TensorFlow：使用 tf.merge_all_summaries() 时出现 PlaceHolder 错误

Question

我收到一个占位符错误。

我不知道这是什么意思，因为我在 sess.run(..., {_y: y, _X: X}) 上正确映射...我在这里提供了一个功能齐全的 MWE 来重现错误：

import tensorflow as tf
import numpy as np

def init_weights(shape):
    return tf.Variable(tf.random_normal(shape, stddev=0.01))

class NeuralNet:
    def __init__(self, hidden):
        self.hidden = hidden

    def __del__(self):
        self.sess.close()

    def fit(self, X, y):
        _X = tf.placeholder('float', [None, None])
        _y = tf.placeholder('float', [None, 1])

        w0 = init_weights([X.shape[1], self.hidden])
        b0 = tf.Variable(tf.zeros([self.hidden]))
        w1 = init_weights([self.hidden, 1])
        b1 = tf.Variable(tf.zeros([1]))

        self.sess = tf.Session()
        self.sess.run(tf.initialize_all_variables())

        h = tf.nn.sigmoid(tf.matmul(_X, w0) + b0)
        self.yp = tf.nn.sigmoid(tf.matmul(h, w1) + b1)

        C = tf.reduce_mean(tf.square(self.yp - y))
        o = tf.train.GradientDescentOptimizer(0.5).minimize(C)

        correct = tf.equal(tf.argmax(_y, 1), tf.argmax(self.yp, 1))
        accuracy = tf.reduce_mean(tf.cast(correct, "float"))
        tf.scalar_summary("accuracy", accuracy)
        tf.scalar_summary("loss", C)

        merged = tf.merge_all_summaries()
        import shutil
        shutil.rmtree('logs')
        writer = tf.train.SummaryWriter('logs', self.sess.graph_def)

        for i in xrange(1000+1):
            if i % 100 == 0:
                res = self.sess.run([o, merged], feed_dict={_X: X, _y: y})
            else:
                self.sess.run(o, feed_dict={_X: X, _y: y})
        return self

    def predict(self, X):
        yp = self.sess.run(self.yp, feed_dict={_X: X})
        return (yp >= 0.5).astype(int)


X = np.array([ [0,0,1],[0,1,1],[1,0,1],[1,1,1]])
y = np.array([[0],[1],[1],[0]]])

m = NeuralNet(10)
m.fit(X, y)
yp = m.predict(X)[:, 0]
print accuracy_score(y, yp)

错误：

I tensorflow/core/common_runtime/local_device.cc:40] Local device intra op parallelism threads: 8
I tensorflow/core/common_runtime/direct_session.cc:58] Direct session inter op parallelism threads: 8
0.847222222222
W tensorflow/core/common_runtime/executor.cc:1076] 0x2340f40 Compute status: Invalid argument: You must feed a value for placeholder tensor 'Placeholder_1' with dtype float
     [[Node: Placeholder_1 = Placeholder[dtype=DT_FLOAT, shape=[], _device="/job:localhost/replica:0/task:0/cpu:0"]()]]
W tensorflow/core/common_runtime/executor.cc:1076] 0x2340f40 Compute status: Invalid argument: You must feed a value for placeholder tensor 'Placeholder' with dtype float
     [[Node: Placeholder = Placeholder[dtype=DT_FLOAT, shape=[], _device="/job:localhost/replica:0/task:0/cpu:0"]()]]
Traceback (most recent call last):
  File "neuralnet.py", line 64, in <module>
    m.fit(X[tr], y[tr, np.newaxis])
  File "neuralnet.py", line 44, in fit
    res = self.sess.run([o, merged], feed_dict={self._X: X, _y: y})
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 368, in run
    results = self._do_run(target_list, unique_fetch_targets, feed_dict_string)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 444, in _do_run
    e.code)
tensorflow.python.framework.errors.InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder_1' with dtype float
     [[Node: Placeholder_1 = Placeholder[dtype=DT_FLOAT, shape=[], _device="/job:localhost/replica:0/task:0/cpu:0"]()]]
Caused by op u'Placeholder_1', defined at:
  File "neuralnet.py", line 64, in <module>
    m.fit(X[tr], y[tr, np.newaxis])
  File "neuralnet.py", line 16, in fit
    _y = tf.placeholder('float', [None, 1])
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/array_ops.py", line 673, in placeholder
    name=name)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gen_array_ops.py", line 463, in _placeholder
    name=name)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/op_def_library.py", line 664, in apply_op
    op_def=op_def)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 1834, in create_op
    original_op=self._default_original_op, op_def=op_def)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 1043, in __init__
    self._traceback = _extract_stack()

如果我删除tf.merge_all_summaries() 或从self.sess.run([o, merged], ...) 中删除merged，那么它运行正常。

这看起来类似于这篇文章： Error when computing summaries in TensorFlow 但是，我没有使用 iPython...

Answer 1

tf.merge_all_summaries() 函数很方便，但也有些危险：它合并了默认图表中的所有摘要，其中包括来自先前（显然未连接）代码调用的任何摘要，这些代码还添加了摘要节点到默认图表。如果旧的摘要节点依赖于旧的占位符，您将收到类似于您在问题中显示的错误（以及previousquestions）。

有两种独立的解决方法：

1. 确保您明确收集要计算的摘要。这就像在您的示例中使用显式 tf.merge_summary() 操作一样简单：

   accuracy_summary = tf.scalar_summary("accuracy", accuracy)
   loss_summary = tf.scalar_summary("loss", C)
   
   merged = tf.merge_summary([accuracy_summary, loss_summary])
   

2. 确保每次创建一组新摘要时，都在新图表中进行。推荐的风格是使用显式的默认图表：

   with tf.Graph().as_default():
     # Build model and create session in this scope.
     #
     # Only summary nodes created in this scope will be returned by a call to
     # `tf.merge_all_summaries()`
   

   或者，如果您使用的是 TensorFlow 的最新开源版本（或即将发布的 0.7.0 版本），您可以调用 tf.reset_default_graph() 来重置图的状态并删除任何旧的汇总节点。