0p2")
skip_at_pos = pos["skip"] if pos else None
skip_range = "-" if not skips else f"[{min(skips) * 100.0:.1f}, {max(skips) * 100.0:.1f}]"
lines = [
"### Forced-Negative Sampling",
"",
"| Setting | WER (%) | ΔWER (pp) | Skip @ θ=+0.2 | Skip range (%) |",
"|---|---:|---:|---:|---:|",
f"| Full ($p_{{\\text{{neg}}}}{{=}}0.3$) | {pct(full['wer']) if full else '-'} | {pp(full['dwer_pp']) if full else '-'} | 100.0% | [0, 100] |",
f"| $-$ Forced-neg ($p_{{\\text{{neg}}}}{{=}}0.0$) | {pct(p0['wer']) if p0 else '-'} | {pp(p0['dwer_pp']) if p0 else '-'} | {('-' if skip_at_pos is None else f'{skip_at_pos * 100.0:.1f}%')} | {skip_range} |",
]
return lines
def table_activation(records: Dict[Tuple[str, str, str], Dict[str, Any]]) -> List[str]:
variants = [(f"activation_{n}", str(n)) for n in range(100, 801, 100)]
lines = [
"### Activation Set Scaling",
"",
"| #utts | FLEURS WER (%) | ΔWER (pp) | VoxPopuli WER (%) | ΔWER (pp) |",
"|---:|---:|---:|---:|---:|",
]
for variant, label in variants:
f = rec(records, variant, "fleurs")
v = rec(records, variant, "voxpopuli")
lines.append(
f"| {label} | {pct(f['wer']) if f else '-'} | {pp(f['dwer_pp']) if f else '-'} | "
f"{pct(v['wer']) if v else '-'} | {pp(v['dwer_pp']) if v else '-'} |"
)
return lines
def table_pneg_sweep(records: Dict[Tuple[str, str, str], Dict[str, Any]]) -> List[str]:
theta_values = {
"full": "-2.0",
"neg0p2": "-0.2",
"zero": "0.0",
"pos0p2": "+0.2",
"pos0p5": "+0.5",
}
lines = [
"### p_neg=0.0 FLEURS Threshold Details",
"",
"| θ | Avg steps | Skip (%) | WER (%) | ΔWER (pp) |",
"|---:|---:|---:|---:|---:|",
]
for theta in ("full", "neg0p2", "zero", "pos0p2", "pos0p5"):
r = rec(records, "pneg0", "fleurs", theta)
if not r:
lines.append(f"| {theta_values[theta]} | - | - | - | - |")
continue
avg = "-" if r["avg_steps"] is None else f"{r['avg_steps']:.2f}"
skip = "-" if r["skip"] is None else f"{r['skip'] * 100.0:.1f}"
lines.append(f"| {theta_values[theta]} | {avg} | {skip} | {pct(r['wer'])} | {pp(r['dwer_pp'])} |")
return lines
def write_report(out_dir: Path, records: Dict[Tuple[str, str, str], Dict[str, Any]]) -> Path:
lines: List[str] = ["# Paper TBD Results", ""]
for section in (
table_component(records),
table_n_sweep(records),
table_pneg(records),
table_pneg_sweep(records),
table_activation(records),
):
lines.extend(section)
lines.append("")
path = out_dir / "paper_tbd_results.md"
path.write_text("\n".join(lines), encoding="utf-8")
return path
def main() -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("out_dir", type=Path)
parser.add_argument("--baseline-dir", type=Path, required=True)
args = parser.parse_args()
records = load_records(args.out_dir, args.baseline_dir)
report = write_report(args.out_dir, records)
print(report)
print(report.read_text(encoding="utf-8"))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|