如何在运行 Tensorflow 推理会话之前批处理多个视频帧

我做了一个项目，基本上使用谷歌对象检测 API 和张量流。

我所做的就是使用预先训练的模型进行推理：这意味着实时对象检测，其中输入是网络摄像头的视频流或使用 OpenCV 的类似内容。

现在我得到了相当不错的性能结果，但我想进一步提高 FPS。

因为我的经验是，Tensorflow 在推理时使用了我的整个内存，但 GPU 使用率根本没有达到最大值（NVIDIA GTX 1050 笔记本电脑上约为 40%，NVIDIA Jetson Tx2 上约为 6%）。

所以我的想法是通过增加每个会话运行中输入的图像批量大小来增加 GPU 使用率。

所以我的问题是：在将输入视频流的多个帧提供给它们之前，如何将它们一起批处理sess.run()?

看看我的代码object_detetection.py在我的 github 仓库上：(https://github.com/GustavZ/realtime_object_detection https://github.com/GustavZ/realtime_object_detection).

如果您能提出一些提示或代码实现，我将非常感激！

import numpy as np
import os
import six.moves.urllib as urllib
import tarfile
import tensorflow as tf
import cv2


# Protobuf Compilation (once necessary)
os.system('protoc object_detection/protos/*.proto --python_out=.')

from object_detection.utils import label_map_util
from object_detection.utils import visualization_utils as vis_util
from stuff.helper import FPS2, WebcamVideoStream

# INPUT PARAMS
# Must be OpenCV readable
# 0 = Default Camera
video_input = 0
visualize = True
max_frames = 300 #only used if visualize==False
width = 640
height = 480
fps_interval = 3
bbox_thickness = 8

# Model preparation
# What model to download.
MODEL_NAME = 'ssd_mobilenet_v1_coco_2017_11_17'
MODEL_FILE = MODEL_NAME + '.tar.gz'
DOWNLOAD_BASE = 'http://download.tensorflow.org/models/object_detection/'
# Path to frozen detection graph. This is the actual model that is used for the object detection.
PATH_TO_CKPT = 'models/' + MODEL_NAME + '/frozen_inference_graph.pb'
# List of the strings that is used to add correct label for each box.
LABEL_MAP = 'mscoco_label_map.pbtxt'
PATH_TO_LABELS = 'object_detection/data/' + LABEL_MAP
NUM_CLASSES = 90

# Download Model    
if not os.path.isfile(PATH_TO_CKPT):
    print('Model not found. Downloading it now.')
    opener = urllib.request.URLopener()
    opener.retrieve(DOWNLOAD_BASE + MODEL_FILE, MODEL_FILE)
    tar_file = tarfile.open(MODEL_FILE)
    for file in tar_file.getmembers():
      file_name = os.path.basename(file.name)
      if 'frozen_inference_graph.pb' in file_name:
        tar_file.extract(file, os.getcwd())
    os.remove('../' + MODEL_FILE)
else:
    print('Model found. Proceed.')

# Load a (frozen) Tensorflow model into memory.
detection_graph = tf.Graph()
with detection_graph.as_default():
  od_graph_def = tf.GraphDef()
  with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:
    serialized_graph = fid.read()
    od_graph_def.ParseFromString(serialized_graph)
    tf.import_graph_def(od_graph_def, name='')

# Loading label map
label_map = label_map_util.load_labelmap(PATH_TO_LABELS)
categories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=NUM_CLASSES, use_display_name=True)
category_index = label_map_util.create_category_index(categories)

# Start Video Stream
video_stream = WebcamVideoStream(video_input,width,height).start()
cur_frames = 0
# Detection
with detection_graph.as_default():
  with tf.Session(graph=detection_graph) as sess:
    # Definite input and output Tensors for detection_graph
    image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
    # Each box represents a part of the image where a particular object was detected.
    detection_boxes = detection_graph.get_tensor_by_name('detection_boxes:0')
    # Each score represent how level of confidence for each of the objects.
    # Score is shown on the result image, together with the class label.
    detection_scores = detection_graph.get_tensor_by_name('detection_scores:0')
    detection_classes = detection_graph.get_tensor_by_name('detection_classes:0')
    num_detections = detection_graph.get_tensor_by_name('num_detections:0')
    # fps calculation
    fps = FPS2(fps_interval).start()
    print ("Press 'q' to Exit")
    while video_stream.isActive():
      image_np = video_stream.read()
      # Expand dimensions since the model expects images to have shape: [1, None, None, 3]
      image_np_expanded = np.expand_dims(image_np, axis=0)
      # Actual detection.
      (boxes, scores, classes, num) = sess.run(
          [detection_boxes, detection_scores, detection_classes, num_detections],
          feed_dict={image_tensor: image_np_expanded})
      # Visualization of the results of a detection.
      vis_util.visualize_boxes_and_labels_on_image_array(
          image_np,
          np.squeeze(boxes),
          np.squeeze(classes).astype(np.int32),
          np.squeeze(scores),
          category_index,
          use_normalized_coordinates=True,
          line_thickness=bbox_thickness)
      if visualize:
          cv2.imshow('object_detection', image_np)
          # Exit Option
          if cv2.waitKey(1) & 0xFF == ord('q'):
              break
      else:
          cur_frames += 1
          if cur_frames >= max_frames:
              break
      # fps calculation
      fps.update()

# End everything
fps.stop()
video_stream.stop()     
cv2.destroyAllWindows()
print('[INFO] elapsed time (total): {:.2f}'.format(fps.elapsed()))
print('[INFO] approx. FPS: {:.2f}'.format(fps.fps()))

嗯，我只是收集batch_size框架并喂养它们：

batch_size = 5
while video_stream.isActive():
  image_np_list = []
  for _ in range(batch_size):
      image_np_list.append(video_stream.read())
      fps.update()
  # Expand dimensions since the model expects images to have shape: [1, None, None, 3]
  image_np_expanded = np.asarray(image_np_list)
  # Actual detection.
  (boxes, scores, classes, num) = sess.run(
      [detection_boxes, detection_scores, detection_classes, num_detections],
      feed_dict={image_tensor: image_np_expanded})

  # Visualization of the results of a detection.
  for i in range(batch_size):
      vis_util.visualize_boxes_and_labels_on_image_array(
          image_np_expanded[i],
          boxes[i],
          classes[i].astype(np.int32),
          scores[i],
          category_index,
          use_normalized_coordinates=True,
          line_thickness=bbox_thickness)
          if visualize:
              cv2.imshow('object_detection', image_np_expanded[i])
              # Exit Option
              if cv2.waitKey(1) & 0xFF == ord('q'):
                  break

当然，如果您正在读取检测结果，则必须在此之后进行相关更改，因为它们现在将具有batch_size rows.

但要小心：在tensorflow 1.4之前（我认为），对象检测API仅支持批量大小为 1 https://github.com/tensorflow/models/issues/1816 in image_tensor，所以除非你升级你的张量流，否则这将不起作用。

另请注意，您生成的 FPS 将是平均值，但同一批次中的帧实际上比不同批次之间的时间更接近（因为您仍然需要等待sess.run()完成）。尽管两个连续帧之间的最大时间应该增加，但平均值仍应明显优于当前的 FPS。

如果您希望帧之间的间隔大致相同，我想您将需要更复杂的工具，例如多线程和队列：一个线程将从流中读取图像并将其存储在队列中，另一个线程将需要他们从队列中出来并打电话sess.run()异步处理它们；它还可以告诉第一个线程根据其自身的计算能力加快或减慢速度。这实施起来比较棘手。