ValueError：仅使用命名参数（标签=...，logits=...，...）调用`softmax_cross_entropy_with_logits`

Question

你能指导如何解决这个问题吗？

with tf.name_scope('loss'):
    #cross_entropy = None
    val = tf.nn.softmax_cross_entropy_with_logits(y_conv, y_)
    cross_entropy = tf.reduce_mean(val)

with tf.name_scope('adam_optimizer'):
    #train_step = None
    train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)

我收到此错误：

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-40-f67d0aecc114> in <module>()
      1 with tf.name_scope('loss'):
      2     #cross_entropy = None
----> 3     val = tf.nn.softmax_cross_entropy_with_logits(y_conv, y_)
      4     cross_entropy = tf.reduce_mean(val)
      5 

~/anaconda/lib/python3.6/site-packages/tensorflow/python/ops/nn_ops.py in softmax_cross_entropy_with_logits(_sentinel, labels, logits, dim, name)
   1576   """
   1577   _ensure_xent_args("softmax_cross_entropy_with_logits", _sentinel,
-> 1578                     labels, logits)
   1579 
   1580   # TODO(pcmurray) Raise an error when the labels do not sum to 1. Note: This

~/anaconda/lib/python3.6/site-packages/tensorflow/python/ops/nn_ops.py in _ensure_xent_args(name, sentinel, labels, logits)
   1531   if sentinel is not None:
   1532     raise ValueError("Only call `%s` with "
-> 1533                      "named arguments (labels=..., logits=..., ...)" % name)
   1534   if labels is None or logits is None:
   1535     raise ValueError("Both labels and logits must be provided.")

ValueError: Only call `softmax_cross_entropy_with_logits` with named arguments (labels=..., logits=..., ...)

另外，tf.__version__ 返回'1.0.0' 我在 OSX Sierra 上有 Anaconda Python 3.6.2。

Answer 1

这是一个简单的解决方法：softmax_cross_entropy_with_logits() 具有三个关键参数：_sentinel、labels 和 logits。哨兵必须为空，需要使用命名参数。

已修复（虽然我不确定 y_conv 或 y_ 在这种情况下是标签还是 logit，因此您可能需要交换它们）：

with tf.name_scope('loss'):
    #cross_entropy = None
    val = tf.nn.softmax_cross_entropy_with_logits(labels = y_conv, logits=y_)
    cross_entropy = tf.reduce_mean(val)

with tf.name_scope('adam_optimizer'):
    #train_step = None
    train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)