dekun/secondary-school-grade-archive
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

OCR Worker 改用 RapidOCR/ONNX，修复 Paddle SIGILL。

Browse Source
This commit is contained in:
dekun
2026-06-28 15:27:59 +08:00
parent 035b65dcc8
commit 4de460c235
5 changed files with 34 additions and 108 deletions
Download Patch File Download Diff File Expand all files Collapse all files
deploy/ocr-common.sh
+6 -10
View File
@@ -36,11 +36,9 @@ install_ocr_worker() {
log_error "未找到 ${worker_dir}"
return 1
fi
log_info "安装/更新 OCR Worker (PaddleOCR)…"
log_info "安装/更新 OCR Worker (RapidOCR/ONNX)…"
chmod +x "${worker_dir}"/*.sh 2>/dev/null || true
local use_gpu
use_gpu="$(detect_ocr_use_gpu)"
OCR_PORT="${OCR_PORT}" OCR_USE_GPU="${use_gpu}" bash "${worker_dir}/install.sh"
OCR_PORT="${OCR_PORT}" bash "${worker_dir}/install.sh"
}
ocr_screen_running() {
@@ -48,7 +46,7 @@ ocr_screen_running() {
}
start_ocr_screen() {
local worker_dir
local worker_dir log_file
worker_dir="$(ocr_worker_dir)"
if [[ ! -x "${worker_dir}/.venv/bin/uvicorn" ]]; then
log_warn "OCR Worker 未安装，跳过 screen 启动"
@@ -59,20 +57,18 @@ start_ocr_screen() {
return 1
fi
local use_gpu log_file
use_gpu="$(detect_ocr_use_gpu)"
log_info "启动 OCR Worker → screen「${OCR_SCREEN_NAME}」(GPU=${use_gpu}, 端口 ${OCR_PORT})"
log_info "启动 OCR Worker → screen「${OCR_SCREEN_NAME}」(RapidOCR, 端口 ${OCR_PORT})"
if ocr_screen_running; then
screen -S "${OCR_SCREEN_NAME}" -X quit 2>/dev/null || true
sleep 1
fi
mkdir -p "${INSTALL_DIR}/logs" 2>/dev/null || true
local log_file="${INSTALL_DIR}/logs/ocr-worker.log"
log_file="${INSTALL_DIR}/logs/ocr-worker.log"
screen -dmS "${OCR_SCREEN_NAME}" bash -c "
cd '${worker_dir}' &&
export OCR_USE_GPU='${use_gpu}' OCR_PORT='${OCR_PORT}' OCR_HOST=0.0.0.0 &&
export OCR_PORT='${OCR_PORT}' OCR_HOST=0.0.0.0 &&
exec bash run.sh >> '${log_file}' 2>&1
"
sleep 2
deploy/ocr-worker/app.py
+19 -68
View File
@@ -1,4 +1,4 @@
"""局域网 OCR 服务：在带 NVIDIA 显卡的机器上运行，供成绩档案系统调用。"""
"""局域网 OCR 服务：RapidOCR(ONNX)，不依赖 Paddle，避免 SIGILL/cuDNN 问题。"""
import logging
import os
@@ -9,19 +9,14 @@ from pathlib import Path
from fastapi import FastAPI, File, Header, HTTPException, UploadFile
from PIL import Image
os.environ.setdefault("OPENCV_IO_ENABLE_OPENEXR", "0")
os.environ.setdefault("FLAGS_use_mkldnn", "0")
logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
logger = logging.getLogger("ocr-worker")
OCR_MAX_SIDE = int(os.getenv("OCR_MAX_SIDE", "1280"))
OCR_API_KEY = os.getenv("OCR_API_KEY", "").strip()
OCR_USE_GPU = os.getenv("OCR_USE_GPU", "true").lower() in {"1", "true", "yes"}
app = FastAPI(title="Grade Archive OCR Worker", version="1.0.0")
app = FastAPI(title="Grade Archive OCR Worker", version="2.0.0")
_engine = None
_engine_mode = "none"
def _check_key(key: str | None) -> None:
@@ -29,54 +24,15 @@ def _check_key(key: str | None) -> None:
raise HTTPException(status_code=401, detail="Invalid OCR API key")
def _create_engine(use_gpu: bool):
from paddleocr import PaddleOCR
def get_engine():
global _engine
if _engine is None:
from rapidocr_onnxruntime import RapidOCR
return PaddleOCR(
use_angle_cls=False,
lang="ch",
show_log=False,
use_gpu=use_gpu,
enable_mkldnn=False,
ir_optim=False,
det_limit_side_len=min(OCR_MAX_SIDE, 1280),
rec_batch_num=8,
)
def get_engine(force_cpu: bool = False):
global _engine, _engine_mode
if _engine is not None and not force_cpu:
return _engine
modes: list[bool] = [False] if force_cpu or not OCR_USE_GPU else [True, False]
last_err: Exception | None = None
for use_gpu in modes:
try:
logger.info("Loading PaddleOCR use_gpu=%s", use_gpu)
_engine = _create_engine(use_gpu)
_engine_mode = "gpu" if use_gpu else "cpu"
logger.info("PaddleOCR ready mode=%s", _engine_mode)
return _engine
except Exception as exc:
last_err = exc
logger.warning("PaddleOCR init failed use_gpu=%s: %s", use_gpu, exc)
_engine = None
_engine_mode = "none"
hint = ""
err_text = str(last_err or "")
if "libGL" in err_text:
hint = " 请执行: sudo bash deploy/install-ocr-deps.sh 后重启 OCR"
elif any(x in err_text.lower() for x in ("cuda", "cudnn", "gpu", "out of memory")):
hint = " 显存不足或 CUDA 异常，可设置 OCR_USE_GPU=false 用 CPU"
raise RuntimeError(f"PaddleOCR 初始化失败: {last_err}{hint}") from last_err
def _reset_engine():
global _engine, _engine_mode
_engine = None
_engine_mode = "none"
logger.info("Loading RapidOCR (ONNX CPU)…")
_engine = RapidOCR()
logger.info("RapidOCR ready")
return _engine
def _bbox_from_box(box: list) -> list[float]:
@@ -110,32 +66,31 @@ def _prepare_image_bytes(content: bytes) -> tuple[bytes, float, float, int, int]
return buf.getvalue(), scale_x, scale_y, orig_w, orig_h
def _run_ocr_impl(content: bytes) -> dict:
def run_ocr_on_bytes(content: bytes) -> dict:
engine = get_engine()
image_bytes, scale_x, scale_y, orig_w, orig_h = _prepare_image_bytes(content)
with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as tmp:
tmp.write(image_bytes)
tmp_path = tmp.name
try:
result = engine.ocr(tmp_path, cls=False)
result, _elapsed = engine(tmp_path)
finally:
Path(tmp_path).unlink(missing_ok=True)
lines: list[dict] = []
if result and result[0]:
for item in result[0]:
if result:
for item in result:
if not item or len(item) < 2:
continue
box, rec = item[0], item[1]
text = rec[0] if rec else ""
conf = float(rec[1]) if rec and len(rec) > 1 else 0.0
box, text = item[0], item[1]
conf = float(item[2]) if len(item) > 2 else 0.0
if not text:
continue
if scale_x != 1.0 or scale_y != 1.0:
box = _scale_box(box, scale_x, scale_y)
lines.append(
{
"text": text,
"text": str(text),
"confidence": conf,
"box": box,
"bbox": _bbox_from_box(box),
@@ -147,17 +102,13 @@ def _run_ocr_impl(content: bytes) -> dict:
"lines": lines,
"width": orig_w,
"height": orig_h,
"engine_mode": _engine_mode,
"engine_mode": "rapidocr-onnx",
}
def run_ocr_on_bytes(content: bytes) -> dict:
return _run_ocr_impl(content)
@app.get("/health")
def health():
return {"status": "ok", "gpu_requested": OCR_USE_GPU, "engine_mode": _engine_mode}
return {"status": "ok", "engine": "rapidocr-onnxruntime"}
@app.post("/api/ocr/regions")
deploy/ocr-worker/install.sh
+6 -27
View File
@@ -1,5 +1,5 @@
#!/usr/bin/env bash
# 在带 NVIDIA 显卡的 Linux 机器上安装 OCR Worker（由 deploy/install.sh 自动调用）
# OCR Worker 安装（RapidOCR / ONNX，无需 Paddle/GPU）
set -euo pipefail
ROOT="$(cd "$(dirname "$0")" && pwd)"
@@ -8,6 +8,7 @@ PORT="${OCR_PORT:-23567}"
PIP_MIRROR="${PIP_MIRROR:-https://pypi.tuna.tsinghua.edu.cn/simple}"
echo "==> OCR Worker 安装目录: ${ROOT}"
echo "==> 引擎: RapidOCR (ONNX CPU，无需 cuDNN/GPU)"
if ! command -v python3 >/dev/null; then
echo "错误: 请先安装 python3"
@@ -15,7 +16,7 @@ if ! command -v python3 >/dev/null; then
fi
if [[ -d "${VENV}" ]]; then
echo "==> 已有虚拟环境，跳过 python3 -m venv"
echo "==> 已有虚拟环境"
else
python3 -m venv "${VENV}"
fi
@@ -23,33 +24,11 @@ fi
# shellcheck disable=SC1091
source "${VENV}/bin/activate"
pip install -U pip wheel -i "${PIP_MIRROR}"
install_paddle() {
local use_gpu="${OCR_USE_GPU:-false}"
if [[ "${use_gpu}" == "true" ]] && command -v nvidia-smi >/dev/null 2>&1 && ldconfig -p 2>/dev/null | grep -q libcudnn; then
local cuda_major
cuda_major="$(nvidia-smi 2>/dev/null | sed -n 's/.*CUDA Version: \([0-9]*\)\.[0-9]*/\1/p' | head -1)"
cuda_major="${cuda_major:-11}"
echo "==> 安装 paddlepaddle-gpu (CUDA ${cuda_major}.x)…"
if [[ "${cuda_major}" -ge 12 ]]; then
pip install paddlepaddle-gpu==2.6.2 -i https://www.paddlepaddle.org.cn/packages/stable/cu123/ \
|| pip install paddlepaddle-gpu==2.6.2 -i https://www.paddlepaddle.org.cn/packages/stable/cu118/
else
pip install paddlepaddle-gpu==2.6.2 -i https://www.paddlepaddle.org.cn/packages/stable/cu118/
fi
else
echo "==> 安装 paddlepaddle CPU（无 cuDNN 或 OCR_USE_GPU=false）…"
pip install paddlepaddle==2.6.2 -i "${PIP_MIRROR}"
fi
pip install 'protobuf>=3.20,<4' -q
}
install_paddle
pip uninstall -y paddlepaddle paddlepaddle-gpu paddleocr 2>/dev/null || true
pip install -r "${ROOT}/requirements.txt" -i "${PIP_MIRROR}"
chmod +x "${ROOT}/run.sh" "${ROOT}/start.sh" 2>/dev/null || true
echo ""
python3 -c "import fastapi, uvicorn, paddle; print('paddle', paddle.__version__, 'OK')"
python3 -c "from rapidocr_onnxruntime import RapidOCR; print('RapidOCR OK')"
echo ""
echo "==> OCR Worker 安装完成。由 deploy/install.sh 通过 screen 自动启动。"
echo " 手动管理: bash $(dirname "$ROOT")/ocr-screen.sh status"
echo "==> 安装完成。管理: bash $(dirname "$ROOT")/ocr-screen.sh status"
deploy/ocr-worker/requirements.txt
+2 -2
View File
@@ -1,6 +1,6 @@
fastapi==0.115.6
uvicorn[standard]==0.34.0
python-multipart==0.0.20
paddleocr==2.9.1
Pillow==11.0.0
# GPU 版 Paddle 请用 install.sh 安装，勿直接 pip install paddlepaddle
rapidocr-onnxruntime>=1.3.0
onnxruntime>=1.16.0
deploy/ocr-worker/run.sh
+1 -1
View File
@@ -6,7 +6,7 @@ ROOT="$(cd "$(dirname "$0")" && pwd)"
VENV="${ROOT}/.venv"
PORT="${OCR_PORT:-23567}"
export OCR_USE_GPU="${OCR_USE_GPU:-true}"
export OCR_USE_GPU="${OCR_USE_GPU:-false}"
export OCR_HOST="${OCR_HOST:-0.0.0.0}"
if [[ ! -d "${VENV}" ]]; then