Question

我想围绕Sonnet模块包装tf.metrics以测量每个批次的性能，以下是我所做的工作：

import tensorflow as tf
import sonnet as snt

class Metrics(snt.AbstractModule):
    def __init__(self, indicator, summaries = None, name = "metrics"):
        super(Metrics, self).__init__(name = name)
        self._indicator = indicator
        self._summaries = summaries

    def _build(self, labels, logits):
        if self._indicator == "accuracy":
            metric, metric_update = tf.metrics.accuracy(labels, logits)
            with tf.control_dependencies([metric_update]):
                outputs = tf.identity(metric)
        elif self._indicator == "precision":
            metric, metric_update = tf.metrics.precision(labels, logits)
            with tf.control_dependencies([metric_update]):
                outputs = tf.identity(metric)
        elif self._indicator == "recall":
            metric, metric_update = tf.metrics.recall(labels, logits)
            with tf.control_dependencies([metric_update]):
                outputs = tf.identity(metric)
        elif self._indicator == "f1_score":
            metric_recall, metric_update_recall = tf.metrics.recall(labels, logits)
            metric_precision, metric_update_precision = tf.metrics.precision(labels, logits)
            with tf.control_dependencies([metric_update_recall, metric_update_precision]):
                outputs = 2.0 / (1.0 / metric_recall + 1.0 / metric_precision)
        else:
            raise ValueError("unsupported metrics")

        if type(self._summaries) == list:
            self._summaries.append(tf.summary.scalar(self._indicator, outputs))

        return outputs

但是，当我想测试模块时，以下代码可以工作：

def test3():
    import numpy as np

    labels = tf.constant([1, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], tf.int32)
    logits = tf.constant([1, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], tf.int32)

    metrics = Metrics("accuracy")
    accuracy = metrics(labels, logits)

    metrics2 = Metrics("f1_score")
    f1_score = metrics2(labels, logits)

    writer = tf.summary.FileWriter("utils-const", tf.get_default_graph())
    with tf.Session() as sess:
        sess.run([tf.global_variables_initializer(), tf.local_variables_initializer()])

        accu, f1 = sess.run([accuracy, f1_score])
        print(accu)
        print(f1)

    writer.close()

但是以下代码不起作用：

def test4():
    from tensorflow.python import debug as tf_debug
    import numpy as np

    tf_labels = tf.placeholder(dtype=tf.int32, shape=[None])
    tf_logits = tf.placeholder(dtype=tf.int32, shape=[None])

    labels = np.array([1, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], np.int32)
    logits = np.array([1, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], np.int32)

    metrics = Metrics("accuracy")
    accuracy = metrics(tf_labels, tf_logits)

    metrics2 = Metrics("f1_score")
    f1_score = metrics2(tf_labels, tf_logits)

    writer = tf.summary.FileWriter("utils-feed", tf.get_default_graph())
    with tf.Session() as sess:
        sess.run([tf.global_variables_initializer(), tf.local_variables_initializer()])

        sess = tf_debug.LocalCLIDebugWrapperSession(sess)

        accu, f1 = sess.run([accuracy, f1_score], feed_dict = {tf_labels: labels, tf_logits: logits})
        print(accu)
        print(f1)

    writer.close()

test3（）的输出是正确的，0.88。 test4（）的输出错误，0.0。但是，它们应该是等价的。

任何人都有任何想法？

Answer 1

您确定tf.constant版本失败了吗？我发现tf.metrics与tf.constant结合使用时有一种奇怪的行为：

import tensorflow as tf

a = tf.constant(1.)
mean_a, mean_a_uop = tf.metrics.mean(a)
with tf.control_dependencies([mean_a_uop]):
  mean_a = tf.identity(mean_a)

sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
tf.local_variables_initializer().run()

for _ in range(10):
  print(sess.run(mean_a))

在GPU上运行时返回

而不是1 s。好像计数落后于一。（我假设第一个值为inf但由于count上的某些条件而为零。另一方面，此代码的占位符版本正在按预期运行。

在CPU上，行为甚至更奇怪，因为输出是非确定性的。输出示例：

0.0
1.0
1.0
0.75
1.0
1.0
0.85714287
0.875
1.0
0.9

看起来像是一个可以登录tensorflow's github repo的错误。（请注意，在常量上使用运行指标不太有用 - 但它仍然是一个错误。）

编辑现在我也偶然发现tf.placeholder的奇怪示例，tf.metrics似乎有一个错误，但不仅限于tf.constant使用它}第

tf.constant和tf.placeholder行为不同

1 个答案: