File size: 14,074 Bytes
7ed86c3 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 | """Inference + benchmark glue for a trained MeterModel.
# Norma benchmark, both tasks, scored in-process (--norma-source hf|git, default hf):
python -m meter.predict --model $STOICHEIA_DATA/runs/meter_joint/best.pt --norma \
[--pred-out data/norma_preds.jsonl] [--norma-source hf]
# work-split scanner dev/test:
python -m meter.predict --model ... --scan-split
# production:
python -m meter.predict --model ... --macronize in.txt out.txt
python -m meter.predict --model ... --scan in.txt out.txt
--norma scores both tasks directly via meter.norma_score (acc, balanced acc, per-class
F1 for macronize; acc, balanced acc, boundary-F1, weight acc for syllabify), split into
dev/test, and additionally dumps macron predictions/probabilities in the old
predictions-jsonl format for anyone still using the legacy ensemble/scorer scripts.
"""
from __future__ import annotations
import argparse, json, os, sys
from collections import defaultdict
from pathlib import Path
import numpy as np
import torch
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from meter.backbone import load_backbone
from meter.dataset import batch_rows, encode_plain, load_records, pack_records
from meter.marks import (ambiguous_mask, bracketize, enforce_circumflex_heavy,
insert_marks, merge_vowelless_syllables,
parse_macron_line, parse_scan_line)
from meter.model import MeterConfig, MeterModel
from meter.norma_data import add_norma_source_arg
from meter.norma_score import scan_metrics
from meter.train import SCAN_DEV_WORKS, SCAN_TEST_WORKS
def load_model(path, device, attn="sdpa"):
sd = torch.load(path, map_location="cpu")
encoder, _ = load_backbone(os.path.expandvars(sd["cfg"]["ckpt"]), device, attn)
mcfg = MeterConfig(**sd["mcfg"])
model = MeterModel(encoder, mcfg).to(device)
model.load_state_dict(sd["model"])
model.eval()
return model, sd
@torch.no_grad()
def predict_records(model, records, T, micro, device):
"""-> per record (mac_argmax, scan_argmax, mac_P(long), scan_probs) with None
entries preserved. mac_P(long) is P(class 0) per letter; scan_probs is (n,4).
scan_argmax has two deterministic corrections applied, in order: (1)
merge_vowelless_syllables -- a predicted syllable span with no vowel gets
folded into the preceding one, keeping its own (usually already-correct)
weight; (2) enforce_circumflex_heavy -- a circumflexed syllable is always
heavy. Every consumer (scoring, --scan production output, Viterbi) gets both
fixed rules for free."""
live = [i for i, r in enumerate(records) if r is not None]
rows, skipped = pack_records(records, T, live)
if skipped:
print(f" WARNING: {skipped} records longer than T={T} skipped", file=sys.stderr)
out = [None] * len(records)
for i in range(0, len(rows), micro):
chunk = rows[i:i + micro]
batch = batch_rows(chunk, records, T, device=device, with_slots=True)
with torch.autocast("cuda", dtype=torch.bfloat16, enabled=device.type == "cuda"):
o = model({k: v for k, v in batch.items() if k != "slots"})
mac_p = torch.softmax(o["mac"].float(), -1).cpu().numpy()
scan_p = torch.softmax(o["scan"].float(), -1).cpu().numpy()
pm = mac_p.argmax(-1)
ps = scan_p.argmax(-1)
for b, slots in enumerate(batch["slots"]):
for ri, c in slots:
n = len(records[ri])
scan_argmax = merge_vowelless_syllables(records[ri].chars, ps[b, c:c + n])
scan_argmax = enforce_circumflex_heavy(records[ri].dia, scan_argmax)
out[ri] = (pm[b, c:c + n], scan_argmax,
mac_p[b, c:c + n, 0], scan_p[b, c:c + n])
return out
def run_norma(model, sd, device, micro, pred_out, norma_source="hf"):
from meter.norma_data import load_norma
from meter.norma_score import MAC_LONG, mac_metrics
T = sd["T"]
norma = load_norma(norma_source)
# -------- macronize: scored directly (in-process; no external scorer needed)
probs_rows = []
with open(pred_out, "w", encoding="utf-8") as f:
for split_name, rows in [("dev", norma["dev"]), ("test", norma["test"])]:
mac_rows = [d for d in rows if d["task"] == "macronize"]
items = [(d["source"], *parse_macron_line(d["text"])) for d in mac_rows]
recs = [encode_plain(p) for _, p, _ in items]
preds = predict_records(model, recs, T, micro, device)
pairs, by_source = [], defaultdict(list)
for (srcname, plain, gold), pr in zip(items, preds):
g = sorted(gold.items())
pred_labels = [(k, int(pr[0][k]) if pr is not None else MAC_LONG) for k, _ in g]
f.write(json.dumps(dict(
split=split_name, source=srcname,
gold=[[k, "", v] for k, v in g],
pred=[[k, "", pv] for k, pv in pred_labels]), ensure_ascii=False) + "\n")
if pr is not None:
probs_rows.append(dict(split=split_name, source=srcname, seq=[
[v, float(pr[2][k])] for k, v in g]))
for (k, gv), (_, pv) in zip(g, pred_labels):
pairs.append((gv, pv))
by_source[srcname].append((gv, pv))
overall = mac_metrics(pairs)
per_source = {s: mac_metrics(ps) for s, ps in by_source.items()}
bals = [m["bal_acc"] for m in per_source.values() if m and m["bal_acc"] is not None]
macro = round(sum(bals) / len(bals), 4) if bals else None
print(f"macron {split_name}: {json.dumps(dict(**overall, macro_bal_acc=macro))}"
if overall else f"macron {split_name}: (empty)")
probs_out = str(pred_out).replace(".jsonl", "") + "_probs.json"
json.dump(probs_rows, open(probs_out, "w"))
print(f"macron P(long) dump -> {probs_out}")
print(f"macron preds -> {pred_out}")
# -------- syllabify: test only (Norma has no syllabify dev rows, on hf or git)
syl_items, syl_recs = [], []
for d in norma["test"]:
if d["task"] != "syllabify":
continue
parsed = parse_scan_line(d["text"])
if parsed is None:
continue
plain, gold = parsed
rec = encode_plain(plain)
if rec is None:
continue
syl_items.append((d["source"], gold))
syl_recs.append(rec)
preds = predict_records(model, syl_recs, T, micro, device)
print("\n=== norma syllabify (test) ===")
pairs, by_src = [], defaultdict(list)
for (srcname, gold), pr in zip(syl_items, preds):
if pr is None:
continue
golds = np.zeros(len(pr[1]), dtype=np.int64)
for k, g in gold.items():
golds[k] = g
rec_pairs = list(zip(golds.tolist(), pr[1].tolist()))
pairs += rec_pairs
by_src[srcname] += rec_pairs
m = scan_metrics(pairs)
bals = [b["bal_acc"] for s in sorted(by_src) if (b := scan_metrics(by_src[s]))]
macro = round(sum(bals) / len(bals), 4) if bals else None
print(f" {json.dumps(dict(**m, macro_bal_acc=macro))}")
def run_scan_split(model, sd, device, micro):
T = sd["T"]
enc_dir = Path(os.path.expandvars(sd["cfg"]["encoded"]))
recs, works = load_records(enc_dir / "scan_corpus.npz")
for split, wanted in (("dev", SCAN_DEV_WORKS), ("test", SCAN_TEST_WORKS)):
sel = [(r, w) for r, w in zip(recs, works) if w in wanted]
preds = predict_records(model, [r for r, _ in sel], T, micro, device)
pairs, by_work = [], defaultdict(list)
for (r, w), pr in zip(sel, preds):
if pr is None:
continue
p = list(zip(r.y_scan.tolist(), pr[1].tolist()))
p = [(g, q) for g, q in p if g != -100]
pairs += p
by_work[w] += p
print(f"\n=== scan {split} (whole verses, by work) ===")
print(f" all: {json.dumps(scan_metrics(pairs))}")
for w in sorted(by_work):
print(f" {w}: {json.dumps(scan_metrics(by_work[w]))}")
def run_viterbi(model, sd, src, device, micro, theta, norma_source="hf"):
"""Meter-constrained decoding: exact-line + char accuracy, raw argmax vs
gated Viterbi (meter forced from the corpus label when a grammar exists,
else auto-detected; Norma syllabify is always auto)."""
import numpy as np
from meter.viterbi import METER_MAP, gated_auto, gated_decode
T = sd["T"]
def decode_set(name, items):
"""items: (group, meter_name|None, gold {ord: lab}, record)"""
recs = [r for _, _, _, r in items]
preds = predict_records(model, recs, T, micro, device)
agg = defaultdict(lambda: np.zeros(6, np.int64))
# [lines, exact_raw, exact_vit, char_ok_raw, char_ok_vit, chars] per group
applied = 0
for (group, mname, gold, rec), pr in zip(items, preds):
if pr is None:
continue
n = len(pr[3])
golds = np.zeros(n, np.int64)
for k, g in gold.items():
golds[k] = g
logp = np.log(np.clip(pr[3], 1e-9, 1.0))
raw = pr[1] # pre-computed argmax, with enforce_circumflex_heavy already applied
grammar = METER_MAP.get(mname or "")
if grammar:
vit, ok = gated_decode(logp.tolist(), grammar, theta)
else:
_, vit, ok = gated_auto(logp.tolist(), theta)
applied += ok
vit = np.asarray(vit)
for g in (group, "ALL"):
a = agg[g]
a[0] += 1
a[1] += int((raw == golds).all())
a[2] += int((vit == golds).all())
a[3] += int((raw == golds).sum())
a[4] += int((vit == golds).sum())
a[5] += n
print(f"\n=== viterbi {name} (theta={theta}, applied {applied}/"
f"{agg['ALL'][0]}) ===")
for g in sorted(agg, key=lambda x: (x != "ALL", x)):
a = agg[g]
print(f" {g}: lines={a[0]} exact raw={a[1]/a[0]:.3f} "
f"vit={a[2]/a[0]:.3f} | char raw={a[3]/a[5]:.4f} vit={a[4]/a[5]:.4f}")
# ---- work-split dev/test from the scanner corpus (grouped by grammar)
from meter.marks import parse_scan_line
for split, works in (("scan-dev", SCAN_DEV_WORKS), ("scan-test", SCAN_TEST_WORKS)):
items = []
for line in open(src / "data/scanner/corpus_v3.tsv", encoding="utf-8"):
parts = line.rstrip("\n").split("\t")
if len(parts) != 3 or parts[0] not in works:
continue
parsed = parse_scan_line(parts[2])
if parsed is None:
continue
rec = encode_plain(parsed[0])
if rec is None:
continue
grammar = METER_MAP.get(parts[1], f"auto({parts[1] or 'lyric'})")
items.append((grammar, parts[1], parsed[1], rec))
decode_set(split, items)
# ---- Norma syllabify (auto meter; test only -- Norma has no syllabify dev rows)
from meter.norma_data import load_norma
items = []
for d in load_norma(norma_source)["test"]:
if d["task"] != "syllabify":
continue
parsed = parse_scan_line(d["text"])
if parsed is None:
continue
rec = encode_plain(parsed[0])
if rec is None:
continue
items.append((d["source"], None, parsed[1], rec))
decode_set("norma-syllabify", items)
def run_file(model, sd, device, micro, mode, infile, outfile):
T = sd["T"]
raw = [l.rstrip("\n") for l in open(infile, encoding="utf-8")]
plains = [parse_macron_line(l)[0] for l in raw] # strips any existing marks
recs = [encode_plain(p) if p.strip() else None for p in plains]
preds = predict_records(model, recs, T, micro, device)
with open(outfile, "w", encoding="utf-8") as f:
for plain, rec, pr in zip(plains, recs, preds):
if rec is None or pr is None:
f.write(plain + "\n")
continue
if mode == "macronize":
amb = ambiguous_mask(rec.chars, rec.boundary, rec.dia)
labels = {int(i): int(pr[0][i]) for i in np.flatnonzero(amb)}
f.write(insert_marks(plain, labels) + "\n")
else:
labels = {i: int(c) for i, c in enumerate(pr[1]) if c > 0}
f.write(bracketize(plain, labels) + "\n")
print(f"{mode}: {len(raw)} lines -> {outfile}")
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--model", required=True)
ap.add_argument("--attn", default="sdpa")
ap.add_argument("--micro", type=int, default=16)
ap.add_argument("--norma", action="store_true")
ap.add_argument("--pred-out", default=None)
ap.add_argument("--scan-split", action="store_true")
ap.add_argument("--viterbi", action="store_true")
ap.add_argument("--theta", type=float, default=0.1)
ap.add_argument("--macronize", nargs=2, metavar=("IN", "OUT"))
ap.add_argument("--scan", nargs=2, metavar=("IN", "OUT"))
add_norma_source_arg(ap)
a = ap.parse_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model, sd = load_model(a.model, device, a.attn)
src = Path(os.path.expandvars(os.environ.get("MACRONIZER_SRC",
"$MACRONIZER_SRC")))
if a.norma:
pred_out = a.pred_out or (Path(sd["cfg"]["out_dir"]) / "norma_pred.jsonl")
run_norma(model, sd, device, a.micro, pred_out, a.norma_source)
if a.scan_split:
run_scan_split(model, sd, device, a.micro)
if a.viterbi:
run_viterbi(model, sd, src, device, a.micro, a.theta, a.norma_source)
if a.macronize:
run_file(model, sd, device, a.micro, "macronize", *a.macronize)
if a.scan:
run_file(model, sd, device, a.micro, "scan", *a.scan)
if __name__ == "__main__":
main()
|