SDK 参考

hailo_ipc_sdk 是 NE503 容器应用的 Python SDK，提供 8 个核心模块。安装方式参见应用参考。

快速开始

from hailo_ipc_sdk import InferenceClient

# 订阅式推理：stream 用 sub（发原始 NV12 帧），model 用设备真实名
inf = InferenceClient()
for seq, result in inf.subscribe(stream="sub", model="hailo_yolov8n_384_640"):
    print(f"Detected {len(result.objects)} objects")

三类调用范式（推理订阅 / 事件发布订阅 / 设备控制）与「名字不能写死、订阅是阻塞迭代器、优雅退出」要点见 SDK 工作流 §3 调用范式；各模块完整 API 见下文 §1–§8。

---

1. inference — AI 推理

数据类

BoundingBox

@dataclass
class BoundingBox:
    x: float
    y: float
    width: float
    height: float
    def to_xyxy(self) -> Tuple[float, float, float, float]
    def to_xywh(self) -> Tuple[float, float, float, float]

DetectedObject

@dataclass
class DetectedObject:
    label: str
    score: float
    bbox: BoundingBox
    class_id: int = 0
    track_id: Optional[int] = None

LandmarkPoint / LandmarkSet

@dataclass
class LandmarkPoint:
    x: float
    y: float
    confidence: float = 1.0

@dataclass
class LandmarkSet:
    type: str
    points: List[LandmarkPoint]

Classification

@dataclass
class Classification:
    type: str       # age, gender, clip 等
    class_id: int
    label: str
    confidence: float

SegmentationMask

@dataclass
class SegmentationMask:
    class_id: int
    label: str
    confidence: float
    bbox: BoundingBox
    mask_rle: bytes
    mask_width: int
    mask_height: int
    def to_numpy_mask() -> np.ndarray

OcrLine / Embedding / DepthMap

@dataclass
class OcrLine:
    text: str
    confidence: float
    bbox: BoundingBox

@dataclass
class Embedding:
    dim: int
    data: List[float]

@dataclass
class DepthMap:
    width: int
    height: int
    data: np.ndarray    # float32 (H, W)

InferenceResult

@dataclass
class InferenceResult:
    frame_sequence: int
    timestamp_ns: int
    objects: List[DetectedObject]
    classifications: List[Classification]
    landmarks: List[LandmarkSet]
    masks: List[SegmentationMask]
    ocr_lines: List[OcrLine]
    embeddings: List[Embedding]
    depth_maps: List[DepthMap]
    raw_outputs: Optional[List[np.ndarray]]
    infer_time_us: int = 0
    queue_time_us: int = 0
    hw_infer_time_us: int = 0       # 纯 NPU 硬件推理耗时（μs），不可用时为 0
    status_message: str = ""        # 诊断信息；"simulation" 表示无真实帧源（降级模式）

    def has_person() -> bool
    def count_by_label(label: str) -> int
    def get_objects_by_label(label: str) -> List[DetectedObject]

ModelInfo

@dataclass
class ModelInfo:
    model_id: str
    model_path: str
    version: str = ""
    inputs: List[Dict]
    outputs: List[Dict]
    estimated_tops: float = 0.0
    estimated_memory: int = 0
    load_timestamp: int = 0

InferenceClient

class InferenceClient:
    def __init__(self, endpoint: Optional[str] = None)
    def connect() -> None
    def close() -> None

推理操作：

方法	签名	说明
infer	infer(image: np.ndarray, model_id: str, timeout_ms: int = 5000, priority: int = 4) -> InferenceResult	单帧推理
infer_with_tensors	infer_with_tensors(model_id: str, inputs: List[np.ndarray], timeout_ms: int = 5000) -> List[np.ndarray]	多输入推理
subscribe	subscribe(stream: str, model: str, fps: int = 10, raw_output_only: bool = False) -> Iterator[Tuple[int, InferenceResult]]	流式推理

模型管理：

方法	签名	说明
register_model	register_model(model_path: str, model_id: Optional[str] = None, ...) -> str	注册模型
unregister_model	unregister_model(model_id: str) -> None	注销模型
list_models	list_models() -> List[ModelInfo]	列出模型
get_model_info	get_model_info(model_id: str) -> Optional[ModelInfo]	模型详情

会话管理：

方法	签名	说明
create_session	create_session(session_id: str, max_qps: int = 0, max_concurrent: int = 0, priority: int = 4) -> str	创建会话
destroy_session	destroy_session(session_id: str) -> None	销毁会话

GenAI：

方法	签名	说明
genai_create_session	genai_create_session(hef_path: str, kind: str = "llm") -> str	创建 GenAI 会话
genai_destroy_session	genai_destroy_session(session_id: str) -> None	销毁会话
genai_generate	genai_generate(session_id: str, messages: List[str], max_tokens: int = 512, ...) -> Iterator[str]	流式生成
genai_abort	genai_abort(session_id: str) -> None	中止生成

其他：

方法	说明
get_stats() -> Dict	系统统计
encode_text(text: str) -> List[float]	CLIP 文本编码
update_postprocess_config(model_id: str, config_json: str) -> bool	更新后处理配置

---

2. media — 视频流

PixelFormat

class PixelFormat(IntEnum):
    NV12 = 0; NV21 = 1; RGB = 2; BGR = 3
    RGBA = 4; BGRA = 5; GRAY8 = 6; YUYV = 7

Frame

@dataclass
class Frame:
    sequence: int
    timestamp_ns: int
    width: int
    height: int
    format: str
    image: np.ndarray
    def to_rgb() -> np.ndarray
    def save(path: str) -> None

StreamInfo

@dataclass
class StreamInfo:
    stream_id: str
    width: int; height: int
    format: str; fps: float
    buffer_count: int

FdMediaClient

零拷贝视频帧获取（Unix Socket + DMA-BUF）：

class FdMediaClient:
    def __init__(self, socket_path: str | None = None)
    def get_frame(stream_id: str, timeout_ms: int = 5000) -> Frame | None
    def subscribe_raw(stream_id: str, skip_frames: bool = True) -> Iterator[Frame]
    def on_frame(stream_id: str, callback: Callable) -> threading.Thread
    def list_streams() -> List[str]          # 如 ['main', 'sub']；推理用 sub
    def close() -> None

EncodedStreamClient

读取编码视频流（H.264/H.265 NAL 单元），用于 RTSP 转推、录像等不需要原始帧的场景。Socket 路径形如 /run/aipc/encoded/main.sock。

@dataclass
class EncodedFrame:
    codec: int            # 0=h264, 1=h265
    flags: int            # bit0 = 关键帧
    pts_ns: int           # 显示时间戳（纳秒）
    width: int
    height: int
    dts_ns: int           # 解码时间戳（纳秒）
    data: bytes           # 编码 NAL 负载
    @property
    def is_keyframe() -> bool
    @property
    def codec_name() -> str       # "h264" / "h265"

class EncodedStreamClient:
    def __init__(self, socket_path: str)
    def get_frame(timeout_ms: int = 5000) -> EncodedFrame | None
    def subscribe(reconnect: bool = True) -> Iterator[EncodedFrame]
    def on_frame(callback: Callable) -> threading.Thread
    def close() -> None

FdMediaClient.get_encoded_stream(stream_id="main") 可便捷返回一个已连接的 EncodedStreamClient。视频流接入（RTSP/WebSocket）详见 Video Integration。

---

3. events — 事件总线

Event / TopicInfo

@dataclass
class Event:
    topic: str
    payload: Dict[str, Any]
    source: str = ""
    event_id: str = ""
    timestamp_ns: int = 0
    metadata: Dict[str, str]
    def to_json() -> str

@dataclass
class TopicInfo:
    topic: str
    subscriber_count: int
    total_messages: int
    last_message_ts: int

EventClient

class EventClient:
    def __init__(self, endpoint: Optional[str] = None)

方法	说明
publish(topic, payload, persistent=False, ttl_ms=None, metadata=None) -> str	发布事件
publish_batch(events, persistent=False) -> None	批量发布
subscribe(topic, filters=None, queue_size=100) -> Iterator[Event]	订阅（支持通配符）
on_event(topic, callback, filters=None) -> Thread	回调订阅
unsubscribe(topic) -> None	取消订阅
list_topics() -> List[TopicInfo]	列出主题
get_topic_info(topic) -> Optional[TopicInfo]	主题详情
get_stats() -> Dict	系统统计
get_topic_stats(topic) -> Dict	主题统计

---

4. device — 设备控制

IrCutMode / DeviceStatus / DeviceEvent

class IrCutMode(Enum):
    AUTO = 0; DAY = 1; NIGHT = 2

@dataclass
class DeviceStatus:
    soc_temp_c: float; mcu_temp_c: float
    light_sensor: int; zoom_pos: int; focus_pos: int
    autofocus_enabled: bool; ircut_mode: IrCutMode
    white_light_level: int; ir_led_on: bool
    mcu_version: str

@dataclass
class DeviceEvent:
    type: EventType    # GPIO_CHANGE, TEMPERATURE_ALERT, ...
    timestamp_ns: int

DeviceClient

class DeviceClient:
    def __init__(self, endpoint: Optional[str] = None)

灯光：set_white_light(level: int)、set_ir_led(on: bool)、set_ircut(mode: IrCutMode)

云台：pan_left/right(speed)、tilt_up/down(speed)、ptz_stop()、save_preset(id)、call_preset(id)

镜头：zoom(speed)、set_zoom_level(level)、focus(speed)、focus_auto(enable)、set_focus_level(level)、get_lens_status()、lens_init()、lens_goto_ratio_distance(zoom_ratio, focus_distance_m)、control_iris(open)、set_iris_target(target)、lens_reset_zero(zoom, focus)

GPIO：gpio_set(pin, value)、gpio_get(pin) -> bool

状态：get_device_status() -> DeviceStatus、subscribe_events() -> Iterator[DeviceEvent]

---

5. app — 应用管理

AppInfo / AppStats / LogLine

@dataclass
class AppInfo:
    id: str; name: str; version: str; state: str
    container_id: str; pid: int
    installed_at: int; started_at: int; restart_count: int

@dataclass
class AppStats:
    cpu_usage_percent: float; memory_usage_bytes: int
    memory_limit_bytes: int; thread_count: int

@dataclass
class LogLine:
    timestamp: int; level: str; message: str

AppClient

class AppClient:
    def __init__(self, endpoint: Optional[str] = None)

方法	说明
install_app(manifest_path, image_path) -> str	安装应用
start_app(app_id)	启动
stop_app(app_id, timeout_seconds=30)	停止
uninstall_app(app_id, keep_logs=True)	卸载
list_apps() -> List[AppInfo]	列出应用
get_app(app_id) -> AppInfo	应用详情
get_app_stats(app_id) -> AppStats	资源统计
get_logs(app_id, max_lines=100, follow=False) -> Iterator[LogLine]	获取日志
get_logs_text(app_id, max_lines=100, follow=False) -> Iterator[str]	文本日志

---

6. plugin — 插件发现

PluginEndpoint

@dataclass
class PluginEndpoint:
    app_id: str; capability_id: str; version: str
    transport: str; socket_path: Optional[str]; state: str
    def connect(**kwargs) -> grpc.Channel
    @property
    def is_available() -> bool

PluginDiscovery

方法	说明
get(capability_id: str) -> Optional[PluginEndpoint]	查找插件
require(capability_id: str, timeout: float = 30.0) -> PluginEndpoint	等待插件可用
list_plugins() -> Dict[str, dict]	列出所有插件
list_capabilities() -> List[str]	列出所有能力
reload() -> None	重新加载
watch(callback: Callable) -> None	监听变化
close() -> None	关闭并释放资源

PluginServer

class PluginServer:
    def __init__(self, plugin_id: str, socket_dir: str = DISCOVERY_DIR)
    def create_server(max_workers: int = 4) -> grpc.Server
    def start() -> None
    def stop(grace: float = 5.0) -> None

---

7. overlay — AI 结果叠加

@dataclass
class OverlayConfig:
    enabled: bool = True
    show_label: bool = True
    show_confidence: bool = True
    line_thickness: int = 2
    box_color: int = 0
    label_color: int = 0
    font_size: int = 0

class OverlayClient:
    def enable(show_label=True, show_confidence=True, line_thickness=2)
    def disable()
    def configure(enabled=True, show_label=True, ...)
    def apply(config: OverlayConfig)

---

8. config — 配置管理

class Config:
    @staticmethod
    def get_app_id() -> str
    @staticmethod
    def get_inference_endpoint() -> str
    @staticmethod
    def get_event_bus_endpoint() -> str
    @staticmethod
    def get_device_control_endpoint() -> str
    @staticmethod
    def get_camera_control_endpoint() -> str
    @staticmethod
    def translate_path_to_host(container_path: str) -> str
    @staticmethod
    def is_debug() -> bool

环境变量

Config 读取的连接端点（AI_RUNTIME_ENDPOINT、EVENT_BUS_ENDPOINT、DEVICE_CONTROL_ENDPOINT、CAMERA_CONTROL_ENDPOINT）及平台自动注入的容器环境变量（APP_ID、AIPC_HOST_PREFIX、SHM_BASE_PATH、LOG_LEVEL 等）完整列表，见 应用参考 §7 环境变量参考。

---

相关文档

• SDK 示例 — 完整应用示例和开发指南
• 应用参考 — 项目创建、app.yaml 配置、部署流程
• 平台服务总览 — AI Runtime、Event Bus 等服务职责与源码指针