实时计算 OpenCV Python 预测的运行时间

Question

我正在尝试获取 TensorFlow 中神经网络预测的运行时间，并在 OpenCV 中使用 Python 运行实时推理。每当分类为真时，我都会在窗口中显示一个文本标签，并且我非常想测量它的真实时间为多少秒。

我的想法是获取流开始时的开始时间，然后获取预测的开始时间并减去这两个时间以获得经过的时间。

但这让我得到了与预期不同的价值观。解决这个问题的最佳方法是什么？

这是我的代码：

import time
import cv2

cap_device = args.device
cap_width = args.width
cap_height = args.height

cap = cv2.VideoCapture(cap_device, cv2.CAP_DSHOW)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, cap_width)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, cap_height)



start_time = time.time() #get start time
class_time = 0 #initialize class time
with mp_pose.Pose(model_complexity = 2) as pose_tracker: #load model
  while cap.isOpened(): #while camera is running 
    ret, frame = cap.read() #read frames in
    ...
    ...
    ...

    with open("labels.txt") as file: #gets labels to show pose class
      lines = file.readlines()
      lines = [line.rstrip() for line in lines]
      
      
          #classification confidence
    for i in range(0, len(prediction[0])): 
          if prediction[0][0] > 0.50 and prediction[0][0]<=1.00:
              draw_info_text(image, lines[0],cap_width,cap_height) 
              class_time = time.time()
          elif prediction[0][1] > 0.50 and prediction[0][1]<=1.00: =
              draw_info_text(image,lines[1],cap_width,cap_height)
        
          elif prediction[0][2] > 0.50 and prediction[0][2]<= 1.00: 
              draw_info_text(image,lines[2],cap_width,cap_height)
            
          elif prediction[0][3] > 0.50 and prediction[0][3] <= 1.00: 
              draw_info_text(image,lines[3],cap_width,cap_height)
                

    image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)

    #show video
    cv2.imshow('webcam', image)
    if cv2.waitKey(1) == ord('q'): 
      break

total_prediction_time = start_time - class_time # calculate the total time pred was true 
print(total_prediction_time)

cap.release()
cv2.destroyAllWindows()

Answer 1

我认为你可以利用 python 装饰器，
这是一个例子：

# Timer decorator


def timer(fn):    
    from time import perf_counter
    def inner(*args, **kwargs):
        start_time = perf_counter()
        to_execute = fn(*args, **kwargs)
        end_time = perf_counter()
        execution_time = end_time - start_time
        print('{0} took {1:.8f}s to execute'.format(fn.__name__, execution_time))
        return to_execute

    return inner

@timer
def function_1():
    for i in range(100000):
        pass

@timer
def function_2():
    for i in range(10000000):
        pass


@timer
def function_3():
    for i in range(10000):
        pass

@timer
def function_4():
    for i in range(100000000):
        pass

function_1()
function_2()
function_3()
function_4()