Sinks

class UDPSink:
    @classmethod
    def init(cls, ip_address: str, port: int) -> "UDPSink":
        """
        Creates `InferencePipeline` predictions sink capable of sending model predictions over network
        using UDP socket.

        As an `inference` user, please use .init() method instead of constructor to instantiate objects.
        Args:
            ip_address (str): IP address to send predictions
            port (int): Port to send predictions

        Returns: Initialised object of `UDPSink` class.
        """
        udp_socket = socket.socket(family=socket.AF_INET, type=socket.SOCK_DGRAM)
        udp_socket.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
        udp_socket.setsockopt(socket.SOL_SOCKET, socket.SO_BROADCAST, 1)
        udp_socket.setsockopt(socket.SOL_SOCKET, socket.SO_RCVBUF, 1)
        udp_socket.setsockopt(socket.SOL_SOCKET, socket.SO_SNDBUF, 65536)
        return cls(
            ip_address=ip_address,
            port=port,
            udp_socket=udp_socket,
        )

    def __init__(self, ip_address: str, port: int, udp_socket: socket.socket):
        self._ip_address = ip_address
        self._port = port
        self._socket = udp_socket

    def send_predictions(
        self,
        predictions: Union[dict, List[Optional[dict]]],
        video_frame: Union[VideoFrame, List[Optional[VideoFrame]]],
    ) -> None:
        """
        Method to send predictions via UDP socket. Useful in combination with `InferencePipeline` as
        a sink for predictions.

        Args:
            predictions (Union[dict, List[Optional[dict]]]): Roboflow predictions, the function support single prediction
                processing and batch processing since version `0.9.18`. Batch predictions elements are optional, but
                should occur at the same position as `video_frame` list. Order is expected to match with `video_frame`.
            video_frame (Union[VideoFrame, List[Optional[VideoFrame]]]): frame of video with its basic metadata emitted
                by `VideoSource` or list of frames from (it is possible for empty batch frames at corresponding positions
                to `predictions` list). Order is expected to match with `predictions`

        Returns: None
        Side effects: Sends serialised `predictions` and `video_frame` metadata via the UDP socket as
            JSON string. It adds key named "inference_metadata" into `predictions` dict (mutating its
            state). "inference_metadata" contain id of the frame, frame grabbing timestamp and message
            emission time in datetime iso format.

        Example:
            ```python
            import cv2
            from inference.core.interfaces.stream.inference_pipeline import InferencePipeline
            from inference.core.interfaces.stream.sinks import UDPSink

            udp_sink = UDPSink.init(ip_address="127.0.0.1", port=9090)

            pipeline = InferencePipeline.init(
                 model_id="your-model/3",
                 video_reference="./some_file.mp4",
                 on_prediction=udp_sink.send_predictions,
            )
            pipeline.start()
            pipeline.join()
            ```
            `UDPSink` used in this way will emit predictions to receiver automatically.
        """
        video_frame = wrap_in_list(element=video_frame)
        predictions = wrap_in_list(element=predictions)
        for single_frame, frame_predictions in zip(video_frame, predictions):
            if single_frame is None:
                continue
            inference_metadata = {
                "source_id": single_frame.source_id,
                "frame_id": single_frame.frame_id,
                "frame_decoding_time": single_frame.frame_timestamp.isoformat(),
                "emission_time": datetime.now().isoformat(),
            }
            frame_predictions["inference_metadata"] = inference_metadata
            serialised_predictions = json.dumps(frame_predictions).encode("utf-8")
            self._socket.sendto(
                serialised_predictions,
                (
                    self._ip_address,
                    self._port,
                ),
            )

class VideoFileSink:
    @classmethod
    def init(
        cls,
        video_file_name: str,
        annotator: Optional[Union[BaseAnnotator, List[BaseAnnotator]]] = None,
        display_size: Optional[Tuple[int, int]] = (1280, 720),
        fps_monitor: Optional[sv.FPSMonitor] = DEFAULT_FPS_MONITOR,
        display_statistics: bool = False,
        output_fps: int = 25,
        quiet: bool = False,
        video_frame_size: Tuple[int, int] = (1280, 720),
    ) -> "VideoFileSink":
        """
        Creates `InferencePipeline` predictions sink capable of saving model predictions into video file.
        It works both for pipelines with single input video and multiple ones.

        As an `inference` user, please use .init() method instead of constructor to instantiate objects.
        Args:
            video_file_name (str): name of the video file to save predictions
            annotator (Union[BaseAnnotator, List[BaseAnnotator]]): instance of class inheriting from supervision BaseAnnotator
                or list of such instances. If nothing is passed chain of `sv.BoxAnnotator()` and `sv.LabelAnnotator()` is used.
            display_size (Tuple[int, int]): tuple in format (width, height) to resize visualisation output. Should
                be set to the same value as `display_size` for InferencePipeline with single video source, otherwise
                it represents the size of single visualisation tile (whole tiles mosaic will be scaled to
                `video_frame_size`)
            fps_monitor (Optional[sv.FPSMonitor]): FPS monitor used to monitor throughput
            display_statistics (bool): Flag to decide if throughput and latency can be displayed in the result image,
                if enabled, throughput will only be presented if `fps_monitor` is not None
            output_fps (int): desired FPS of output file
            quiet (bool): Flag to decide whether to log progress
            video_frame_size (Tuple[int, int]): The size of frame in target video file.

        Attributes:
            on_prediction (Callable[[dict, VideoFrame], None]): callable to be used as a sink for predictions

        Returns: Initialized object of `VideoFileSink` class.

        Example:
            ```python
            import cv2
            from inference import InferencePipeline
            from inference.core.interfaces.stream.sinks import VideoFileSink

            video_sink = VideoFileSink.init(video_file_name="output.avi")

            pipeline = InferencePipeline.init(
                model_id="your-model/3",
                video_reference="./some_file.mp4",
                on_prediction=video_sink.on_prediction,
            )
            pipeline.start()
            pipeline.join()
            video_sink.release()
            ```

            `VideoFileSink` used in this way will save predictions to video file automatically.
        """
        return cls(
            video_file_name=video_file_name,
            annotator=annotator,
            display_size=display_size,
            fps_monitor=fps_monitor,
            display_statistics=display_statistics,
            output_fps=output_fps,
            quiet=quiet,
            video_frame_size=video_frame_size,
        )

    def __init__(
        self,
        video_file_name: str,
        annotator: Union[BaseAnnotator, List[BaseAnnotator]],
        display_size: Optional[Tuple[int, int]],
        fps_monitor: Optional[sv.FPSMonitor],
        display_statistics: bool,
        output_fps: int,
        quiet: bool,
        video_frame_size: Tuple[int, int],
    ):
        self._video_file_name = video_file_name
        self._annotator = annotator
        self._display_size = display_size
        self._fps_monitor = fps_monitor
        self._display_statistics = display_statistics
        self._output_fps = output_fps
        self._quiet = quiet
        self._frame_idx = 0
        self._video_frame_size = video_frame_size
        self._video_writer: Optional[cv2.VideoWriter] = None
        self.on_prediction = partial(
            render_boxes,
            annotator=self._annotator,
            display_size=self._display_size,
            fps_monitor=self._fps_monitor,
            display_statistics=self._display_statistics,
            on_frame_rendered=self._save_predictions,
        )

    def release(self) -> None:
        """
        Releases VideoWriter object.
        """
        if self._video_writer is not None and self._video_writer.isOpened():
            self._video_writer.release()

    def _save_predictions(
        self,
        frame: Union[ImageWithSourceID, List[ImageWithSourceID]],
    ) -> None:
        if self._video_writer is None:
            self._initialise_sink()
        if issubclass(type(frame), list):
            frame = create_tiles(images=[i[1] for i in frame])
        else:
            frame = frame[1]
        if (frame.shape[1], frame.shape[0]) != self._video_frame_size:
            frame = letterbox_image(image=frame, desired_size=self._video_frame_size)
        self._video_writer.write(frame)
        if not self._quiet:
            print(f"Writing frame {self._frame_idx}", end="\r")
        self._frame_idx += 1

    def _initialise_sink(self) -> None:
        self._video_writer = cv2.VideoWriter(
            self._video_file_name,
            cv2.VideoWriter_fourcc(*"MJPG"),
            self._output_fps,
            self._video_frame_size,
        )

    def __enter__(self) -> "VideoFileSink":
        return self

    def __exit__(self, exc_type, exc_val, exc_tb) -> None:
        self.release()