batuhan28's picture
Upload v4/code/train.py with huggingface_hub
1e010ba verified
Raw History Blame Contribute Delete
33.2 kB
import os, io, re, json, time, math, pickle, sys, subprocess, shutil, datetime, glob, gc, uuid, warnings, logging
from pathlib import Path
from concurrent.futures import ThreadPoolExecutor
warnings.filterwarnings("ignore")
os.environ["HYDRA_FULL_ERROR"] = "0"
import numpy as np
import torch
import torch.nn.functional as F
import soundfile as sf
import sentencepiece as spm
import pyarrow.parquet as pq
from torch.utils.data import IterableDataset, DataLoader
from huggingface_hub import HfApi, HfFileSystem
import lightning.pytorch as pl
from lightning.pytorch.strategies import DDPStrategy
from lightning.pytorch.callbacks import ModelCheckpoint
from nemo.collections.asr.models import EncDecCTCModelBPE
from nemo.utils import logging as nemo_logging
nemo_logging.setLevel(logging.ERROR)
logging.getLogger("lightning.pytorch").setLevel(logging.ERROR)
sys.path.insert(0, str(Path(__file__).parent))
import lib
R = Path(__file__).parent
STATUS = R / "status.txt"
C = json.loads((R / "config.json").read_text())
rank = int(os.environ.get("RANK", 0))
world = int(os.environ.get("WORLD_SIZE", 1))
api = HfApi(token=os.environ["HF_TOKEN"])
fs = HfFileSystem(token=os.environ["HF_TOKEN"])
pl.seed_everything(42, workers=True)
# ==============================================================================
# ⚡ A100 EXTREME DONANIM BAYRAKLARI
# ==============================================================================
torch.set_num_threads(2)
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cudnn.benchmark = True
torch.set_float32_matmul_precision("high")
USE_BF16 = bool(C.get("bf16", True) and torch.cuda.is_bf16_supported())
AMP_DTYPE = torch.bfloat16 if USE_BF16 else torch.float16
PRECISION_STR = "bf16-mixed" if USE_BF16 else "16-mixed"
SOURCES = list(C["share"])
lang_of = lambda s: "en" if s == "librispeech" else "tr"
SCRATCH = Path(f"/tmp/shard_scratch_r{rank}")
SCRATCH.mkdir(parents=True, exist_ok=True)
def say(*a):
ts = time.strftime("%H:%M:%S")
msg = " ".join(str(x) for x in a)
print(f"[{ts} r{rank}] {msg}", flush=True)
if rank == 0:
try:
STATUS.write_text(f"[{ts}] {msg}", encoding="utf-8")
except Exception:
pass
def read(p, cols):
"""64MB Bloklarla Ultra-Hızlı İndir -> Çok Çekirdekli Arrow ile RAM'e Al -> Diskten Anında Sil."""
err = None
for k in range(4):
tmp_file = SCRATCH / f"{uuid.uuid4().hex}_{Path(p).name}"
try:
with fs.open(f"datasets/{C['data']}/{p}", "rb", block_size=64 * 1024 * 1024) as r_src, open(tmp_file, "wb") as w_dst:
shutil.copyfileobj(r_src, w_dst, length=64 * 1024 * 1024)
table = pq.read_table(tmp_file, columns=cols, use_threads=True)
out = table.to_pydict()
del table
return out
except Exception as e:
err = e
time.sleep(2 * (k + 1))
finally:
if tmp_file.exists():
try:
tmp_file.unlink()
except OSError:
pass
raise RuntimeError(f"Shard okunamadı: {p}") from err
def atomic_json(obj, path):
path = Path(path); tmp = path.with_suffix(".tmp")
tmp.write_text(json.dumps(obj, ensure_ascii=False)); tmp.replace(path)
def upload(path, name):
if C["upload"]:
try:
api.upload_file(path_or_fileobj=str(path), path_in_repo=f"v4/{name}", repo_id=C["repo"])
except Exception as e:
say("yükleme başarısız:", name, repr(e)[:150])
say(f"🚀 EXTREME A100 Süreci Başlatıldı (Batch={C['seconds']}sn | Clips={C['max_clips']} | Workers={C['workers']} | {PRECISION_STR})...")
if world > 1:
torch.cuda.set_device(int(os.environ["LOCAL_RANK"]))
torch.distributed.init_process_group("nccl", timeout=datetime.timedelta(hours=3))
def barrier():
if world > 1:
torch.distributed.barrier()
# ---------------------------------------------------------------- 1) shard listesi
if rank == 0 and not (R / "groups.json").exists():
say("📂 Adım 1/5: HuggingFace veri deposundaki Parquet shard listesi çekiliyor...")
files = api.list_repo_files(C["data"], repo_type="dataset")
allg = {s: sorted(p for p in files if p.startswith(f"data/{s}/") and p.endswith(".parquet")) for s in SOURCES}
assert all(allg.values()), "Bir eğitim kaynağı eksik: " + str([s for s in allg if not allg[s]])
atomic_json({"train": {s: v[1:] for s, v in allg.items()}, "dev": {s: v[0] for s, v in allg.items()}},
R / "groups.json")
say("✅ Shard listesi kaydedildi:", {s: len(v) for s, v in allg.items()})
barrier()
G = json.loads((R / "groups.json").read_text())
groups, devshard = G["train"], G["dev"]
# ---------------------------------------------------------------- 2) FLEURS dev/test
def load_fleurs(split):
dest = R / f"fleurs_{split}.pkl"
if dest.exists():
return pickle.load(dest.open("rb"))
say(f"🌍 Adım 2/5: Google FLEURS Türkçe '{split}' veri seti indiriliyor ve sesleri çözülüyor...")
from datasets import load_dataset, Audio
ds = None
for kwargs in (
dict(path="google/fleurs", name="tr_tr", split=split, trust_remote_code=True),
dict(path="google/fleurs", name="tr_tr", split=split),
dict(path="xtreme_s", name="fleurs.tr_tr", split=split),
):
try:
ds = load_dataset(**kwargs)
break
except Exception:
continue
if ds is None:
from huggingface_hub import hf_hub_download
import tarfile, csv
tsv_name = "dev.tsv" if split == "validation" else f"{split}.tsv"
tar_name = "dev.tar.gz" if split == "validation" else f"{split}.tar.gz"
tsv_p = hf_hub_download("google/fleurs", f"data/tr_tr/{tsv_name}", repo_type="dataset")
tar_p = hf_hub_download("google/fleurs", f"data/tr_tr/audio/{tar_name}", repo_type="dataset")
trans_map = {}
with open(tsv_p, "r", encoding="utf-8") as f:
for row in csv.reader(f, delimiter="\t"):
if len(row) >= 3:
trans_map[row[1].strip()] = row[2].strip()
out = []
with tarfile.open(tar_p, "r:gz") as tar:
members = [m for m in tar.getmembers() if m.isfile() and m.name.endswith(".wav")]
for idx, m in enumerate(members, 1):
fname = Path(m.name).name
if fname in trans_map:
raw = tar.extractfile(m).read()
out.append((lib.decode_audio(raw), trans_map[fname]))
if idx % 200 == 0 or idx == len(members):
say(f" -> FLEURS {split} (doğrudan tar.gz) ses çözümü: {idx}/{len(members)}")
with dest.open("wb") as f:
pickle.dump(out, f)
return out
ds = ds.cast_column("audio", Audio(decode=False))
out = []
tot = len(ds)
for idx, ex in enumerate(ds, 1):
a = ex["audio"]
raw = a.get("bytes") or Path(a["path"]).read_bytes()
out.append((lib.decode_audio(raw), ex["transcription"]))
if idx % 200 == 0 or idx == tot:
say(f" -> FLEURS {split} ses çözümü: {idx}/{tot}")
with dest.open("wb") as f:
pickle.dump(out, f)
shutil.rmtree(Path.home() / ".cache/huggingface/datasets", ignore_errors=True)
return out
if rank == 0:
load_fleurs("validation")
atomic_json([t for _, t in load_fleurs("test")], R / "fleurs_test_refs.json")
barrier()
rawdev = load_fleurs("validation")
rawtest_refs = json.loads((R / "fleurs_test_refs.json").read_text())
ref_clean = lambda t: lib.clean_text(t, "tr", strict=False)
blocked = {ref_clean(t) for _, t in rawdev} | {ref_clean(t) for t in rawtest_refs}
del rawtest_refs
# ---------------------------------------------------------------- 3) tokenizer (BPE) + başlangıç modeli
if rank == 0 and not (R / "init.nemo").exists():
tokdir = R / "tokenizer"; tokdir.mkdir(exist_ok=True)
rng = np.random.default_rng(0)
per_source = {s: [] for s in SOURCES}
jobs = []
for s in SOURCES:
paths = groups[s]
pick = [paths[i] for i in np.linspace(0, len(paths) - 1, min(8, len(paths))).astype(int)]
for p in pick:
jobs.append((s, p))
say(f"🔤 Adım 3/5: Tokenizer eğitimi için {len(jobs)} shard'dan 12 paralel kolla metin örnekleniyor...")
def fetch_tok_texts(item):
s, p = item
d = read(p, ["text"])
out = [t for t in (lib.clean_text(x, lang_of(s)) for x in d["text"]) if t and t not in blocked]
return s, p, out
done_cnt = 0
with ThreadPoolExecutor(12) as ex:
for s, p, out in ex.map(fetch_tok_texts, jobs):
per_source[s].extend(out)
done_cnt += 1
if done_cnt % 4 == 0 or done_cnt == len(jobs):
say(f" -> Tokenizer metin indirme ilerlemesi: [{done_cnt}/{len(jobs)}] shard tamamlandı ({s})")
lines = []
for s in SOURCES:
n = min(len(per_source[s]), int(C["share"][s] * 2_000_000))
lines += list(rng.choice(per_source[s], size=n, replace=False)) if n < len(per_source[s]) else per_source[s]
assert lines, "Tokenizer metni yok."
say(f"🔤 SentencePiece BPE ({C['vocab']} token) toplam {len(lines):,} cümle üzerinde eğitiliyor...")
(tokdir / "text.txt").write_text("\n".join(lines), encoding="utf-8")
del per_source, lines; gc.collect()
spm.SentencePieceTrainer.train(
input=str(tokdir / "text.txt"), model_prefix=str(tokdir / "tokenizer"), model_type="bpe",
vocab_size=C["vocab"], character_coverage=1.0, normalization_rule_name="identity",
bos_id=-1, eos_id=-1, pad_id=-1, hard_vocab_limit=False, num_threads=C["cpus"],
required_chars="abcçdefgğhıijklmnoöprsştuüvyzqwx", input_sentence_size=2_000_000,
shuffle_input_sentence=True, minloglevel=2)
(tokdir / "text.txt").unlink(missing_ok=True)
sp0 = spm.SentencePieceProcessor(model_file=str(tokdir / "tokenizer.model"))
(tokdir / "vocab.txt").write_text("\n".join(sp0.id_to_piece(i) for i in range(sp0.vocab_size())))
say(f"📦 Taban model ({C['base']}) indiriliyor ve Türkçe BPE sözlüğü enjekte ediliyor...")
m = EncDecCTCModelBPE.from_pretrained(C["base"], map_location="cpu")
m.change_vocabulary(new_tokenizer_dir=str(tokdir), new_tokenizer_type="bpe")
m.save_to(str(R / "init.nemo")); del m
say("☁️ init.nemo yedekleniyor...")
upload(R / "init.nemo", "init.nemo")
barrier()
# ---------------------------------------------------------------- 4) model (2400s Dev Batch + OOM Korumalı Forward + FusedAdamW)
say("🧠 Adım 4/5: init.nemo belleğe yükleniyor...")
class ASR(EncDecCTCModelBPE):
group_names = ["fleurs"] + SOURCES + ["fleurs_rest"]
def configure_optimizers(self):
if USE_BF16:
opt = torch.optim.AdamW(self.parameters(), lr=C["lr"], betas=(0.9, 0.98), weight_decay=C["wd"], fused=True)
else:
opt = torch.optim.AdamW(self.parameters(), lr=C["lr"], betas=(0.9, 0.98), weight_decay=C["wd"], foreach=True)
total, warm = C["steps"], C["warmup"]
decay = int(total * 0.25)
def f(s):
if s < warm:
return (s + 1) / warm
if s < total - decay:
return 1.0
return max(0.05, 1.0 - 0.95 * (s - (total - decay)) / decay)
return {"optimizer": opt, "lr_scheduler": {"scheduler": torch.optim.lr_scheduler.LambdaLR(opt, f),
"interval": "step"}}
def on_before_optimizer_step(self, optimizer):
torch.nn.utils.clip_grad_norm_(self.parameters(), max_norm=1.0)
def _chunk_loss(self, x, xl, y, yl):
proc_x, proc_xl = self.preprocessor(input_signal=x, length=xl)
if self.spec_augmentation is not None and self.training:
proc_x = self.spec_augmentation(input_spec=proc_x, length=proc_xl)
encoded, el = self.encoder(audio_signal=proc_x, length=proc_xl)
lp = self.decoder(encoder_output=encoded)
loss = F.ctc_loss(lp.float().transpose(0, 1), y, el.long(), yl, blank=lp.shape[-1] - 1,
reduction="none", zero_infinity=True)
return loss.sum()
def training_step(self, batch, batch_idx):
x, xl, y, yl = batch
tot_tokens = yl.sum().clamp(min=1)
try:
return self._chunk_loss(x, xl, y, yl) / tot_tokens
except torch.cuda.OutOfMemoryError:
# Eğer 2400 sn'lik ekstrem bir batch 40GB sınırını aşarsa çökmez, ikiye bölüp işler!
torch.cuda.empty_cache()
mid = max(1, x.shape[0] // 2)
l1 = self._chunk_loss(x[:mid], xl[:mid], y[:mid], yl[:mid])
l2 = self._chunk_loss(x[mid:], xl[mid:], y[mid:], yl[mid:])
return (l1 + l2) / tot_tokens
def on_validation_epoch_start(self):
self._acc = torch.zeros(len(self.group_names), 6, device=self.device)
self.full_pass = self.global_step % C["full_every"] == 0
def validation_step(self, batch, batch_idx, dataloader_idx=0):
if dataloader_idx == 2 and not self.full_pass:
return
x, xl, y, yl, g, refs = batch
lp, el, _ = self.forward(input_signal=x, input_signal_length=xl)
lpf = lp.float(); V = lpf.shape[-1] - 1
nll = F.ctc_loss(lpf.transpose(0, 1), y, el.long(), yl, blank=V, reduction="none", zero_infinity=True)
pred = lpf.argmax(-1).cpu(); elc = el.cpu(); gl = g.tolist()
for i in range(len(refs)):
z = pred[i, : int(elc[i])].unique_consecutive().tolist()
hyp = self.tokenizer.ids_to_text([v for v in z if v != V])
we, wn, ce, cn = lib.wer_counts([refs[i]], [hyp])
self._acc[gl[i]] += torch.tensor([we, wn, ce, cn, float(nll[i]), float(yl[i])], device=self.device)
def on_validation_epoch_end(self):
acc = self._acc
if world > 1:
torch.distributed.all_reduce(acc)
rep = {}
for k, name in enumerate(self.group_names):
we, wn, ce, cn, nl, nt = acc[k].tolist()
if wn > 0:
rep[name] = dict(wer=100 * we / wn, cer=100 * ce / max(cn, 1), nll=nl / max(nt, 1))
if self.full_pass and "fleurs" in rep and "fleurs_rest" in rep:
a, b = acc[0], acc[len(self.group_names) - 1]
s_ = (a + b).tolist()
rep["fleurs_full"] = dict(wer=100 * s_[0] / max(s_[1], 1), cer=100 * s_[2] / max(s_[3], 1),
nll=s_[4] / max(s_[5], 1))
self.domain_report = rep
f = rep.get("fleurs", dict(wer=100.0, cer=100.0, nll=99.0))
for key in ("wer", "cer", "nll"):
self.log(f"dev_{key}", float(f[key]), sync_dist=(world > 1), rank_zero_only=False, batch_size=1)
model = ASR.restore_from(str(R / "init.nemo"), map_location="cpu")
tok = model.tokenizer
if hasattr(model.encoder, "sync_max_audio_length"):
model.encoder.sync_max_audio_length = False
V = tok.vocab_size
say("✅ Model yüklendi -> Vokabüler:", V, "| Parametre (M):", round(sum(p.numel() for p in model.parameters()) / 1e6, 1))
# ---------------------------------------------------------------- 5) veri
def pack(items):
ws, ids = zip(*items)
x = torch.zeros(len(items), max(map(len, ws)))
y = torch.zeros(len(items), max(map(len, ids)), dtype=torch.long)
for i, (w, t) in enumerate(items):
x[i, : len(w)] = torch.from_numpy(w.astype("float32") / 32767.0) if w.dtype == np.int16 else torch.from_numpy(w)
y[i, : len(t)] = torch.tensor(t)
return x, torch.tensor(list(map(len, ws))), y, torch.tensor(list(map(len, ids)))
class Stream(IterableDataset):
def __iter__(self):
wi = torch.utils.data.get_worker_info()
wid, nw = (wi.id, wi.num_workers) if wi else (0, 1)
worker = rank * nw + wid
rng = np.random.default_rng(42 + 1000 * C["launch"] + worker)
cols = ["audio", "text", "duration"]
for b in lib.batch_stream(groups, C["share"], C["hours"], worker, world * nw, rng,
lambda p: read(p, cols), tok.text_to_ids, blocked, C["seconds"], C["max_clips"],
lang_of=lang_of, tel_p=C["tel_p"], edge_p=C["edge_p"],
min_buffer=C.get("min_buffer", 3200),
ram_cache_shards=C.get("ram_cache_shards", 4),
status_fn=say if worker == 0 else None):
yield pack(b)
def dev_items(raw, g):
out = []
for w, t in raw:
ref = ref_clean(t)
ids = tok.text_to_ids(ref)
if ids:
out.append((w, ids, ref, g))
out.sort(key=lambda z: len(z[0]))
return out
def pack_dev(items):
x, xl, y, yl = pack([(w, ids) for w, ids, _, _ in items])
return x, xl, y, yl, torch.tensor([g for *_, g in items]), [r for _, _, r, _ in items]
if rank == 0 and not (R / "domain_dev.pkl").exists():
out = []
for idx_s, s in enumerate(SOURCES, 1):
say(f"📊 Adım 5/5: Kaynak bazlı doğrulama kümesi hazırlanıyor [{idx_s}/{len(SOURCES)}]: {s}...")
d = read(devshard[s], ["audio", "text", "duration"])
rows = []
for a, t, dur in zip(d["audio"], d["text"], d["duration"]):
text = lib.clean_text(t, lang_of(s))
if a and a.get("bytes") and dur and 1 <= dur <= 20 and text and lib.label_ok(text, dur) and text not in blocked:
rows.append((a["bytes"], text))
pick = np.random.default_rng(0).permutation(len(rows))[: C["dev_per_source"]]
out += [(lib.decode_audio(rows[i][0]), rows[i][1], s) for i in pick]
del d, rows; gc.collect()
with (R / "domain_dev.pkl").open("wb") as f:
pickle.dump(out, f)
barrier()
dom_raw = pickle.load((R / "domain_dev.pkl").open("rb"))
dom = []
for w, text, s in dom_raw:
ids = tok.text_to_ids(text)
if ids:
dom.append((w, ids, text, 1 + SOURCES.index(s)))
dom.sort(key=lambda z: len(z[0]))
del dom_raw
train = DataLoader(Stream(), batch_size=None, num_workers=C["workers"], prefetch_factor=C.get("prefetch", 6),
pin_memory=True, persistent_workers=(C["workers"] > 0))
fl_all = dev_items(rawdev, 0)
sel = set(lib.pick_representative([len(w) / 16000 for w, *_ in fl_all], [len(r.split()) for _, _, r, _ in fl_all],
[len(r) for _, _, r, _ in fl_all], C["dev_subset"]))
fl_sub = [it for i, it in enumerate(fl_all) if i in sel]
fl_rest = [(w, ids, r, 1 + len(SOURCES)) for i, (w, ids, r, _) in enumerate(fl_all) if i not in sel]
say("🎯 Hazırlık tamamlandı! FLEURS dev:", len(fl_all), "| Alt küme:", len(fl_sub), "-> A100 Extreme Eğitimi başlıyor...")
dev = [DataLoader(fl_sub, batch_size=64, collate_fn=pack_dev, pin_memory=True),
DataLoader(dom, batch_size=64, collate_fn=pack_dev, pin_memory=True),
DataLoader(fl_rest, batch_size=64, collate_fn=pack_dev, pin_memory=True)]
# ---------------------------------------------------------------- 6) callback'ler
class Report(pl.Callback):
def on_train_start(self, t, m):
self.f = (R / "metrics.jsonl").open("a", buffering=1) if t.is_global_zero else None
if t.is_global_zero:
say(f"🔥 A100 EXTREME ({PRECISION_STR} | Batch={C['seconds']}sn) eğitimi aktif!")
def on_train_batch_end(self, t, m, outputs, batch, batch_idx):
if self.f and (t.global_step == 1 or t.global_step % 5 == 0):
loss = outputs["loss"] if isinstance(outputs, dict) else outputs
lr = t.optimizers[0].param_groups[0]["lr"]
self.f.write(json.dumps(dict(step=t.global_step, loss=float(loss.detach()), lr=lr, time=time.time())) + "\n")
if t.global_step <= 15 or t.global_step % 25 == 0:
say(f"⚡ Adım {t.global_step:,}/{C['steps']:,} | Loss: {float(loss.detach()):.3f} | LR: {lr:.2e} | Klip/Batch: {batch[0].shape[0]}")
def on_validation_end(self, t, m):
if t.sanity_checking or not t.is_global_zero or not self.f:
return
rep = getattr(m, "domain_report", {})
d = dict(step=t.global_step, time=time.time(), domains=rep)
d.update({k: float(v) for k, v in t.callback_metrics.items() if k in ("dev_wer", "dev_cer", "dev_nll")})
self.f.write(json.dumps(d) + "\n")
say("DEV", {k: (round(v, 2) if isinstance(v, float) else v) for k, v in d.items() if k != "domains"},
{k: round(v["wer"], 1) for k, v in rep.items()})
def on_train_end(self, t, m):
if self.f:
self.f.close()
class Backup(pl.Callback):
fut = None
def on_train_batch_end(self, t, m, *a):
s = t.global_step
if not s or s % C["save"]:
return
if self.fut is not None:
self.fut.result(); self.fut = None
tmp = R / "last.tmp.ckpt"
t.save_checkpoint(str(tmp))
if t.is_global_zero:
tmp.replace(R / "last.ckpt")
if C["upload"]:
try:
self.fut = api.upload_file(path_or_fileobj=str(R / "last.ckpt"), path_in_repo="v4/last.ckpt",
repo_id=C["repo"], run_as_future=True)
except Exception as e:
say("yedek yüklenemedi", repr(e)[:120])
def on_train_end(self, t, m):
if self.fut is not None:
self.fut.result()
best = ModelCheckpoint(dirpath=str(R / "best"), filename="best-{step}-{dev_wer:.3f}", monitor="dev_wer", mode="min",
save_top_k=C["topk"], save_weights_only=True, enable_version_counter=False)
strategy = DDPStrategy(find_unused_parameters=False) if world > 1 else "auto"
trainer = pl.Trainer(
accelerator="gpu", devices=world, strategy=strategy,
precision=PRECISION_STR, max_steps=C["steps"], max_epochs=-1, val_check_interval=C["val"],
check_val_every_n_epoch=None, num_sanity_val_steps=0, gradient_clip_val=None, logger=False,
enable_progress_bar=False, callbacks=[Report(), best, Backup()])
model.set_trainer(trainer)
# Eğer önceki küçük batch'ten kalan last.ckpt varsa optimizer/LR zamanlayıcısı uyuşmazlığı olmaması için
# sadece ağırlıklarını yükleyip yeni hızlı 8500 adımlık takvimle devam etmesini sağlıyoruz:
last = R / "last.ckpt"
ckpt_arg = None
if last.exists():
try:
ck = torch.load(str(last), map_location="cpu", weights_only=False)
prev_step = int(ck.get("global_step", 0))
if prev_step > 0 and prev_step < C["steps"]:
# Aynı takvimdeyse doğrudan devam et, eski 30k takvimindeyse ağırlıkları al
model.load_state_dict(ck["state_dict"], strict=True)
say(f"♻️ Önceki last.ckpt ağırlıkları yüklendi (Eski adım: {prev_step}).")
del ck; gc.collect()
except Exception as e:
say("last.ckpt okuma bilgisi:", repr(e)[:120])
trainer.fit(model, train_dataloaders=train, val_dataloaders=dev, ckpt_path=ckpt_arg)
barrier()
shutil.rmtree(SCRATCH, ignore_errors=True)
if world > 1:
torch.distributed.destroy_process_group()
if rank != 0:
sys.exit(0)
# =============================================================== yalnız rank 0: Greedy Soup, test, LM, ONNX
SCRATCH.mkdir(parents=True, exist_ok=True)
dev0 = torch.device("cuda:0")
def infer(m, raw, bs=64):
m = m.to(dev0).eval()
order = sorted(range(len(raw)), key=lambda i: len(raw[i][0]))
out = [None] * len(raw)
with torch.inference_mode(), torch.autocast("cuda", dtype=AMP_DTYPE):
for k in range(0, len(order), bs):
idx = order[k: k + bs]
ws = [raw[i][0] for i in idx]
x = torch.zeros(len(ws), max(map(len, ws)))
for j, w in enumerate(ws):
x[j, : len(w)] = torch.from_numpy(w)
lp, el, _ = m.forward(input_signal=x.to(dev0, non_blocking=True),
input_signal_length=torch.tensor([len(w) for w in ws], device=dev0))
lp = lp.float().cpu().numpy()
for j, i in enumerate(idx):
out[i] = (lp[j, : int(el[j])].copy(), ref_clean(raw[i][1]))
return out
def greedy(lps):
V_ = lps[0][0].shape[1] - 1
refs, hyps = [], []
for lp, ref in lps:
a = lp.argmax(1)
ids = [int(v) for k, v in enumerate(a) if v != V_ and (k == 0 or v != a[k - 1])]
refs.append(ref); hyps.append(tok.ids_to_text(ids))
we, wn, ce, cn = lib.wer_counts(refs, hyps)
return 100 * we / wn, 100 * ce / cn
ckpts = sorted(glob.glob(str(R / "best" / "*.ckpt")))
say("🥣 Greedy Soup değerlendirmesi başlıyor, checkpoint sayısı:", len(ckpts))
sd_list = [torch.load(p, map_location="cpu", weights_only=False)["state_dict"] for p in ckpts]
final_sd = {k: v.detach().cpu().clone() for k, v in model.state_dict().items()}
cands = {"last": final_sd}
scores = {}
model.load_state_dict(final_sd, strict=True)
scores["last"] = greedy(infer(model, rawdev))[0]
say("dev WER last:", round(scores["last"], 2))
if sd_list:
wers = [float(re.search(r"dev_wer=(\d+\.\d+)", p).group(1)) if re.search(r"dev_wer=(\d+\.\d+)", p) else 1e9 for p in ckpts]
order = np.argsort(wers)
best_sd = sd_list[order[0]]
cands["best_single"] = best_sd
model.load_state_dict(best_sd, strict=True)
scores["best_single"] = greedy(infer(model, rawdev))[0]
say("dev WER best_single:", round(scores["best_single"], 2))
soup_pool = [best_sd]
best_soup_wer = scores["best_single"]
soup_sd = best_sd
for idx in order[1:]:
trial_pool = soup_pool + [sd_list[idx]]
trial_sd = {
k: (sum(sd[k].float() for sd in trial_pool) / len(trial_pool)).to(best_sd[k].dtype)
if best_sd[k].is_floating_point() else best_sd[k]
for k in best_sd
}
model.load_state_dict(trial_sd, strict=True)
w_score = greedy(infer(model, rawdev))[0]
if w_score < best_soup_wer:
soup_pool = trial_pool
best_soup_wer = w_score
soup_sd = trial_sd
say(f"greedy soup eklendi ({len(soup_pool)} model) -> dev WER: {w_score:.2f}")
cands["soup"] = soup_sd
scores["soup"] = best_soup_wer
final_name = min(scores, key=scores.get)
model.load_state_dict(cands[final_name], strict=True)
say("🏆 Seçilen final model:", final_name, "| WER:", round(scores[final_name], 2))
model.save_to(str(R / "turkish_ctc.nemo"))
upload(R / "turkish_ctc.nemo", "turkish_ctc.nemo")
del sd_list, cands; gc.collect()
rawtest = load_fleurs("test")
lp_dev, lp_test = infer(model, rawdev), infer(model, rawtest)
res = dict(selected=final_name, dev_scores=scores,
dev_greedy_wer_cer=greedy(lp_dev), test_greedy_wer_cer=greedy(lp_test))
atomic_json(res, R / "results.json")
say("FINAL FLEURS (greedy)", res)
sp = tok.tokenizer
(R / "tokenizer").mkdir(exist_ok=True)
(R / "tokenizer" / "tokenizer.model").write_bytes(sp.serialized_model_proto())
def lse_norm(lp, T):
z = lp / T
z = z - z.max(1, keepdims=True)
return (z - np.log(np.exp(z).sum(1, keepdims=True))).astype("float32")
try:
bindir = lib.ensure_kenlm(say)
assert bindir, "KenLM ikilileri (lmplz) derlenemedi"
corpus = Path("/tmp/lm_corpus.txt")
extra = [Path(p) for p in C.get("extra_text", []) if Path(p).exists()]
def lm_texts(args):
s, p = args
d = read(p, ["text", "duration"])
out = []
for t, dur in zip(d["text"], d["duration"]):
c = lib.clean_text(t, lang_of(s))
if c and c not in blocked and dur and lib.label_ok(c, dur):
out.append(lib.lm_line(sp, c))
del d
return out
jobs = [(s, p) for s in SOURCES for p in groups[s]]
say(f"📚 KenLM 6-gram dil modeli için {len(jobs)} shard 12 kolla paralel taranıyor...")
n_lines, j_done = 0, 0
with corpus.open("w", encoding="utf-8") as fo, ThreadPoolExecutor(12) as ex:
for ch in ex.map(lm_texts, jobs):
j_done += 1
if ch:
fo.write("\n".join(ch) + "\n")
n_lines += len(ch)
if j_done % 25 == 0 or j_done == len(jobs):
say(f" -> LM metin tarama: [{j_done}/{len(jobs)}] shard ({n_lines:,} satır)")
for p in extra:
for line in p.read_text(encoding="utf-8", errors="ignore").splitlines():
c = lib.clean_text(line, "tr")
if c and c not in blocked:
fo.write(lib.lm_line(sp, c) + "\n"); n_lines += 1
say("🔨 KenLM binary (24G RAM ile) derleniyor, toplam satır:", n_lines)
lib.build_kenlm(corpus, R / "lm6.bin", bindir, order=C["lm_order"], mem=C["lm_mem"])
corpus.unlink(missing_ok=True)
upload(R / "lm6.bin", "lm6.bin")
dec = lib.LMDecoder(str(R / "lm6.bin"), V, alpha=0.5, beta=0.0, beam=C["beam"])
say("🔍 LM Decoder motoru:", dec.backend)
def lm_wer(lps, alpha, beta, T):
dec.set(alpha=alpha, beta=beta)
refs, hyps = [], []
for lp, ref in lps:
hyps.append(tok.ids_to_text(dec.decode(lse_norm(lp, T)))); refs.append(ref)
we, wn, ce, cn = lib.wer_counts(refs, hyps)
return 100 * we / wn, 100 * ce / cn
grid = {}
for alpha in (0.35, 0.55, 0.75):
for beta in (-0.2, 0.0, 0.3):
for T in (1.0, 1.25):
key = f"{alpha}_{beta}_{T}"
grid[key] = lm_wer(lp_dev, alpha, beta, T)[0]
say("LM dev", key, round(grid[key], 2))
bk = min(grid, key=grid.get)
ba, bb, bt = map(float, bk.split("_"))
res.update(lm_backend=dec.backend, lm_grid=grid, lm_alpha=ba, lm_beta=bb, lm_T=bt,
dev_lm_wer=grid[bk], test_lm_wer_cer=lm_wer(lp_test, ba, bb, bt))
atomic_json(res, R / "results.json")
say("FINAL FLEURS (LM)", res)
except Exception as e:
res["lm_error"] = repr(e)[:300]; atomic_json(res, R / "results.json"); say("LM aşaması atlandı:", repr(e)[:300])
upload(R / "results.json", "results.json")
shutil.rmtree(SCRATCH, ignore_errors=True)
try:
say("📦 ONNX (FP32 + INT8) CPU paketi dışa aktarılıyor...")
import onnxruntime as ort
from onnxruntime.quantization import quantize_dynamic, QuantType
cpu = model.cpu().eval()
fe = cpu.preprocessor.featurizer
fb = fe.fb[0].detach().numpy().astype("float32"); win = fe.window.detach().numpy().astype("float32")
kw = dict(n_fft=int(fe.n_fft), hop=int(fe.hop_length), preemph=float(getattr(fe, "preemph", 0.97) or 0))
onnx_fp, onnx_q = R / "turkish_ctc.onnx", R / "turkish_ctc.int8.onnx"
cpu.export(str(onnx_fp))
quantize_dynamic(str(onnx_fp), str(onnx_q), weight_type=QuantType.QInt8)
np.savez(R / "frontend.npz", fb=fb, window=win, **{k: np.array(v) for k, v in kw.items()})
w0 = rawdev[0][0]
with torch.no_grad():
ref_f, ref_l = cpu.preprocessor(input_signal=torch.from_numpy(w0)[None], length=torch.tensor([len(w0)]))
ref_f = ref_f[0, :, : int(ref_l[0])].numpy()
parity = {}
for pm in ("constant", "reflect"):
f = lib.featurize_np(w0, fb, win, pad_mode=pm, **kw)
parity[pm] = float(np.abs(f[:, : ref_f.shape[1]] - ref_f[:, : f.shape[1]]).max()) if f.shape[0] == ref_f.shape[0] else 1e9
pad_mode = min(parity, key=parity.get)
info = dict(feature_parity_max_abs=parity, pad_mode=pad_mode)
def run(path, threads, clips):
so = ort.SessionOptions(); so.intra_op_num_threads = threads
sess = ort.InferenceSession(str(path), so, providers=["CPUExecutionProvider"])
in_names = [i.name for i in sess.get_inputs()]
refs, hyps, dur, t_used = [], [], 0.0, 0.0
for w, ref in clips:
f = lib.featurize_np(w, fb, win, pad_mode=pad_mode, **kw)
t0 = time.time()
feed = {in_names[0]: f[None]}
if len(in_names) > 1:
feed[in_names[1]] = np.array([f.shape[1]], dtype="int64")
o = sess.run(None, feed)[0][0]
t_used += time.time() - t0 + 0.0
a = o.argmax(1)
ids = [int(v) for k, v in enumerate(a) if v != V and (k == 0 or v != a[k - 1])]
refs.append(ref_clean(ref)); hyps.append(tok.ids_to_text(ids)); dur += len(w) / 16000
we, wn, _, _ = lib.wer_counts(refs, hyps)
return dict(wer=100 * we / wn, rtfx_model_only=dur / max(t_used, 1e-9))
clips = rawdev[:60]
for name, p in (("fp32", onnx_fp), ("int8", onnx_q)):
info[name] = dict(size_mb=round(p.stat().st_size / 2 ** 20, 1),
threads1=run(p, 1, clips), threads4=run(p, 4, clips))
res["cpu_export"] = info; atomic_json(res, R / "results.json")
say("CPU EXPORT", info)
for p in (onnx_fp, onnx_q, R / "frontend.npz", R / "tokenizer" / "tokenizer.model"):
upload(p, p.name)
upload(R / "results.json", "results.json")
except Exception as e:
res["cpu_export_error"] = repr(e)[:300]; atomic_json(res, R / "results.json"); say("ONNX aşaması atlandı:", repr(e)[:300])
say("🏁 TÜM AŞAMALAR BİTTİ")