Question

当我为RNN的tf.metrics.mean_per_class_accuracy操作运行一个张量流会话时，我的代码将返回FailedPreconditionError，如：

FailedPreconditionError: Attempting to use uninitialized value mean_accuracy/count
   [[Node: mean_accuracy/count/read = Identity[T=DT_FLOAT, 
      _class=["loc:@mean_accuracy/count"], 
      _device="/job:localhost/replica:0/task:0/device:CPU:0"](mean_accuracy/count)]]

我想知道是否在操作平均精度之前初始化张量，然后初始化权重和偏差张量，以使其与训练过的张量值不同。

我的代码如下：

    import tensorflow as tf
    import numpy as np

    x_raw = np.random.rand(10,3,3)
    y_raw = np.random.randint(0,high=2,size=10)
    x_test = np.random.rand(1,3,3)
    y_test = np.random.randint(0,high=2,size=1)

    tf.reset_default_graph()
    x = tf.placeholder(tf.float32, shape=[None, 3, 3])
    y = tf.placeholder(tf.float32, shape=[None, 2])

    def rnn(x):
        l = {'w': tf.Variable(tf.random_normal([5, 2])),
             'b': tf.Variable(tf.random_normal([2]))}
        x = tf.transpose(x, (1,0,2))
        x = tf.reshape(x, (-1, 3))
        x = tf.split(x, 3, axis=0)
        cell = tf.nn.rnn_cell.BasicLSTMCell(5)
        out, stt = tf.contrib.rnn.static_rnn(cell, x, dtype=tf.float32)
        o = tf.matmul(out[-1], l['w'])+l['b']
        return o

    pr = rnn(x)
    co = tf.nn.softmax_cross_entropy_with_logits_v2(labels=y, logits=pr)
    om = tf.train.AdamOptimizer(learning_rate=0.01).minimize(co)
    cr = tf.equal(tf.argmax(pr,1), tf.argmax(y,1))
    ac0 = tf.reduce_mean(tf.cast(cr,tf.float32))
    ac1 = tf.metrics.mean_per_class_accuracy(y,pr,2)

    sess = tf.Session()
    sess.run(tf.global_variables_initializer())

    c, _ = sess.run([co,om], feed_dict={x:x_raw[:5], y:tf.Session().run(tf.one_hot(y_raw[:5],2))})
    print(sess.run(ac0, feed_dict={x:x_test, y:tf.Session().run(tf.one_hot(y_test,2))}))
    print(sess.run(ac1, feed_dict={x:x_test, y:tf.Session().run(tf.one_hot(y_test,2))}))

当我的控制台接近用于操作“ ac1”张量的线时，将引发错误消息。

Answer 1

tf.metrics.mean_per_class_accuracy在tf.GraphKeys.LOCAL_VARIABLES中创建变量，而不是在GLOBAL_VARIABLES中创建变量。这意味着它们不会由您的tf.global_variables_initializer()初始化。

幸运的是，有一个相应的tf.local_variables_initializer()可以帮助您。将其与您的全局初始化程序一起放入，应该很好！

sess.run([tf.global_variables_initializer(), tf.local_variables_initializer()])

Rf的tf.mean_per_class_accuracy中的未初始化值问题

1 个答案: