python - tensorflow boolean_mask 如何在两个张量之间进行掩码？

Question

我有如下代码：

def yolo_filter_boxes(box_confidence, boxes, box_class_probs, threshold = .6):
    """Filters YOLO boxes by thresholding on object and class confidence.

    Arguments:
    box_confidence -- tensor of shape (3, 3, 5, 1)
    boxes -- tensor of shape (3, 3, 5, 4)
    box_class_probs -- tensor of shape (3, 3, 5, 80)
    threshold -- real value, if [ highest class probability score < threshold], then get rid of the corresponding box

    Returns:
    scores -- tensor of shape (None,), containing the class probability score for selected boxes
    boxes -- tensor of shape (None, 4), containing (b_x, b_y, b_h, b_w) coordinates of selected boxes
    classes -- tensor of shape (None,), containing the index of the class detected by the selected boxes

    Note: "None" is here because you don't know the exact number of selected boxes, as it depends on the threshold. 
    For example, the actual output size of scores would be (10,) if there are 10 boxes.
    """

    # Step 1: Compute box scores
    box_scores = np.multiply(box_confidence, box_class_probs)

    # Step 2: Find the box_classes thanks to the max box_scores, keep track of the corresponding score
    box_classes = K.argmax(box_scores, -1)
    box_class_scores = K.max(box_scores, -1)

    # Step 3: Create a filtering mask based on "box_class_scores" by using "threshold". The mask should have the
    # same dimension as box_class_scores, and be True for the boxes you want to keep (with probability >= threshold)
    filtering_mask = K.greater_equal(box_class_scores,threshold)
    # Step 4: Apply the mask to scores, boxes and classes

    print(filtering_mask.shape)
    print(filtering_mask.eval())

    print(box_class_scores.shape)
    print(box_class_scores.eval())
    scores = tf.boolean_mask(box_class_scores, filtering_mask)
    print(scores.eval())
    boxes = tf.boolean_mask(boxes, filtering_mask)
    classes = tf.boolean_mask(box_classes, filtering_mask)

    return scores, boxes, classes


with tf.Session() as test_a:
    box_confidence = tf.random_normal([3, 3, 5, 1], mean=1, stddev=4, seed = 1)
    boxes = tf.random_normal([3, 3, 5, 4], mean=1, stddev=4, seed = 1)
    box_class_probs = tf.random_normal([3, 3, 5, 80], mean=1, stddev=4, seed = 1)
    scores, boxes, classes = yolo_filter_boxes(box_confidence, boxes, box_class_probs, threshold = 0.5)
    print("scores[2] = " + str(scores[2].eval()))
    print("boxes[2] = " + str(boxes[2].eval()))
    print("classes[2] = " + str(classes[2].eval()))
    print("scores.shape = " + str(scores.shape))
    print("boxes.shape = " + str(boxes.shape))
    print("classes.shape = " + str(classes.shape))

这是输出：

(3, 3, 5)
[[[ True  True  True  True  True]
  [ True  True  True  True  True]
  [ True False  True  True  True]]

 [[ True  True  True  True  True]
  [ True  True  True  True  True]
  [ True  True  True  True  True]]

 [[ True  True  True  True False]
  [ True  True  True  True  True]
  [ True  True  True  True  True]]]
(3, 3, 5)
[[[  45.00004959   21.20238304   17.39275742   26.73288918   49.47431946]
  [  22.16205978   27.96604347   12.38916492   33.66600418   62.04590225]
  [ 113.03194427    2.68868852    6.33391762   45.17211914   10.5103178 ]]

 [[   8.22186852   35.88579941   48.54780579   12.48789883   32.40937042]
  [  75.73269653   17.52830696   62.99983597   29.0468502    42.82471848]
  [  72.42234039  108.19727325   36.93912888   40.9789238    36.91137314]]

 [[   1.57321405    3.35663748   16.33576775    5.16499805   19.43038177]
  [  48.13769913   68.20082092   47.06818008    1.82166731   67.30760956]
  [  33.01203537   63.93298721    9.71860027   49.06838989   60.74739456]]]
[  22.63684464   10.29589462   58.76845551   74.67560577   20.25722504
   47.24279022    6.96320772   22.59087944   86.61974335    1.05248117
   57.47060394   92.50878143   16.8335762    23.29385757   78.58971405
    6.95861435   65.61254883   45.47106171   43.53435135   10.0660677
   60.34520721   28.5535984    15.9668026    45.14865494    5.49425364
    2.35473752   29.40540886    2.5579865    46.96302032    9.39739799
   45.78501892   49.42660904   34.68322754   40.72031784   58.91592407
   35.39850616   56.24537277    6.80519342    9.52552414  138.54457092
   14.07888412   56.37608719   69.59171295   25.83714676]
scores[2] = 62.0051
boxes[2] = [-1.89158893  0.7749185   3.57417917 -0.05729628]
classes[2] = 36
scores.shape = (?,)
boxes.shape = (?, 4)
classes.shape = (?,)

我有一个简单的问题。结果是怎么scores来的？它有 44 个元素，同时filtering_mask有box_class_scores45 个元素（3 * 3 * 5），而 filtering_mask 有 2 个错误值，必须使分数为 43 个元素。即使 filter_mask 有 1 个 false 值，分数中的数字都不匹配box_class_scores。任何人都可以向我解释如何scores计算

标签： pythontensorflowkeras

Answer 1

问题不在于像您期望的那样起作用的掩蔽。问题是您在图表中使用随机值，其行为可能有点令人惊讶。每次调用eval()实际上都是run在默认会话中调用。问题在于 TensorFlow 中随机值的工作方式。每次run在会话上调用时，都会生成一个新的随机值。这意味着每次调用都会根据和的eval不同值产生结果。有可能解决它的方法，要么根本不使用随机值生成器作为输入，要么在同一次调用中评估所有输出（而不是box_confidenceboxesbox_class_probsruneval）。由于您似乎正在编写测试代码，解决它的一种简单方法是将输入替换为由 NumPy 随机值制成的常量。

import tensorflow as tf
import numpy as np

def yolo_filter_boxes(box_confidence, boxes, box_class_probs, threshold = .6):
    # ...

with tf.Session() as test_a:
    np.random.seed(1)
    box_confidence = tf.constant(np.random.normal(loc=1, scale=4, size=[3, 3, 5, 1]), dtype=tf.float32)
    boxes = tf.constant(np.random.normal(loc=1, scale=4, size=[3, 3, 5, 4]), dtype=tf.float32)
    box_class_probs = tf.constant(np.random.normal(loc=1, scale=4, size=[3, 3, 5, 80]), dtype=tf.float32
    scores, boxes, classes = yolo_filter_boxes(box_confidence, boxes, box_class_probs, threshold = 0.5)
    print("scores[2] = " + str(scores[2].eval()))
    print("boxes[2] = " + str(boxes[2].eval()))
    print("classes[2] = " + str(classes[2].eval()))
    print("scores.shape = " + str(scores.shape))
    print("boxes.shape = " + str(boxes.shape))
    print("classes.shape = " + str(classes.shape))

或者您仍然可以使用 TensorFlow 随机数，但使用变量作为输入。变量的不同之处在于它们只在初始化时评估它们的初始值，然后它们在会话之间保持它们的值（直到它再次被改变），所以你不会每次都生成新的随机值。

import tensorflow as tf

def yolo_filter_boxes(box_confidence, boxes, box_class_probs, threshold = .6):
    # ...

with tf.Session() as test_a:
    box_confidence = tf.Variable(tf.random_normal([3, 3, 5, 1], mean=1, stddev=4, seed = 1)))
    boxes = tf.Variable(tf.random_normal([3, 3, 5, 4], mean=1, stddev=4, seed = 1))
    box_class_probs = tf.Variable(tf.random_normal([3, 3, 5, 80], mean=1, stddev=4, seed = 1))
    # You must initialize the variables
    test_a.run(tf.global_variables_initializer())
    scores, boxes, classes = yolo_filter_boxes(box_confidence, boxes, box_class_probs, threshold = 0.5)
    print("scores[2] = " + str(scores[2].eval()))
    print("boxes[2] = " + str(boxes[2].eval()))
    print("classes[2] = " + str(classes[2].eval()))
    print("scores.shape = " + str(scores.shape))
    print("boxes.shape = " + str(boxes.shape))
    print("classes.shape = " + str(classes.shape))