python - 如何使用 Tensorflow Slim 库执行迭代推理

Question

有很多示例展示了如何使用 tf.contrib.slim 库对从网络下载的单个图像进行分类。事实上tensorflow github 提供了这个。然而，我正在努力理解循环执行此操作的最佳方法。任何使用 Tensorflow 进行分类的应用程序都必须对多于一批的图像进行分类。推理过程包括构建图表并从检查点加载权重。当迭代运行时，一遍又一遍地重复这些步骤似乎很浪费。事实上，当我尝试这个基本方法时，我可以看到分配给 python 的内存在每次迭代中都在持续增长。有人可以帮助建议如何修改基本示例以实现重复/迭代推理吗？这是我当前有效的方法，但显然浪费了内存资源(此代码使内存有限的机器崩溃，新图像定期转储到全局框架中): def 分类():

def classification():
  global frame
  global count

  slim = tf.contrib.slim
  image_size = inception_v4.inception_v4.default_image_size
  names = imagenet.create_readable_names_for_imagenet_labels()
  checkpoints_dir = '../../checkpoints'

  # Don't classify the first few frames
  while count < 5:
    pass

  while True:
    start = count
    with tf.Graph().as_default():
      image = tf.convert_to_tensor(frame,dtype=tf.float32)
      processed_image = inception_preprocessing.preprocess_image(image, image_size, image_size, is_training=False)
      processed_images  = tf.expand_dims(processed_image, 0)
      # processed_images will be a 1x299x299x3 tensor of float32

    # Create the model, use the default arg scope to configure the batch norm parameters.
    with slim.arg_scope(inception_v4.inception_v4_arg_scope()):
      logits, _ = inception_v4.inception_v4(processed_images, num_classes=1001, is_training=False)
      probabilities = tf.nn.softmax(logits)

    init_fn = slim.assign_from_checkpoint_fn(
      os.path.join(checkpoints_dir, 'inception_v4.ckpt'),
      slim.get_model_variables('InceptionV4'))

    with tf.Session() as sess:
      init_fn(sess)
      np_image, probabilities = sess.run([image, probabilities])
      probabilities = probabilities[0, 0:]
      sorted_inds = [i[0] for i in sorted(enumerate(-probabilities), key=lambda x:x[1])]

    for i in range(5):
      index = sorted_inds[i]
      print('Probability %0.2f%% => [%s]' % (probabilities[index] * 100, names[index]))

    end = count
    print "Classification latency = %d frames" % (end-start)

Answer 1

我让它工作起来，仍然会欣赏其他人的一些智慧。我的解决方案是使用占位符作为输入来构建图表。然后可以将视频帧输入到使用 feed_dict 的 session 运行方法中。这允许我将 while 循环放在对 session 运行的调用周围。使用这种方法的延迟是我分享的原始方法的1/10，并且内存指纹稳定。这是我用于对网络摄像头的视频帧进行分类的完整代码。请注意，它有一个问题。我没有干净地退出线程的机制。 Ctrl+C 不会终止脚本。另请注意，要运行此程序，您需要克隆 github tensorflow 模型存储库，并在 ../../checkpoints 下载并解压预训练权重。

import cv2
import os
import time
import numpy as np
from threading import Thread
import tensorflow as tf
from datasets import imagenet
from nets import inception_v4
from preprocessing import inception_preprocessing
######################################################

# Global Variables Shared by threads
frame = None
count = 0

######################################################
def capture():
######################################################
  global frame
  global count

  video_capture = cv2.VideoCapture(0)

  while True:
    # Capture frame-by-frame
    ret, frame_bgr = video_capture.read()

    # Display the resulting frame
    cv2.imshow('Video', frame_bgr)

    # Convert to RGB format (Inception expects RGB not BGR color channels)
    frame = cv2.cvtColor(frame_bgr,cv2.COLOR_BGR2RGB)

    # Increment frame counter (Used only to calculate latency)
    count += 1

    # Kill loop when user hits q
    if cv2.waitKey(1) & 0xFF == ord('q'):
      break

  # When everything is done, release the capture
  video_capture.release()
  cv2.destroyAllWindows()
######################################################

######################################################
def classification():
######################################################
  global frame
  global count

  slim = tf.contrib.slim
  image_size = inception_v4.inception_v4.default_image_size
  names = imagenet.create_readable_names_for_imagenet_labels()
  checkpoints_dir = '../../checkpoints'

  # Don't classify the None Object
  time.sleep(5)

  with tf.Graph().as_default():
    image = tf.placeholder(tf.uint8,[480,640,3])
    processed_image = inception_preprocessing.preprocess_image(image, 
    image_size, image_size, is_training=False)
    processed_images  = tf.expand_dims(processed_image, 0)
    # processed_images will be a 1x299x299x3 tensor of float32

    # Create the model, use the default arg scope to configure the batch norm parameters.
    with slim.arg_scope(inception_v4.inception_v4_arg_scope()):
      logits, _ = inception_v4.inception_v4(processed_images, num_classes=1001, is_training=False)
      probs = tf.nn.softmax(logits)

    init_fn = slim.assign_from_checkpoint_fn(
      os.path.join(checkpoints_dir, 'inception_v4.ckpt'),
      slim.get_model_variables('InceptionV4'))

    with tf.Session() as sess:
      init_fn(sess)

      while True:
        start = count
        probabilities = sess.run(probs,feed_dict={image: frame})
        probabilities = probabilities[0, 0:]
        sorted_inds = [i[0] for i in sorted(enumerate(-probabilities), key=lambda x:x[1])]

        for i in range(5):
          index = sorted_inds[i]
          print('Probability %0.2f%% => [%s]' % (probabilities[index] * 100, names[index]))

        end = count
        print "Classification latency = %d frames" % (end-start)

    # How to end this thread cleanly?
######################################################

# Start the threads
capture_thread = Thread(target=capture)
classify_thread = Thread(target=classification)
capture_thread.start()
classify_thread.start()