prosperious's picture
Add 1000 files (ocr-tech/tests/test_audit_limbu_book_missed_line_review_coverage.py ...)
d82db76 verified
Raw History Blame Contribute Delete
15.9 kB
"""Fine-tune the official devanagari_PP-OCRv5_mobile_rec with PyTorch (CUDA on the T4, MPS on k2).
Paddle's CPU training on macOS arm64 runs at ~2 images/s (the PPLCNetV3 depthwise convs have
no fast CPU kernel there), so training uses the Hugging Face torch port of the same network,
initialised from PaddlePaddle/devanagari_PP-OCRv5_mobile_rec_safetensors, with CTC loss.
Sources are ``name:root:labels:weight:aug`` where ``weight`` is the per-epoch sampling
multiplier (0.5 = a random half each epoch, 2 = every line twice) and ``aug`` the
probability of photometric/geometric augmentation. Batches are bucketed by resized width
and padded exactly like PaddleX inference, so there is no train/serve squeeze mismatch.
Model selection uses ``--dev`` sets only; test sets are never read here.
``--guard-dev NAME`` turns a dev set into a no-regression guard: a checkpoint is eligible as
"best" only while every guard set's codepoint CER stays within ``--guard-tol`` (absolute) of
its step-0 value, and eligible checkpoints are ranked by the mean CER of the non-guard sets.
Without guards, a 5-point handwriting gain outweighs a 1-point print loss in the plain mean.
"""
from __future__ import annotations
import argparse
import json
import math
import random
import sys
import time
import unicodedata
from pathlib import Path
import cv2
import numpy as np
import torch
from PIL import Image
from torch.utils.data import DataLoader, Dataset, Sampler
from common import (
IMG_H,
MIN_W,
DEVICE_CHOICES,
batch_tensor,
ctc_greedy_decode,
ctc_loss,
encode_label,
load_character_list,
load_model,
logits,
pick_device,
read_bgr,
resized_width,
)
from eval_metrics import evaluate
from selection import guard_regressions, split_dev_sets, target_score
MAX_TRAIN_W = 960
def augment(image: np.ndarray, rng: random.Random) -> np.ndarray:
"""Conservative capture degradations that never crop glyphs."""
img = image
if rng.random() < 0.3: # extra margin, replicated border
top, bottom = rng.randint(0, 4), rng.randint(0, 4)
left, right = rng.randint(0, 8), rng.randint(0, 8)
img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_REPLICATE)
if rng.random() < 0.25:
h, w = img.shape[:2]
m = cv2.getRotationMatrix2D((w / 2, h / 2), rng.uniform(-2.0, 2.0), 1.0)
img = cv2.warpAffine(img, m, (w, h), borderMode=cv2.BORDER_REPLICATE)
if rng.random() < 0.15:
kernel = np.ones((2, 2), np.uint8)
img = cv2.erode(img, kernel) if rng.random() < 0.5 else cv2.dilate(img, kernel)
if rng.random() < 0.3: # low-resolution capture
h, w = img.shape[:2]
scale = rng.uniform(0.4, 0.8)
small = cv2.resize(img, (max(4, int(w * scale)), max(4, int(h * scale))), interpolation=cv2.INTER_AREA)
img = cv2.resize(small, (w, h), interpolation=cv2.INTER_LINEAR)
if rng.random() < 0.25:
img = cv2.GaussianBlur(img, (3, 3) if rng.random() < 0.7 else (5, 5), 0)
if rng.random() < 0.4:
alpha, beta = rng.uniform(0.6, 1.3), rng.uniform(-30, 30)
img = cv2.convertScaleAbs(img, alpha=alpha, beta=beta)
if rng.random() < 0.2:
tint = np.array([rng.uniform(0.8, 1.0) for _ in range(3)], dtype=np.float32)
img = np.clip(img.astype(np.float32) * tint, 0, 255).astype(np.uint8)
if rng.random() < 0.3:
noise = np.random.default_rng(rng.randint(0, 2**31)).normal(0, rng.uniform(2, 10), img.shape)
img = np.clip(img.astype(np.float32) + noise, 0, 255).astype(np.uint8)
if rng.random() < 0.3:
ok, buf = cv2.imencode(".jpg", img, [cv2.IMWRITE_JPEG_QUALITY, rng.randint(30, 85)])
decoded = cv2.imdecode(buf, cv2.IMREAD_COLOR) if ok else None
if decoded is not None:
img = decoded
return img
class LineSet(Dataset):
def __init__(self, items: list[dict], train: bool, seed: int) -> None:
self.items = items
self.train = train
self.seed = seed
def __len__(self) -> int:
return len(self.items)
def __getitem__(self, index: int):
item = self.items[index]
image = read_bgr(Path(item["path"]))
if self.train and item["aug"] > 0:
rng = random.Random((self.seed * 1_000_003 + index * 7919 + int(time.time_ns() % 1_000_000)) % (2**31))
if rng.random() < item["aug"]:
image = augment(image, rng)
width = min(resized_width(image), MAX_TRAIN_W if self.train else 3200)
return image, width, item["ids"], item["text"]
def collate(batch):
images, widths, ids, texts = zip(*batch)
return batch_tensor(list(images), list(widths)), list(ids), list(texts)
class BucketSampler(Sampler[list[int]]):
"""Width-bucketed batches; a fresh weighted epoch sample every iteration."""
def __init__(self, groups: list[tuple[list[int], float]], widths: list[int], batch_size: int, seed: int) -> None:
self.groups, self.widths, self.batch_size, self.seed = groups, widths, batch_size, seed
self.epoch = 0
def epoch_indices(self) -> list[int]:
rng = random.Random(self.seed + self.epoch)
out: list[int] = []
for indices, weight in self.groups:
whole, frac = int(weight), weight - int(weight)
out += indices * whole
out += rng.sample(indices, int(round(frac * len(indices))))
rng.shuffle(out)
return out
def __iter__(self):
rng = random.Random(self.seed * 31 + self.epoch)
order = self.epoch_indices()
window = self.batch_size * 64
batches: list[list[int]] = []
for start in range(0, len(order), window):
chunk = sorted(order[start:start + window], key=lambda i: self.widths[i])
batches += [chunk[i:i + self.batch_size] for i in range(0, len(chunk), self.batch_size)]
rng.shuffle(batches)
self.epoch += 1
return iter(b for b in batches if len(b) == self.batch_size)
def __len__(self) -> int:
return len(self.epoch_indices()) // self.batch_size
def load_source(spec: str, char_to_index: dict[str, int], max_len: int, stats: dict) -> tuple[list[dict], float]:
name, root, labels, weight, aug = spec.split(":")
root_path = Path(root)
items = []
for line in Path(labels).read_text(encoding="utf-8").splitlines():
if "\t" not in line:
continue
rel, text = line.split("\t", 1)
text = unicodedata.normalize("NFC", text).strip()
if not text or len(text) > max_len:
stats[f"{name}_skipped_length"] = stats.get(f"{name}_skipped_length", 0) + 1
continue
ids = encode_label(text, char_to_index)
if ids is None:
stats[f"{name}_skipped_oov"] = stats.get(f"{name}_skipped_oov", 0) + 1
continue
items.append({"path": str(root_path / rel), "text": text, "ids": ids, "aug": float(aug), "source": name})
stats[f"{name}_kept"] = len(items)
return items, float(weight)
def image_widths(items: list[dict]) -> list[int]:
widths = []
for item in items:
with Image.open(item["path"]) as im:
w, h = im.size
widths.append(min(MAX_TRAIN_W, max(MIN_W, int(math.ceil(IMG_H * w / float(h))))))
return widths
def set_train_mode(model: torch.nn.Module) -> None:
"""Train mode with BatchNorm running statistics frozen at their pretrained values.
Width-bucketed batches have very different padding fractions, so live BN statistics
(momentum 0.1, i.e. the last ~10 batches) swing with whatever bucket came last and made
the step-4000 v1b eval 3x worse on the synthetic dev set. Affine BN weights still train.
"""
model.train()
for module in model.modules():
if isinstance(module, torch.nn.modules.batchnorm._BatchNorm):
module.eval()
@torch.no_grad()
def run_eval(model, loader, chars, device) -> dict:
model.eval()
pairs = []
for pixel_values, _, texts in loader:
probs = model(pixel_values.to(device)).last_hidden_state
preds = ctc_greedy_decode(probs.float().cpu(), chars)
pairs += [(p, t) for (p, _), t in zip(preds, texts, strict=True)]
set_train_mode(model)
return evaluate(pairs)
def main() -> int:
p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
p.add_argument("--model-dir", type=Path, required=True)
p.add_argument("--out", type=Path, required=True)
p.add_argument("--train", action="append", required=True, help="name:root:labels:weight:aug")
p.add_argument("--dev", action="append", required=True, help="name:root:labels (max rows via --dev-max)")
p.add_argument("--dev-max", type=int, default=2000)
p.add_argument("--guard-dev", action="append", default=[],
help="dev set name that must not regress beyond --guard-tol vs step 0 (repeatable)")
p.add_argument("--guard-tol", type=float, default=0.002, help="allowed absolute cp CER increase on guard sets")
p.add_argument("--epochs", type=int, default=4)
p.add_argument("--batch-size", type=int, default=64)
p.add_argument("--grad-accum", type=int, default=1,
help="split each batch into this many micro-batches (same gradient, less GPU memory)")
p.add_argument("--lr", type=float, default=1e-4)
p.add_argument("--weight-decay", type=float, default=1e-4)
p.add_argument("--warmup-steps", type=int, default=500)
p.add_argument("--eval-every", type=int, default=1500)
p.add_argument("--max-len", type=int, default=25)
p.add_argument("--workers", type=int, default=6)
p.add_argument("--seed", type=int, default=20260924)
p.add_argument("--device", choices=DEVICE_CHOICES, default="auto")
args = p.parse_args()
torch.manual_seed(args.seed)
device = pick_device(args.device)
if not 1 <= args.grad_accum <= args.batch_size:
raise SystemExit(f"--grad-accum must be between 1 and --batch-size, got {args.grad_accum}")
chars = load_character_list(args.model_dir)
char_to_index = {c: i for i, c in enumerate(chars) if i > 0}
args.out.mkdir(parents=True, exist_ok=True)
stats: dict = {}
items: list[dict] = []
groups: list[tuple[list[int], float]] = []
for spec in args.train:
source_items, weight = load_source(spec, char_to_index, args.max_len, stats)
groups.append((list(range(len(items), len(items) + len(source_items))), weight))
items += source_items
dev_sets = {}
for spec in args.dev:
name, root, labels = spec.split(":")
dev_items, _ = load_source(f"{name}:{root}:{labels}:1:0", char_to_index, 10_000, stats)
dev_items = random.Random(args.seed).sample(dev_items, min(args.dev_max, len(dev_items)))
dev_sets[name] = DataLoader(LineSet(dev_items, train=False, seed=args.seed), batch_size=64,
shuffle=False, num_workers=2, collate_fn=collate)
print(json.dumps(stats, indent=1), flush=True)
try:
target_sets = split_dev_sets(list(dev_sets), args.guard_dev)
except ValueError as exc:
raise SystemExit(f"--guard-dev: {exc}") from exc
print("reading image sizes...", flush=True)
widths = image_widths(items)
sampler = BucketSampler(groups, widths, args.batch_size, args.seed)
loader = DataLoader(LineSet(items, train=True, seed=args.seed), batch_sampler=sampler,
num_workers=args.workers, collate_fn=collate, persistent_workers=True, prefetch_factor=4)
steps_per_epoch = len(sampler)
total_steps = steps_per_epoch * args.epochs
model = load_model(args.model_dir, device)
set_train_mode(model)
optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
def lr_at(step: int) -> float:
if step < args.warmup_steps:
return args.lr * (step + 1) / args.warmup_steps
progress = (step - args.warmup_steps) / max(1, total_steps - args.warmup_steps)
return args.lr * 0.5 * (1 + math.cos(math.pi * min(1.0, progress)))
log = (args.out / "train_log.jsonl").open("a", encoding="utf-8")
guard_baseline: dict[str, float] = {}
def evaluate_and_log(step: int, best: float | None) -> float | None:
results = {name: run_eval(model, dl, chars, device) for name, dl in dev_sets.items()}
cer = {name: r["cer_codepoint"] for name, r in results.items()}
score = target_score(cer, target_sets)
if not guard_baseline:
guard_baseline.update({name: cer[name] for name in args.guard_dev})
regressions = guard_regressions(cer, guard_baseline, args.guard_tol)
record = {"step": step, "dev_mean_cp_cer": float(np.mean([r["cer_codepoint"] for r in results.values()])),
"target_mean_cp_cer": score, "eligible": not regressions, "guard_regressions": regressions,
**{f"{k}_{m}": v[m] for k, v in results.items() for m in ("cer_codepoint", "cer_grapheme", "exact_line_acc")}}
log.write(json.dumps(record) + "\n"); log.flush()
print(json.dumps(record), flush=True)
if regressions:
print(f"step {step}: not eligible as best, guard sets regressed: {regressions}", flush=True)
return best
if best is None or score < best:
model.save_pretrained(args.out / "best")
(args.out / "best" / "preprocessor_config.json").write_text(
(args.model_dir / "preprocessor_config.json").read_text(encoding="utf-8"), encoding="utf-8")
(args.out / "best" / "selected.json").write_text(json.dumps(record, indent=1), encoding="utf-8")
return score
return best
print(f"device={device} train_items={len(items)} steps/epoch={steps_per_epoch} total={total_steps}", flush=True)
best = evaluate_and_log(0, None) # the unmodified starting point, so "best" can only improve on it
step = 0
started = time.time()
for epoch in range(args.epochs):
for pixel_values, ids, _ in loader:
for group in optimizer.param_groups:
group["lr"] = lr_at(step)
optimizer.zero_grad(set_to_none=True)
# Micro-batches weighted by their share of the batch: with mean-reduced CTC and frozen
# BatchNorm statistics the summed gradient equals the full-batch gradient exactly.
loss_value, finite = 0.0, True
for part in torch.arange(len(ids)).chunk(args.grad_accum):
rows = part.tolist()
part_loss = ctc_loss(logits(model, pixel_values[part].to(device)), [ids[i] for i in rows])
if not torch.isfinite(part_loss):
finite = False
break
(part_loss * (len(rows) / len(ids))).backward()
loss_value += part_loss.item() * len(rows) / len(ids)
if not finite:
print(f"step {step}: non-finite loss, batch skipped", flush=True)
optimizer.zero_grad(set_to_none=True)
step += 1
continue
torch.nn.utils.clip_grad_norm_(model.parameters(), 5.0)
optimizer.step()
step += 1
if step % 50 == 0:
rate = step * args.batch_size / (time.time() - started)
print(f"epoch {epoch} step {step}/{total_steps} loss {loss_value:.4f} lr {lr_at(step):.2e} {rate:.0f} img/s", flush=True)
if step % args.eval_every == 0:
best = evaluate_and_log(step, best)
model.save_pretrained(args.out / "last")
best = evaluate_and_log(step, best)
print(f"done: best target mean cp CER {best:.4f} (targets {target_sets}, guards {args.guard_dev})", flush=True)
return 0
if __name__ == "__main__":
sys.exit(main())