Upload 43 files

.gitattributes

@@ -41,3 +41,13 @@ Models/QCS6490/cutoff_yolov5s_int16_htp_snpe2.dlc filter=lfs diff=lfs merge=lfs
 Models/QCS6490/cutoff_yolov5s_int8_htp_snpe2.dlc filter=lfs diff=lfs merge=lfs -text
 Models/QCS8550/cutoff_yolov5s_int16_htp_snpe2.dlc filter=lfs diff=lfs merge=lfs -text
 Models/QCS8550/cutoff_yolov5s_int8_htp_snpe2.dlc filter=lfs diff=lfs merge=lfs -text
+models/8550/cutoff_yolov5s_int16_htp_snpe2/cutoff_yolov5s_int16_htp_snpe2.dlc filter=lfs diff=lfs merge=lfs -text
+models/8550/cutoff_yolov5s_int16_qnn/libcutoff_yolov5s_int16.qnn.aarch64.gcc7_5.so filter=lfs diff=lfs merge=lfs -text
+models/8550/cutoff_yolov5s_int16_qnn/libcutoff_yolov5s_int16.qnn.aarch64.gcc9_4.so filter=lfs diff=lfs merge=lfs -text
+models/8550/cutoff_yolov5s_int16_qnn/libcutoff_yolov5s_int16.qnn.aarch64.ndk.so filter=lfs diff=lfs merge=lfs -text
+models/8550/cutoff_yolov5s_int16_qnn/libcutoff_yolov5s_int16.qnn.x86.so filter=lfs diff=lfs merge=lfs -text
+models/8550/cutoff_yolov5s_int8_htp_snpe2/cutoff_yolov5s_int8_htp_snpe2.dlc filter=lfs diff=lfs merge=lfs -text
+models/8550/cutoff_yolov5s_int8_qnn/libcutoff_yolov5s_int8.qnn.aarch64.gcc7_5.so filter=lfs diff=lfs merge=lfs -text
+models/8550/cutoff_yolov5s_int8_qnn/libcutoff_yolov5s_int8.qnn.aarch64.gcc9_4.so filter=lfs diff=lfs merge=lfs -text
+models/8550/cutoff_yolov5s_int8_qnn/libcutoff_yolov5s_int8.qnn.aarch64.ndk.so filter=lfs diff=lfs merge=lfs -text
+models/8550/cutoff_yolov5s_int8_qnn/libcutoff_yolov5s_int8.qnn.x86.so filter=lfs diff=lfs merge=lfs -text

examples/python/qnn_yolov5_multi.py

@@ -0,0 +1,350 @@
+# import sys
+import os
+import time
+
+import aidlite
+import cv2
+import numpy as np
+
+OBJ_CLASS_NUM = 80
+NMS_THRESH = 0.45
+BOX_THRESH = 0.5
+MODEL_SIZE = 640
+
+OBJ_NUMB_MAX_SIZE = 64
+PROP_BOX_SIZE = (5 + OBJ_CLASS_NUM)
+STRIDE8_SIZE = (MODEL_SIZE / 8)
+STRIDE16_SIZE = (MODEL_SIZE / 16)
+STRIDE32_SIZE = (MODEL_SIZE / 32)
+
+anchors = [[10, 13, 16, 30, 33, 23],
+           [30, 61, 62, 45, 59, 119],
+           [116, 90, 156, 198, 373, 326]]
+
+coco_class = [
+    'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
+    'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 'elephant',
+    'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard',
+    'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket', 'bottle',
+    'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli',
+    'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', 'dining table', 'toilet',
+    'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', 'refrigerator',
+    'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush']
+
+
+def eqprocess(image, size1, size2):
+    h, w, _ = image.shape
+    mask = np.zeros((size1, size2, 3), dtype=np.float32)
+    scale1 = h / size1
+    scale2 = w / size2
+    if scale1 > scale2:
+        scale = scale1
+    else:
+        scale = scale2
+    img = cv2.resize(image, (int(w / scale), int(h / scale)))
+    mask[:int(h / scale), :int(w / scale), :] = img
+    return mask, scale
+
+
+def xywh2xyxy(x):
+    '''
+    Box (center x, center y, width, height) to (x1, y1, x2, y2)
+    '''
+    y = np.copy(x)
+    y[:, 0] = x[:, 0] - x[:, 2] / 2  # top left x
+    y[:, 1] = x[:, 1] - x[:, 3] / 2  # top left y
+    y[:, 2] = x[:, 0] + x[:, 2] / 2  # bottom right x
+    y[:, 3] = x[:, 1] + x[:, 3] / 2  # bottom right y
+    return y
+
+
+def xyxy2xywh(box):
+    '''
+    Box (left_top x, left_top y, right_bottom x, right_bottom y) to (left_top x, left_top y, width, height)
+    '''
+    box[:, 2:] = box[:, 2:] - box[:, :2]
+    return box
+
+
+def NMS(dets, scores, thresh):
+    '''
+    单类NMS算法
+    dets.shape = (N, 5), (left_top x, left_top y, right_bottom x, right_bottom y, Scores)
+    '''
+    x1 = dets[:, 0]
+    y1 = dets[:, 1]
+    x2 = dets[:, 2]
+    y2 = dets[:, 3]
+    areas = (y2 - y1 + 1) * (x2 - x1 + 1)
+    keep = []
+    index = scores.argsort()[::-1]
+    while index.size > 0:
+        i = index[0]  # every time the first is the biggst, and add it directly
+        keep.append(i)
+        x11 = np.maximum(x1[i], x1[index[1:]])  # calculate the points of overlap
+        y11 = np.maximum(y1[i], y1[index[1:]])
+        x22 = np.minimum(x2[i], x2[index[1:]])
+        y22 = np.minimum(y2[i], y2[index[1:]])
+        w = np.maximum(0, x22 - x11 + 1)  # the weights of overlap
+        h = np.maximum(0, y22 - y11 + 1)  # the height of overlap
+        overlaps = w * h
+        ious = overlaps / (areas[i] + areas[index[1:]] - overlaps)
+        idx = np.where(ious <= thresh)[0]
+        index = index[idx + 1]  # because index start from 1
+
+    return keep
+
+
+def clip_coords(boxes, img_shape):
+    # Clip bounding xyxy bounding boxes to image shape (height, width)
+    boxes[:, 0].clip(0, img_shape[1], out=boxes[:, 0])  # x1
+    boxes[:, 1].clip(0, img_shape[0], out=boxes[:, 1])  # y1
+    boxes[:, 2].clip(0, img_shape[1], out=boxes[:, 2])  # x2
+    boxes[:, 3].clip(0, img_shape[0], out=boxes[:, 3])  # y2
+
+
+def detect_postprocess(prediction, img0shape, img1shape, conf_thres=0.25, iou_thres=0.45):
+    '''
+    检测输出后处理
+    prediction: aidlite模型预测输出
+    img0shape: 原始图片shape
+    img1shape: 输入图片shape
+    conf_thres: 置信度阈值
+    iou_thres: IOU阈值
+    return: list[np.ndarray(N, 5)], 对应类别的坐标框信息, xywh、conf
+    '''
+    h, w, _ = img1shape
+    valid_condidates = prediction[prediction[..., 4] > conf_thres]
+    valid_condidates[:, 5:] *= valid_condidates[:, 4:5]
+    valid_condidates[:, :4] = xywh2xyxy(valid_condidates[:, :4])
+
+    max_det = 300
+    max_wh = 7680
+    max_nms = 30000
+    valid_condidates[:, 4] = valid_condidates[:, 5:].max(1)
+    valid_condidates[:, 5] = valid_condidates[:, 5:].argmax(1)
+    sort_id = np.argsort(valid_condidates[:, 4])[::-1]
+    valid_condidates = valid_condidates[sort_id[:max_nms]]
+    boxes, scores = valid_condidates[:, :4] + valid_condidates[:, 5:6] * max_wh, valid_condidates[:, 4]
+    index = NMS(boxes, scores, iou_thres)[:max_det]
+    out_boxes = valid_condidates[index]
+    clip_coords(out_boxes[:, :4], img0shape)
+    out_boxes[:, :4] = xyxy2xywh(out_boxes[:, :4])
+    print("检测到{}个区域".format(len(out_boxes)))
+    return out_boxes
+
+
+def draw_detect_res(img, det_pred):
+    '''
+    检测结果绘制
+    '''
+    img = img.astype(np.uint8)
+    color_step = int(255 / len(coco_class))
+    for i in range(len(det_pred)):
+        x1, y1, x2, y2 = [int(t) for t in det_pred[i][:4]]
+        score = det_pred[i][4]
+        cls_id = int(det_pred[i][5])
+
+        print(i + 1, [x1, y1, x2, y2], score, coco_class[cls_id])
+
+        cv2.putText(img, f'{coco_class[cls_id]}', (x1, y1 - 6), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255), 1)
+        cv2.rectangle(img, (x1, y1), (x2 + x1, y2 + y1), (0, int(cls_id * color_step), int(255 - cls_id * color_step)),
+                      thickness=2)
+
+    return img
+
+
+class Detect():
+    # YOLOv5 Detect head for detection models
+    def __init__(self, nc=80, anchors=(), stride=[], image_size=640):  # detection layer
+        super().__init__()
+        self.nc = nc  # number of classes
+        self.no = nc + 5  # number of outputs per anchor
+        self.stride = stride
+        self.nl = len(anchors)  # number of detection layers
+        self.na = len(anchors[0]) // 2  # number of anchors
+        self.grid, self.anchor_grid = [0] * self.nl, [0] * self.nl
+        self.anchors = np.array(anchors, dtype=np.float32).reshape(self.nl, -1, 2)
+
+        base_scale = image_size // 8
+        for i in range(self.nl):
+            self.grid[i], self.anchor_grid[i] = self._make_grid(base_scale // (2 ** i), base_scale // (2 ** i), i)
+
+    def _make_grid(self, nx=20, ny=20, i=0):
+        y, x = np.arange(ny, dtype=np.float32), np.arange(nx, dtype=np.float32)
+        yv, xv = np.meshgrid(y, x)
+        yv, xv = yv.T, xv.T
+        # add grid offset, i.e. y = 2.0 * x - 0.5
+        grid = np.stack((xv, yv), 2)
+        grid = grid[np.newaxis, np.newaxis, ...]
+        grid = np.repeat(grid, self.na, axis=1) - 0.5
+        anchor_grid = self.anchors[i].reshape((1, self.na, 1, 1, 2))
+        anchor_grid = np.repeat(anchor_grid, repeats=ny, axis=2)
+        anchor_grid = np.repeat(anchor_grid, repeats=nx, axis=3)
+        return grid, anchor_grid
+
+    def sigmoid(self, arr):
+        return 1 / (1 + np.exp(-arr))
+
+    def __call__(self, x):
+        z = []  # inference output
+        for i in range(self.nl):
+            bs, _, ny, nx = x[i].shape
+            x[i] = x[i].reshape(bs, self.na, self.no, ny, nx).transpose(0, 1, 3, 4, 2)
+            y = self.sigmoid(x[i])
+            y[..., 0:2] = (y[..., 0:2] * 2. + self.grid[i]) * self.stride[i]  # xy
+            y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+            z.append(y.reshape(bs, self.na * nx * ny, self.no))
+
+        return np.concatenate(z, 1)
+
+
+def main():
+    # argvs = len(sys.argv)
+    # if argvs < 2:
+    #     print("python3 ./qnn_yolov5_multi.py 0 for use DSP")
+    #     print("python3 ./qnn_yolov5_multi.py 1 for use CPU")
+    #     print("python3 ./qnn_yolov5_multi.py 2 for use GPU")
+    #     return False
+    # acc_type = int(sys.argv[1])
+
+    print("Start main ... ...")
+    aidlite.set_log_level(aidlite.LogLevel.INFO)
+    aidlite.log_to_stderr()
+    print(f"Aidlite library version : {aidlite.get_library_version()}")
+    print(f"Aidlite python library version : {aidlite.get_py_library_version()}")
+
+    config = aidlite.Config.create_instance()
+    if config is None:
+        print("Create config failed !")
+        return False
+
+    # if acc_type == 1:
+    #     config.accelerate_type = aidlite.AccelerateType.TYPE_CPU
+    # elif acc_type == 2:
+    #     config.accelerate_type = aidlite.AccelerateType.TYPE_GPU
+    #     config.is_quantify_model = 0
+    # elif acc_type == 0:
+    #     config.accelerate_type = aidlite.AccelerateType.TYPE_DSP
+    #     config.is_quantify_model = 1
+    # else:
+    #     print("python3 ./qnn_yolov5_multi.py 0 for use DSP")
+    #     print("python3 ./qnn_yolov5_multi.py 1 for use CPU")
+    #     print("python3 ./qnn_yolov5_multi.py 2 for use GPU")
+    #     return False
+
+    config.implement_type = aidlite.ImplementType.TYPE_LOCAL
+    config.framework_type = aidlite.FrameworkType.TYPE_QNN
+    config.accelerate_type = aidlite.AccelerateType.TYPE_DSP
+    config.is_quantify_model = 1
+
+    model_path = "/home/aidlux/modelzoo/examples/data/cutoff_yolov5s_int8_qnn/cutoff_yolov5s_int8.qnn.serialized.bin"
+    model = aidlite.Model.create_instance(model_path)
+    if model is None:
+        print("Create model failed !")
+        return False
+    input_shapes = [[1, MODEL_SIZE, MODEL_SIZE, 3]]
+    output_shapes = [[1, 20, 20, 255], [1, 40, 40, 255], [1, 80, 80, 255]]
+    model.set_model_properties(input_shapes, aidlite.DataType.TYPE_FLOAT32,
+                               output_shapes, aidlite.DataType.TYPE_FLOAT32)
+
+    interpreter = aidlite.InterpreterBuilder.build_interpretper_from_model_and_config(model, config)
+    if interpreter is None:
+        print("build_interpretper_from_model_and_config failed !")
+        return None
+    result = interpreter.init()
+    if result != 0:
+        print(f"interpreter init failed !")
+        return False
+    result = interpreter.load_model()
+    if result != 0:
+        print("interpreter load model failed !")
+        return False
+    print("detect model load success!")
+    stride8 = stride16 = stride32 = None
+
+    image_path = os.path.join("../data/bus.jpg")
+    frame = cv2.imread(image_path)
+    frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+    img_input, scale = eqprocess(frame, MODEL_SIZE, MODEL_SIZE)
+    img_input = img_input / 255
+    img_input = img_input.astype(np.float32)
+
+    sum_time_0 = 0.0
+    sum_time_1 = 0.0
+    sum_time_2 = 0.0
+    _counter = 1
+    for idx in range(_counter):
+        st0 = time.time()
+        input_tensor_data = img_input.data
+        result = interpreter.set_input_tensor(0, input_tensor_data)
+        if result != 0:
+            print("interpreter set_input_tensor() failed")
+            return False
+        et0 = time.time()
+        dur0 = (et0 - st0) * 1000
+        sum_time_0 += dur0
+        print(f"current [{idx}] set_input_tensor cost time :{dur0} ms")
+
+        st1 = time.time()
+        result = interpreter.invoke()
+        if result != 0:
+            print("interpreter set_input_tensor() failed")
+            return False
+        et1 = time.time()
+        dur1 = (et1 - st1) * 1000
+        sum_time_1 += dur1
+        print(f"current [{idx}] invoke cost time :{dur1} ms")
+
+        st2 = time.time()
+
+        stride8 = interpreter.get_output_tensor(0)
+        if stride8 is None:
+            print("sample : interpreter->get_output_tensor() 0 failed !")
+            return False
+        print(f"len(stride8 {len(stride8)}")
+
+        stride16 = interpreter.get_output_tensor(1)
+        if stride16 is None:
+            print("sample : interpreter->get_output_tensor() 2 failed !")
+            return False
+        print(f"len(stride16 {len(stride16)}")
+
+        stride32 = interpreter.get_output_tensor(2)
+        if stride32 is None:
+            print("sample : interpreter->get_output_tensor() 1 failed !")
+            return False
+        print(f"len(stride32 {len(stride32)}")
+        et2 = time.time()
+        dur2 = (et2 - st2) * 1000
+        sum_time_2 += dur2
+        print(f"current [{idx}] get_output_tensor cost time :{dur2} ms")
+
+    print(
+        f"repeat [{_counter}] times , input[{sum_time_0 * 1000}]ms --- invoke[{sum_time_1 * 1000}]ms --- output[{sum_time_2 * 1000}]ms --- sum[{(sum_time_0 + sum_time_1 + sum_time_2) * 1000}]ms")
+
+    stride = [8, 16, 32]
+    yolo_head = Detect(OBJ_CLASS_NUM, anchors, stride, MODEL_SIZE)
+    # 后处理部分reshape需要知道模型的output_shapes
+    validCount0 = stride8.reshape(*output_shapes[2]).transpose(0, 3, 1, 2)
+    validCount1 = stride16.reshape(*output_shapes[1]).transpose(0, 3, 1, 2)
+    validCount2 = stride32.reshape(*output_shapes[0]).transpose(0, 3, 1, 2)
+    pred = yolo_head([validCount0, validCount1, validCount2])
+    det_pred = detect_postprocess(pred, frame.shape, [MODEL_SIZE, MODEL_SIZE, 3], conf_thres=0.5, iou_thres=0.45)
+    det_pred[np.isnan(det_pred)] = 0.0
+    det_pred[:, :4] = det_pred[:, :4] * scale
+    res_img = draw_detect_res(frame, det_pred)
+
+    result = interpreter.destory()
+    if result != 0:
+        print(f"interpreter set_input_tensor() failed")
+        return False
+    frame_bgr = cv2.cvtColor(res_img, cv2.COLOR_RGB2BGR)
+    result_img_path = f"{os.path.splitext(os.path.abspath(__file__))[0]}.jpg"
+    cv2.imwrite(result_img_path, frame_bgr)
+    print(f"The result image has been saved to : {result_img_path}")
+    return True
+
+
+if __name__ == "__main__":
+    main()

examples/python/snpe2_yolov5_multi.py

@@ -0,0 +1,338 @@
+import os
+import struct
+import sys
+import time
+
+import aidlite
+import cv2
+import numpy as np
+
+FLOAT_SIZE = struct.calcsize('f')
+OBJ_CLASS_NUM = 80
+MODEL_SIZE = 640
+anchors = [[10, 13, 16, 30, 33, 23],
+           [30, 61, 62, 45, 59, 119],
+           [116, 90, 156, 198, 373, 326]]
+
+coco_class = [
+    'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
+    'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 'elephant',
+    'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard',
+    'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket', 'bottle',
+    'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli',
+    'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', 'dining table', 'toilet',
+    'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', 'refrigerator',
+    'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush']
+
+
+def eqprocess(image, size1, size2):
+    h, w, _ = image.shape
+    mask = np.zeros((size1, size2, 3), dtype=np.float32)
+    scale1 = h / size1
+    scale2 = w / size2
+    if scale1 > scale2:
+        scale = scale1
+    else:
+        scale = scale2
+    img = cv2.resize(image, (int(w / scale), int(h / scale)))
+    mask[:int(h / scale), :int(w / scale), :] = img
+    return mask, scale
+
+
+def xywh2xyxy(x):
+    '''
+    Box (center x, center y, width, height) to (x1, y1, x2, y2)
+    '''
+    y = np.copy(x)
+    y[:, 0] = x[:, 0] - x[:, 2] / 2  # top left x
+    y[:, 1] = x[:, 1] - x[:, 3] / 2  # top left y
+    y[:, 2] = x[:, 0] + x[:, 2] / 2  # bottom right x
+    y[:, 3] = x[:, 1] + x[:, 3] / 2  # bottom right y
+    return y
+
+
+def xyxy2xywh(box):
+    '''
+    Box (left_top x, left_top y, right_bottom x, right_bottom y) to (left_top x, left_top y, width, height)
+    '''
+    box[:, 2:] = box[:, 2:] - box[:, :2]
+    return box
+
+
+def NMS(dets, scores, thresh):
+    '''
+    单类NMS算法
+    dets.shape = (N, 5), (left_top x, left_top y, right_bottom x, right_bottom y, Scores)
+    '''
+    x1 = dets[:, 0]
+    y1 = dets[:, 1]
+    x2 = dets[:, 2]
+    y2 = dets[:, 3]
+    areas = (y2 - y1 + 1) * (x2 - x1 + 1)
+    keep = []
+    index = scores.argsort()[::-1]
+    while index.size > 0:
+        i = index[0]  # every time the first is the biggst, and add it directly
+        keep.append(i)
+        x11 = np.maximum(x1[i], x1[index[1:]])  # calculate the points of overlap
+        y11 = np.maximum(y1[i], y1[index[1:]])
+        x22 = np.minimum(x2[i], x2[index[1:]])
+        y22 = np.minimum(y2[i], y2[index[1:]])
+        w = np.maximum(0, x22 - x11 + 1)  # the weights of overlap
+        h = np.maximum(0, y22 - y11 + 1)  # the height of overlap
+        overlaps = w * h
+        ious = overlaps / (areas[i] + areas[index[1:]] - overlaps)
+        idx = np.where(ious <= thresh)[0]
+        index = index[idx + 1]  # because index start from 1
+
+    return keep
+
+
+def clip_coords(boxes, img_shape):
+    # Clip bounding xyxy bounding boxes to image shape (height, width)
+    boxes[:, 0].clip(0, img_shape[1], out=boxes[:, 0])  # x1
+    boxes[:, 1].clip(0, img_shape[0], out=boxes[:, 1])  # y1
+    boxes[:, 2].clip(0, img_shape[1], out=boxes[:, 2])  # x2
+    boxes[:, 3].clip(0, img_shape[0], out=boxes[:, 3])  # y2
+
+
+def detect_postprocess(prediction, img0shape, img1shape, conf_thres=0.25, iou_thres=0.45):
+    '''
+    检测输出后处理
+    prediction: aidlite模型预测输出
+    img0shape: 原始图片shape
+    img1shape: 输入图片shape
+    conf_thres: 置信度阈值
+    iou_thres: IOU阈值
+    return: list[np.ndarray(N, 5)], 对应类别的坐标框信息, xywh、conf
+    '''
+    h, w, _ = img1shape
+    valid_condidates = prediction[prediction[..., 4] > conf_thres]
+    valid_condidates[:, 5:] *= valid_condidates[:, 4:5]
+    valid_condidates[:, :4] = xywh2xyxy(valid_condidates[:, :4])
+
+    max_det = 300
+    max_wh = 7680
+    max_nms = 30000
+    valid_condidates[:, 4] = valid_condidates[:, 5:].max(1)
+    valid_condidates[:, 5] = valid_condidates[:, 5:].argmax(1)
+    sort_id = np.argsort(valid_condidates[:, 4])[::-1]
+    valid_condidates = valid_condidates[sort_id[:max_nms]]
+    boxes, scores = valid_condidates[:, :4] + valid_condidates[:, 5:6] * max_wh, valid_condidates[:, 4]
+    index = NMS(boxes, scores, iou_thres)[:max_det]
+    out_boxes = valid_condidates[index]
+    clip_coords(out_boxes[:, :4], img0shape)
+    out_boxes[:, :4] = xyxy2xywh(out_boxes[:, :4])
+    print("检测到{}个区域".format(len(out_boxes)))
+    return out_boxes
+
+
+def draw_detect_res(img, det_pred):
+    '''
+    检测结果绘制
+    '''
+    img = img.astype(np.uint8)
+    color_step = int(255 / len(coco_class))
+    for i in range(len(det_pred)):
+        x1, y1, x2, y2 = [int(t) for t in det_pred[i][:4]]
+        score = det_pred[i][4]
+        cls_id = int(det_pred[i][5])
+
+        print(i + 1, [x1, y1, x2, y2], score, coco_class[cls_id])
+
+        cv2.putText(img, f'{coco_class[cls_id]}', (x1, y1 - 6), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255), 1)
+        cv2.rectangle(img, (x1, y1), (x2 + x1, y2 + y1), (0, int(cls_id * color_step), int(255 - cls_id * color_step)),
+                      thickness=2)
+
+    return img
+
+
+class Detect():
+    # YOLOv5 Detect head for detection models
+    def __init__(self, nc=80, anchors=(), stride=[], image_size=640):  # detection layer
+        super().__init__()
+        self.nc = nc  # number of classes
+        self.no = nc + 5  # number of outputs per anchor
+        self.stride = stride
+        self.nl = len(anchors)  # number of detection layers
+        self.na = len(anchors[0]) // 2  # number of anchors
+        self.grid, self.anchor_grid = [0] * self.nl, [0] * self.nl
+        self.anchors = np.array(anchors, dtype=np.float32).reshape(self.nl, -1, 2)
+
+        base_scale = image_size // 8
+        for i in range(self.nl):
+            self.grid[i], self.anchor_grid[i] = self._make_grid(base_scale // (2 ** i), base_scale // (2 ** i), i)
+
+    def _make_grid(self, nx=20, ny=20, i=0):
+        y, x = np.arange(ny, dtype=np.float32), np.arange(nx, dtype=np.float32)
+        yv, xv = np.meshgrid(y, x)
+        yv, xv = yv.T, xv.T
+        # add grid offset, i.e. y = 2.0 * x - 0.5
+        grid = np.stack((xv, yv), 2)
+        grid = grid[np.newaxis, np.newaxis, ...]
+        grid = np.repeat(grid, self.na, axis=1) - 0.5
+        anchor_grid = self.anchors[i].reshape((1, self.na, 1, 1, 2))
+        anchor_grid = np.repeat(anchor_grid, repeats=ny, axis=2)
+        anchor_grid = np.repeat(anchor_grid, repeats=nx, axis=3)
+        return grid, anchor_grid
+
+    def sigmoid(self, arr):
+        return 1 / (1 + np.exp(-arr))
+
+    def __call__(self, x):
+        z = []  # inference output
+        for i in range(self.nl):
+            bs, _, ny, nx = x[i].shape
+            x[i] = x[i].reshape(bs, self.na, self.no, ny, nx).transpose(0, 1, 3, 4, 2)
+            y = self.sigmoid(x[i])
+            y[..., 0:2] = (y[..., 0:2] * 2. + self.grid[i]) * self.stride[i]  # xy
+            y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+            z.append(y.reshape(bs, self.na * nx * ny, self.no))
+
+        return np.concatenate(z, 1)
+
+
+def main():
+    acc_type = 0
+    aidlite.set_log_level(aidlite.LogLevel.INFO)
+    aidlite.log_to_stderr()
+
+    # image process
+    image_path = r"../data/bus.jpg"
+    frame = cv2.imread(image_path)
+    if frame is None:
+        print(f"{image_path} not exists")
+        return False
+    frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+    img_input, scale = eqprocess(frame, MODEL_SIZE, MODEL_SIZE)
+    img_input = img_input / 255
+    img_input = img_input.astype(np.float32)
+
+    config = aidlite.Config.create_instance()
+    if config is None:
+        print("build_interpretper_from_model_and_config failed !")
+        return False
+
+    # model init
+    acc_type_dsp = True
+    model_path = r"../data/cutoff_yolov5s_int8_htp_snpe2/cutoff_yolov5s_int8_htp_snpe2.dlc"
+    outNode_names = "/model.24/m.1/Conv,/model.24/m.2/Conv,/model.24/m.0/Conv"
+    config.accelerate_type = aidlite.AccelerateType.TYPE_DSP
+    config.is_quantify_model = 1
+
+
+    config.implement_type = aidlite.ImplementType.TYPE_LOCAL
+    config.framework_type = aidlite.FrameworkType.TYPE_SNPE2
+    config.snpe_out_names = outNode_names.split(',')
+    print("---------------------------")
+    print(config.snpe_out_names)
+
+    model = aidlite.Model.create_instance(model_path)
+    if model is None:
+        print("Create model failed !")
+        return False
+
+    input_shapes = [[1, MODEL_SIZE, MODEL_SIZE, 3]]
+    output_shapes = [[1, 40, 40, 255], [1, 20, 20, 255], [1, 80, 80, 255]]
+    model.set_model_properties(input_shapes, aidlite.DataType.TYPE_FLOAT32, output_shapes,
+                               aidlite.DataType.TYPE_FLOAT32)
+
+    snpe2_interpreter = aidlite.InterpreterBuilder.build_interpretper_from_model_and_config(model, config)
+    if snpe2_interpreter is None:
+        print("build_interpretper_from_model_and_config failed !")
+        return None
+    result = snpe2_interpreter.init()
+    if result != 0:
+        print(f"interpreter init failed !")
+        return False
+    result = snpe2_interpreter.load_model()
+    if result != 0:
+        print("interpreter load model failed !")
+        return False
+    print("detect model load success!")
+
+    stride8 = stride16 = stride32 = None
+    sum_time_0 = 0.0
+    sum_time_1 = 0.0
+    sum_time_2 = 0.0
+    _counter = 1
+    for idx in range(_counter):
+        st0 = time.time()
+
+        if acc_type_dsp:
+            time.sleep(0.04)  # 40milliseconds
+
+        input_tensor_data = img_input.data
+        result = snpe2_interpreter.set_input_tensor(0, input_tensor_data)
+        if result != 0:
+            print("interpreter set_input_tensor() failed")
+            return False
+        et0 = time.time()
+        dur0 = (et0 - st0) * 1000
+        sum_time_0 += dur0
+        print(f"current [{idx}] set_input_tensor cost time :{dur0} ms")
+
+        st1 = time.time()
+        result = snpe2_interpreter.invoke()
+        if result != 0:
+            print("interpreter set_input_tensor() failed")
+            return False
+        et1 = time.time()
+        dur1 = (et1 - st1) * 1000
+        sum_time_1 += dur1
+        print(f"current [{idx}] invoke cost time :{dur1} ms")
+
+        st2 = time.time()
+
+        if acc_type_dsp:
+            time.sleep(0.02)  # 40milliseconds
+
+        stride8 = snpe2_interpreter.get_output_tensor(0)
+        if stride8 is None:
+            print("sample : interpreter->get_output_tensor() 0 failed !")
+            return False
+        print(f"len(stride8 {len(stride8)}")
+
+        stride16 = snpe2_interpreter.get_output_tensor(1)
+        if stride16 is None:
+            print("sample : interpreter->get_output_tensor() 2 failed !")
+            return False
+        print(f"len(stride16 {len(stride16)}")
+
+        stride32 = snpe2_interpreter.get_output_tensor(2)
+        if stride32 is None:
+            print("sample : interpreter->get_output_tensor() 1 failed !")
+            return False
+        print(f"len(stride32 {len(stride32)}")
+        et2 = time.time()
+        dur2 = (et2 - st2) * 1000
+        sum_time_2 += dur2
+        print(f"current [{idx}] get_output_tensor cost time :{dur2} ms")
+
+    print(
+        f"repeat [{_counter}] time , input[{sum_time_0}]ms --- invoke[{sum_time_1}]ms --- output[{sum_time_2}]ms --- sum[{sum_time_0 + sum_time_1 + sum_time_2}]ms")
+
+    stride = [8, 16, 32]
+    yolo_head = Detect(OBJ_CLASS_NUM, anchors, stride, MODEL_SIZE)
+    validCount0 = stride8.reshape(*output_shapes[2]).transpose(0, 3, 1, 2)
+    validCount1 = stride16.reshape(*output_shapes[0]).transpose(0, 3, 1, 2)
+    validCount2 = stride32.reshape(*output_shapes[1]).transpose(0, 3, 1, 2)
+    pred = yolo_head([validCount0, validCount1, validCount2])
+    det_pred = detect_postprocess(pred, frame.shape, [MODEL_SIZE, MODEL_SIZE, 3], conf_thres=0.5, iou_thres=0.45)
+    det_pred[np.isnan(det_pred)] = 0.0
+    det_pred[:, :4] = det_pred[:, :4] * scale
+    res_img = draw_detect_res(frame, det_pred)
+
+    result = snpe2_interpreter.destory()
+    if result != 0:
+        print(f"interpreter set_input_tensor() failed")
+        return False
+    frame_bgr = cv2.cvtColor(res_img, cv2.COLOR_RGB2BGR)
+    result_img_path = f"{os.path.splitext(os.path.abspath(__file__))[0]}.jpg"
+    cv2.imwrite(result_img_path, frame_bgr)
+    print(f"The result image has been saved to : {result_img_path}")
+    return True
+
+
+if __name__ == "__main__":
+    main()

models/8550/cutoff_yolov5s_int16_htp_snpe2/cutoff_yolov5s.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f3268665cee28d5932e69a42d32ea1585cd85bc2438751c2df92e66c53fbbf7
+size 28953517

models/8550/cutoff_yolov5s_int16_htp_snpe2/cutoff_yolov5s_int16_htp_snpe2.dlc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d689002a3e27c63edc944f474f25e2c331c05ca6b1ab4f85a260743f18e9648f
+size 15175256

models/8550/cutoff_yolov5s_int16_htp_snpe2/shell_info.txt

@@ -0,0 +1,9 @@
+2024-07-08 17:21:00,089 - INFO : Optimization started.
+2024-07-08 17:21:00,090 - INFO : [ONNX-SIM] Clean ONNX Model input node.
+2024-07-08 17:21:01,319 - INFO : [CUTOFF] Cutoff model...
+2024-07-08 17:21:07,597 - INFO : [ONNX2DLC] Start converting to DLC, expected to take 9 s.
+2024-07-08 17:21:10,976 - INFO : [ONNX2DLC] DLC conversion is done.
+2024-07-08 17:21:10,976 - INFO : [DLC-QUAN] Start quantizing model, expected to take 70 s.
+2024-07-08 17:21:17,875 - INFO : [DLC-QUAN] quantization is done.
+2024-07-08 17:21:19,220 - INFO : [ONNX2DLC] DLC conversion is done.
+2024-07-08 17:21:19,220 - INFO : Model optimization done.

models/8550/cutoff_yolov5s_int16_qnn/cutoff_yolov5s.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:202ad25e2ad38b3889926aa5ecb9144d0ce10b34756d23f61a3d6e430acc269f
+size 7331840

models/8550/cutoff_yolov5s_int16_qnn/cutoff_yolov5s.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f3268665cee28d5932e69a42d32ea1585cd85bc2438751c2df92e66c53fbbf7
+size 28953517

models/8550/cutoff_yolov5s_int16_qnn/cutoff_yolov5s_int16.qnn.serialized.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:122a955a4b0bc1c3f553e6102a034601e0f8b290e79599c8ed4f9099711f8846
+size 7645656

models/8550/cutoff_yolov5s_int16_qnn/libcutoff_yolov5s_int16.qnn.aarch64.gcc7_5.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:214eade70f7ed41ec4addc2fb9356b19aea0d6c1d354a83674b5f705a2ebe8d3
+size 7608368

models/8550/cutoff_yolov5s_int16_qnn/libcutoff_yolov5s_int16.qnn.aarch64.gcc9_4.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:847636599d0cc6c51e2252b97177ffd58a7355cb21d113545095391459f0329f
+size 7604360

models/8550/cutoff_yolov5s_int16_qnn/libcutoff_yolov5s_int16.qnn.aarch64.ndk.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb91df7035d426bbf8addc2859598d28326ddb5411018b54f66718e513e23af9
+size 7870472

models/8550/cutoff_yolov5s_int16_qnn/libcutoff_yolov5s_int16.qnn.x86.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a64c0b874c5f4330b874455320f7e63def71674c88ec7278da336b43e6388d5f
+size 7660448

models/8550/cutoff_yolov5s_int16_qnn/qnn_model_info.json

@@ -0,0 +1 @@
+{"model_name": "cutoff_yolov5s_int16.qnn.serialized.bin", "data_type": "w8a16", "inputDimensions": {"images": [1, 640, 640, 3]}, "outputDimensions": {}, "htp_socs_name": "QCS8550", "htp_socs_type": "sm8550", "qnn_sdk": "qnn_sdk_2.16.4.231110"}

models/8550/cutoff_yolov5s_int16_qnn/shell_info.txt

@@ -0,0 +1,9 @@
+2024-07-08 18:22:25,075 - INFO : Optimization started.
+2024-07-08 18:22:25,076 - INFO : [ONNX-SIM] Clean ONNX Model input node.
+2024-07-08 18:22:26,130 - INFO : [CUTOFF] Cutoff model...
+2024-07-08 18:22:28,402 - INFO : [ONNX2QNN] Start converting to QNN.
+2024-07-08 18:24:11,047 - INFO : [CONVERT-QNN] Convert model is done.
+2024-07-08 18:24:31,125 - INFO : [COMPILE-QNN] Compile model is done.
+2024-07-08 18:24:33,089 - INFO : [CREATE-CONTEXT-BINARY] Create context binary is done.
+2024-07-08 18:24:33,291 - INFO : [ONNX2QNN] QNN conversion is done.
+2024-07-08 18:24:33,292 - INFO : Model optimization done.

models/8550/cutoff_yolov5s_int8_htp_snpe2/cutoff_yolov5s.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f3268665cee28d5932e69a42d32ea1585cd85bc2438751c2df92e66c53fbbf7
+size 28953517

models/8550/cutoff_yolov5s_int8_htp_snpe2/cutoff_yolov5s_int8_htp_snpe2.dlc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:934d69f4e315f4a011350072d1b0206efdb505ecae051fa5dbb3581f87907e85
+size 14956262

models/8550/cutoff_yolov5s_int8_htp_snpe2/shell_info.txt

@@ -0,0 +1,9 @@
+2024-07-08 16:25:41,225 - INFO : Optimization started.
+2024-07-08 16:25:41,225 - INFO : [ONNX-SIM] Clean ONNX Model input node.
+2024-07-08 16:25:43,400 - INFO : [CUTOFF] Cutoff model...
+2024-07-08 16:25:49,055 - INFO : [ONNX2DLC] Start converting to DLC, expected to take 11 s.
+2024-07-08 16:25:52,205 - INFO : [ONNX2DLC] DLC conversion is done.
+2024-07-08 16:25:52,205 - INFO : [DLC-QUAN] Start quantizing model, expected to take 87 s.
+2024-07-08 16:25:58,803 - INFO : [DLC-QUAN] quantization is done.
+2024-07-08 16:26:00,104 - INFO : [ONNX2DLC] DLC conversion is done.
+2024-07-08 16:26:00,105 - INFO : Model optimization done.

models/8550/cutoff_yolov5s_int8_qnn/cutoff_yolov5s.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:407d11ec97f758a43ee56926e755526701fc8356fb43225f8b01eaaf8f314cbe
+size 7311360

models/8550/cutoff_yolov5s_int8_qnn/cutoff_yolov5s.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f3268665cee28d5932e69a42d32ea1585cd85bc2438751c2df92e66c53fbbf7
+size 28953517

models/8550/cutoff_yolov5s_int8_qnn/cutoff_yolov5s_int8.qnn.serialized.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:603bad2305c9c2a2c9ce5678eb2919867ae5157502dbe190f370e4c0085482f0
+size 7530968

models/8550/cutoff_yolov5s_int8_qnn/libcutoff_yolov5s_int8.qnn.aarch64.gcc7_5.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cdd4984b91d23bf7c0334dfcc0bfd3287cb847c53f848949821ac51101e6bba
+size 7573464

models/8550/cutoff_yolov5s_int8_qnn/libcutoff_yolov5s_int8.qnn.aarch64.gcc9_4.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3248baf6bafb88c338c628e4150f9f04b26a2efdc17e55a0962d2eeffbe0fad
+size 7581744

models/8550/cutoff_yolov5s_int8_qnn/libcutoff_yolov5s_int8.qnn.aarch64.ndk.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f7d40d7a95f2ef375ae034ebb6f792a7e7916b99e6315980dfda4485645dc26
+size 7839680

models/8550/cutoff_yolov5s_int8_qnn/libcutoff_yolov5s_int8.qnn.x86.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fa4f0995efcd5d5c415775be5059a1b28ecf248d27bc81b8dd2e6533bec3e18
+size 7633744

models/8550/cutoff_yolov5s_int8_qnn/qnn_model_info.json

@@ -0,0 +1 @@
+{"model_name": "cutoff_yolov5s_int8.qnn.serialized.bin", "data_type": "w8a8", "inputDimensions": {"images": [1, 640, 640, 3]}, "outputDimensions": {}, "htp_socs_name": "QCS8550", "htp_socs_type": "sm8550", "qnn_sdk": "qnn_sdk_2.16.4.231110"}

models/8550/cutoff_yolov5s_int8_qnn/shell_info.txt

@@ -0,0 +1,9 @@
+2024-07-08 18:07:31,799 - INFO : Optimization started.
+2024-07-08 18:07:31,799 - INFO : [ONNX-SIM] Clean ONNX Model input node.
+2024-07-08 18:07:33,343 - INFO : [CUTOFF] Cutoff model...
+2024-07-08 18:07:35,697 - INFO : [ONNX2QNN] Start converting to QNN.
+2024-07-08 18:09:19,450 - INFO : [CONVERT-QNN] Convert model is done.
+2024-07-08 18:09:39,385 - INFO : [COMPILE-QNN] Compile model is done.
+2024-07-08 18:09:41,192 - INFO : [CREATE-CONTEXT-BINARY] Create context binary is done.
+2024-07-08 18:09:41,388 - INFO : [ONNX2QNN] QNN conversion is done.
+2024-07-08 18:09:41,388 - INFO : Model optimization done.