Toybrick

RV1126推理时间异常

blue

注册会员

积分
51
发表于 2022-5-21 11:44:29    查看: 313|回复: 2 | [复制链接]    打印 | 显示全部楼层
我现在是在rv1126上运行YOLOv5的demo,有时候会遇到推理时间过长达到20S的问题,有时候就正常运行,时间是70ms,我的rknn-toolkit-lite是1.7.1,npu也对应是最新的,这个问题是怎么回事呢,万分感谢!
回复

使用道具 举报

blue

注册会员

积分
51
 楼主| 发表于 2022-5-21 11:45:11 | 显示全部楼层

"""将RKNN模型部署在RV1126平台并运行测试,目前只是简单推理"""
#环境:rv1126(Python3.7)
import time
import cv2
import numpy as np
from rknnlite.api import RKNNLite

RKNN_MODEL = 'best1.rknn'
IMG_PATH = 'bus.jpg'  # 修改推理图片

QUANTIZE_ON = True

BOX_THRESH = 0.25
NMS_THRESH = 0.45
IMG_SIZE = 640
# CLASSES = ("1")
# CLASSES = ("pedestrians","crowd","partially","ignore","riders")
# CLASSES = ("person", "bicycle", "car","motorbike ","aeroplane ","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","sofa",
#            "pottedplant","bed","diningtable","toilet ","tvmonitor","laptop","mouse","remote ","keyboard ","cell phone","microwave ",
#            "oven ","toaster","sink","refrigerator ","book","clock","vase","scissors ","teddy bear ","hair drier", "toothbrush ")
CLASSES = ('aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog',
        'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor')
def sigmoid(x):
    return 1 / (1 + np.exp(-x))


def xywh2xyxy(x):
    # Convert [x, y, w, h] 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 process(input, mask, anchors):
    anchors = [anchors[i] for i in mask]
    grid_h, grid_w = map(int, input.shape[0:2])

    box_confidence = sigmoid(input[..., 4])
    box_confidence = np.expand_dims(box_confidence, axis=-1)

    box_class_probs = sigmoid(input[..., 5:])

    box_xy = sigmoid(input[..., :2]) * 2 - 0.5

    col = np.tile(np.arange(0, grid_w), grid_w).reshape(-1, grid_w)
    row = np.tile(np.arange(0, grid_h).reshape(-1, 1), grid_h)
    col = col.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2)
    row = row.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2)
    grid = np.concatenate((col, row), axis=-1)
    box_xy += grid
    box_xy *= int(IMG_SIZE / grid_h)

    box_wh = pow(sigmoid(input[..., 2:4]) * 2, 2)
    box_wh = box_wh * anchors

    box = np.concatenate((box_xy, box_wh), axis=-1)

    return box, box_confidence, box_class_probs


def filter_boxes(boxes, box_confidences, box_class_probs):
    """Filter boxes with box threshold. It's a bit different with origin yolov5 post process!

    # Arguments
        boxes: ndarray, boxes of objects.
        box_confidences: ndarray, confidences of objects.
        box_class_probs: ndarray, class_probs of objects.

    # Returns
        boxes: ndarray, filtered boxes.
        classes: ndarray, classes for boxes.
        scores: ndarray, scores for boxes.
    """
    box_classes = np.argmax(box_class_probs, axis=-1)
    box_class_scores = np.max(box_class_probs, axis=-1)
    pos = np.where(box_confidences[..., 0] >= BOX_THRESH)

    boxes = boxes[pos]
    classes = box_classes[pos]
    scores = box_class_scores[pos]

    return boxes, classes, scores


def nms_boxes(boxes, scores):
    """Suppress non-maximal boxes.

    # Arguments
        boxes: ndarray, boxes of objects.
        scores: ndarray, scores of objects.

    # Returns
        keep: ndarray, index of effective boxes.
    """
    x = boxes[:, 0]
    y = boxes[:, 1]
    w = boxes[:, 2] - boxes[:, 0]
    h = boxes[:, 3] - boxes[:, 1]

    areas = w * h
    order = scores.argsort()[::-1]

    keep = []
    while order.size > 0:
        i = order[0]
        keep.append(i)

        xx1 = np.maximum(x[i], x[order[1:]])
        yy1 = np.maximum(y[i], y[order[1:]])
        xx2 = np.minimum(x[i] + w[i], x[order[1:]] + w[order[1:]])
        yy2 = np.minimum(y[i] + h[i], y[order[1:]] + h[order[1:]])

        w1 = np.maximum(0.0, xx2 - xx1 + 0.00001)
        h1 = np.maximum(0.0, yy2 - yy1 + 0.00001)
        inter = w1 * h1

        ovr = inter / (areas[i] + areas[order[1:]] - inter)
        inds = np.where(ovr <= NMS_THRESH)[0]
        order = order[inds + 1]
    keep = np.array(keep)
    return keep


def yolov5_post_process(input_data):
    masks = [[0, 1, 2], [3, 4, 5], [6, 7, 8]]
    anchors = [[10, 13], [16, 30], [33, 23], [30, 61], [62, 45],
               [59, 119], [116, 90], [156, 198], [373, 326]]

    boxes, classes, scores = [], [], []
    for input, mask in zip(input_data, masks):
        b, c, s = process(input, mask, anchors)
        b, c, s = filter_boxes(b, c, s)
        boxes.append(b)
        classes.append(c)
        scores.append(s)

    boxes = np.concatenate(boxes)
    boxes = xywh2xyxy(boxes)
    classes = np.concatenate(classes)
    scores = np.concatenate(scores)

    nboxes, nclasses, nscores = [], [], []
    for c in set(classes):
        inds = np.where(classes == c)
        b = boxes[inds]
        c = classes[inds]
        s = scores[inds]

        keep = nms_boxes(b, s)

        nboxes.append(b[keep])
        nclasses.append(c[keep])
        nscores.append(s[keep])

    if not nclasses and not nscores:
        return None, None, None

    boxes = np.concatenate(nboxes)
    classes = np.concatenate(nclasses)
    scores = np.concatenate(nscores)

    return boxes, classes, scores


def draw(image, boxes, scores, classes):
    """Draw the boxes on the image.

    # Argument:
        image: original image.
        boxes: ndarray, boxes of objects.
        classes: ndarray, classes of objects.
        scores: ndarray, scores of objects.
        all_classes: all classes name.
    """
    for box, score, cl in zip(boxes, scores, classes):
        top, left, right, bottom = box
        print('class: {}, score: {}'.format(CLASSES[cl], score))
        print('box coordinate left,top,right,down: [{}, {}, {}, {}]'.format(top, left, right, bottom))

        top = int(top)
        left = int(left)
        right = int(right)
        bottom = int(bottom)

        cv2.rectangle(image, (top, left), (right, bottom), (255, 0, 0), 2)
        cv2.putText(image, '{0} {1:.2f}'.format(CLASSES[cl], score),
                    (top, left - 6),
                    cv2.FONT_HERSHEY_SIMPLEX,
                    0.6, (0, 0, 255), 2)


def letterbox(im, new_shape=(640, 640), color=(0, 0, 0)):
    # Resize and pad image while meeting stride-multiple constraints
    shape = im.shape[:2]  # current shape [height, width]
    if isinstance(new_shape, int):
        new_shape = (new_shape, new_shape)

    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])

    # Compute padding
    ratio = r, r  # width, height ratios
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding

    dw /= 2  # divide padding into 2 sides
    dh /= 2

    if shape[::-1] != new_unpad:  # resize
        im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    return im, ratio, (dw, dh)


def load_model():
    """载入model"""

    # Create RKNN object
    rknn_lite = RKNNLite()

    print('--> list devices:')
    rknn_lite.list_devices()
    print('done')

    print('--> query support target platform')
    rknn_lite.list_support_target_platform(rknn_model=RKNN_MODEL)
    print('done')

    print('--> Load RKNN model')
    ret = rknn_lite.load_rknn(RKNN_MODEL)
    if ret != 0:
        print('Load RKNN model failed')
        exit(ret)
    print('done')

    # init runtime environment
    print('--> Init runtime environment')
    # 指定平台
    # ret = rknn_lite.init_runtime(target='rv1126')
    ret = rknn_lite.init_runtime()
    print(ret)
    if ret != 0:
        print('Init runtime environment failed')
        exit(ret)
    print('done')
    return rknn_lite


if __name__ == '__main__':

    # rknn导入
    rknn = load_model()

    # Set inputs
    img = cv2.imread(IMG_PATH)  #  改变输入,可以获取摄像头数据源
    # img = cv2.resize(img, (640,640))
    img, ratio, (dw, dh) = letterbox(img, new_shape=(IMG_SIZE, IMG_SIZE))
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

    # print("--> get sdk version:")
    # sdk_version = rknn_lite.get_sdk_version()
    # print(sdk_version)
    # print("done")

    # Inference
    print('--> Running model')
    _force_builtin_perm = False
    t0 = time.time()
    outputs = rknn.inference(inputs=[img], inputs_pass_through=[0 if not _force_builtin_perm else 1])
    print("inference time:\t", time.time() - t0)
    # post process
    input0_data = outputs[0]
    input1_data = outputs[1]
    input2_data = outputs[2]

    input0_data = input0_data.reshape([3, -1] + list(input0_data.shape[-2:]))
    input1_data = input1_data.reshape([3, -1] + list(input1_data.shape[-2:]))
    input2_data = input2_data.reshape([3, -1] + list(input2_data.shape[-2:]))

    input_data = list()
    input_data.append(np.transpose(input0_data, (2, 3, 0, 1)))
    input_data.append(np.transpose(input1_data, (2, 3, 0, 1)))
    input_data.append(np.transpose(input2_data, (2, 3, 0, 1)))

    boxes, classes, scores = yolov5_post_process(input_data)

    img_1 = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
    if boxes is not None:
        draw(img_1, boxes, scores, classes)

    cv2.imwrite('result.jpg', img_1)
    # cv2.imshow("post process result", img_1)
    # cv2.waitKeyEx(0)

    rknn.release()
回复

使用道具 举报

blue

注册会员

积分
51
 楼主| 发表于 2022-5-21 11:46:00 | 显示全部楼层
blue 发表于 2022-5-21 11:45
"""将RKNN模型部署在RV1126平台并运行测试,目前只是简单推理"""
#环境:rv1126(Python3.7)
import time

这是运行在1126上的代码
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则

产品中心 购买渠道 开源社区 Wiki教程 资料下载 关于Toybrick


快速回复 返回顶部 返回列表