blakeblackshear.frigate/frigate/detectors/plugins/openvino.py

import logging
import numpy as np
import openvino.runtime as ov

from frigate.detectors.detection_api import DetectionApi
from frigate.detectors.detector_config import BaseDetectorConfig, ModelTypeEnum
from typing import Literal
from pydantic import Extra, Field


logger = logging.getLogger(__name__)

DETECTOR_KEY = "openvino"


class OvDetectorConfig(BaseDetectorConfig):
    type: Literal[DETECTOR_KEY]
    device: str = Field(default=None, title="Device Type")


class OvDetector(DetectionApi):
    type_key = DETECTOR_KEY

    def __init__(self, detector_config: OvDetectorConfig):
        self.ov_core = ov.Core()
        self.ov_model = self.ov_core.read_model(detector_config.model.path)
        self.ov_model_type = detector_config.model.model_type

        self.h = detector_config.model.height
        self.w = detector_config.model.width

        self.interpreter = self.ov_core.compile_model(
            model=self.ov_model, device_name=detector_config.device
        )

        logger.info(f"Model Input Shape: {self.interpreter.input(0).shape}")
        self.output_indexes = 0

        while True:
            try:
                tensor_shape = self.interpreter.output(self.output_indexes).shape
                logger.info(f"Model Output-{self.output_indexes} Shape: {tensor_shape}")
                self.output_indexes += 1
            except:
                logger.info(f"Model has {self.output_indexes} Output Tensors")
                break
        if self.ov_model_type == ModelTypeEnum.yolox:
            self.num_classes = tensor_shape[2] - 5
            logger.info(f"YOLOX model has {self.num_classes} classes")
            self.set_strides_grids()

    def set_strides_grids(self):
        grids = []
        expanded_strides = []

        strides = [8, 16, 32]

        hsizes = [self.h // stride for stride in strides]
        wsizes = [self.w // stride for stride in strides]

        for hsize, wsize, stride in zip(hsizes, wsizes, strides):
            xv, yv = np.meshgrid(np.arange(wsize), np.arange(hsize))
            grid = np.stack((xv, yv), 2).reshape(1, -1, 2)
            grids.append(grid)
            shape = grid.shape[:2]
            expanded_strides.append(np.full((*shape, 1), stride))
        self.grids = np.concatenate(grids, 1)
        self.expanded_strides = np.concatenate(expanded_strides, 1)

    def detect_raw(self, tensor_input):
        infer_request = self.interpreter.create_infer_request()
        infer_request.infer([tensor_input])

        if self.ov_model_type == ModelTypeEnum.ssd:
            results = infer_request.get_output_tensor()

            detections = np.zeros((20, 6), np.float32)
            i = 0
            for object_detected in results.data[0, 0, :]:
                if object_detected[0] != -1:
                    logger.debug(object_detected)
                if object_detected[2] < 0.1 or i == 20:
                    break
                detections[i] = [
                    object_detected[1],  # Label ID
                    float(object_detected[2]),  # Confidence
                    object_detected[4],  # y_min
                    object_detected[3],  # x_min
                    object_detected[6],  # y_max
                    object_detected[5],  # x_max
                ]
                i += 1
            return detections
        elif self.ov_model_type == ModelTypeEnum.yolox:
            out_tensor = infer_request.get_output_tensor()
            # [x, y, h, w, box_score, class_no_1, ..., class_no_80],
            results = out_tensor.data
            results[..., :2] = (results[..., :2] + self.grids) * self.expanded_strides
            results[..., 2:4] = np.exp(results[..., 2:4]) * self.expanded_strides
            image_pred = results[0, ...]

            class_conf = np.max(
                image_pred[:, 5 : 5 + self.num_classes], axis=1, keepdims=True
            )
            class_pred = np.argmax(image_pred[:, 5 : 5 + self.num_classes], axis=1)
            class_pred = np.expand_dims(class_pred, axis=1)

            conf_mask = (image_pred[:, 4] * class_conf.squeeze() >= 0.3).squeeze()
            # Detections ordered as (x1, y1, x2, y2, obj_conf, class_conf, class_pred)
            dets = np.concatenate((image_pred[:, :5], class_conf, class_pred), axis=1)
            dets = dets[conf_mask]

            ordered = dets[dets[:, 5].argsort()[::-1]][:20]

            detections = np.zeros((20, 6), np.float32)
            i = 0

            for object_detected in ordered:
                if i < 20:
                    detections[i] = [
                        object_detected[6],  # Label ID
                        object_detected[5],  # Confidence
                        (object_detected[1] - (object_detected[3] / 2))
                        / self.h,  # y_min
                        (object_detected[0] - (object_detected[2] / 2))
                        / self.w,  # x_min
                        (object_detected[1] + (object_detected[3] / 2))
                        / self.h,  # y_max
                        (object_detected[0] + (object_detected[2] / 2))
                        / self.w,  # x_max
                    ]
                    i += 1
                else:
                    break
            return detections