diff --git a/README.md b/README.md index 03c99a0..1094b27 100644 --- a/README.md +++ b/README.md @@ -152,6 +152,33 @@ batch 8 + cpu_threads 4 : 约 114 秒 50% → 80% 文字识别 约 50 秒 ← 主要瓶颈 ``` +注意:同一配置在不同时间实测波动较大(同一图片分别测到 66 秒和 128 秒),比较参数时请在相同时间段多测几次,不要只看单次结果。 + +### 是否使用多核 + +不是单核推理。PaddleX 默认 `cpu_threads = 10`(可用 `cpu_threads` 配置项覆盖),20 逻辑核机器上的实测 CPU 占用: + +```text +检测阶段 CPU 约 94%~105% (约 1 个核) +识别阶段 CPU 约 534%~686% (约 5.3~6.9 个核) +平均约 553%,即约 5.5 个核,20 核未跑满 +``` + +结论: + +- 识别阶段是多线程的,约占 5~7 个核; +- 检测阶段接近单核,这是关闭 oneDNN 后检测耗时约 13 秒的主要原因; +- 20 核没有跑满,说明并行效率不高,瓶颈在「未启用 oneDNN 的 CPU 算子」,而不是核心数不够。 + +线程数对耗时的影响(同图、`batch 8`): + +```text +cpu_threads=4 约 114 秒 +cpu_threads=10 约 60 秒 +``` + +可调整 `cpu_threads` 实测,建议不超过物理核心数,并用 `benchmark_ocr.py` 对比。 + 关闭 oneDNN 是必须的:`enable_mkldnn=True` 会触发 PaddlePaddle 3.3.1 的缺陷并直接失败: ```text @@ -509,6 +536,7 @@ log_level: INFO | `text_det_unclip_ratio` | `1.5` | 检测框扩张比例 | | `text_rec_score_thresh` | `0.0` | 识别结果阈值 | | `text_recognition_batch_size` | `1` | 识别批量大小;内存足够时可提高 | +| `cpu_threads` | `10` | CPU 推理线程数(Paddle `cpu_math_library_num_threads`) | | `log_level` | `INFO` | 日志级别:`DEBUG`、`INFO`、`WARNING`、`ERROR` | 接口文档:启动后访问 。 diff --git a/config.yaml b/config.yaml index fb2be3a..221de17 100644 --- a/config.yaml +++ b/config.yaml @@ -34,6 +34,11 @@ text_rec_score_thresh: 0.0 # 识别批量大小 text_recognition_batch_size: 1 +# CPU 推理线程数(Paddle 的 cpu_math_library_num_threads)。 +# PaddleX 默认 10;不是单核推理,但关闭 oneDNN 后多线程加速有限。 +# 建议不要超过物理核心数;调大不一定更快,请用 benchmark_ocr.py 实测。 +cpu_threads: 10 + # 日志级别:DEBUG / INFO / WARNING / ERROR log_level: INFO diff --git a/scripts/benchmark_ocr.py b/scripts/benchmark_ocr.py index 5dd8edc..96fb9c0 100644 --- a/scripts/benchmark_ocr.py +++ b/scripts/benchmark_ocr.py @@ -23,11 +23,22 @@ if __name__ == "__main__": text_det_unclip_ratio=settings.text_det_unclip_ratio, text_rec_score_thresh=settings.text_rec_score_thresh, text_recognition_batch_size=settings.text_recognition_batch_size, + cpu_threads=settings.cpu_threads, ) service.prepare() + + stages: list[str] = [] + + def on_progress(percent: int, message: str) -> None: + stages.append(f"{percent}% {message}") + started = perf_counter() - lines = service.recognize(image_path) + lines = service.recognize(image_path, on_progress=on_progress) elapsed_ms = (perf_counter() - started) * 1000 print(f"image={image_path}") + print(f"cpu_threads={settings.cpu_threads}") + print(f"det={settings.det_model_name} rec={settings.rec_model_name}") print(f"elapsed_ms={elapsed_ms:.2f}") print(f"segments={len(lines)}") + for stage in stages: + print(f"stage={stage}") diff --git a/scripts/download_models.py b/scripts/download_models.py index 63d94ae..198d3ff 100644 --- a/scripts/download_models.py +++ b/scripts/download_models.py @@ -19,5 +19,6 @@ if __name__ == "__main__": text_det_unclip_ratio=settings.text_det_unclip_ratio, text_rec_score_thresh=settings.text_rec_score_thresh, text_recognition_batch_size=settings.text_recognition_batch_size, + cpu_threads=settings.cpu_threads, ).prepare() print(f"Models are ready in: {settings.model_dir.resolve()}") diff --git a/src/ocr_server/app.py b/src/ocr_server/app.py index a294171..bbaa792 100644 --- a/src/ocr_server/app.py +++ b/src/ocr_server/app.py @@ -28,6 +28,7 @@ def create_app(settings: Settings | None = None) -> FastAPI: text_det_unclip_ratio=settings.text_det_unclip_ratio, text_rec_score_thresh=settings.text_rec_score_thresh, text_recognition_batch_size=settings.text_recognition_batch_size, + cpu_threads=settings.cpu_threads, ) tasks = OCRTaskQueue(create_ocr, worker_count=settings.ocr_workers) diff --git a/src/ocr_server/config.py b/src/ocr_server/config.py index 2624960..b27ae19 100644 --- a/src/ocr_server/config.py +++ b/src/ocr_server/config.py @@ -37,6 +37,7 @@ class Settings(BaseModel): text_det_unclip_ratio: float = Field(default=1.5, ge=0.1, le=5.0) text_rec_score_thresh: float = Field(default=0.0, ge=0.0, le=1.0) text_recognition_batch_size: int = Field(default=1, ge=1, le=64) + cpu_threads: int = Field(default=10, ge=1, le=64) log_level: str = "INFO" def __init__(self, **overrides: Any) -> None: diff --git a/src/ocr_server/ocr.py b/src/ocr_server/ocr.py index 7b69ad9..dab4c98 100644 --- a/src/ocr_server/ocr.py +++ b/src/ocr_server/ocr.py @@ -41,6 +41,7 @@ class OCRService: text_det_unclip_ratio: float = 1.5, text_rec_score_thresh: float = 0.0, text_recognition_batch_size: int = 1, + cpu_threads: int = 10, ) -> None: self.lang = lang self.device = device @@ -54,6 +55,7 @@ class OCRService: self._text_det_unclip_ratio = text_det_unclip_ratio self._text_rec_score_thresh = text_rec_score_thresh self._text_recognition_batch_size = text_recognition_batch_size + self._cpu_threads = cpu_threads self._engine: Any | None = None self._lock = threading.Lock() self._inference_lock = threading.Lock() @@ -130,10 +132,11 @@ class OCRService: ) from exc with self._process_model_lock: logger.info( - "加载 OCR 引擎: det=%s rec=%s device=%s model_dir=%s", + "加载 OCR 引擎: det=%s rec=%s device=%s cpu_threads=%d model_dir=%s", self._det_model_name, self._rec_model_name, self.device, + self._cpu_threads, self.model_dir, ) self._engine = PaddleOCR( @@ -150,6 +153,7 @@ class OCRService: text_det_unclip_ratio=self._text_det_unclip_ratio, text_rec_score_thresh=self._text_rec_score_thresh, text_recognition_batch_size=self._text_recognition_batch_size, + cpu_threads=self._cpu_threads, ) self._install_progress_hooks(self._engine) logger.info("OCR 引擎加载完成")