python - 无法使用给定 session 评估张量 : the tensor's graph is different from the session's graph

Question

我只是不知道出了什么问题......我之前尝试过 InteractiveSession() 并传递了一个显式 session ，但这个错误没有得到解决......我是 tensorflow 的新手......请帮忙。

cost=-tf.reduce_sum(y*tf.log(y_))
train_step=tf.train.AdamOptimizer(LEARNING_RATE).minimize(cost)
correct_pred=tf.equal(tf.argmax(y,1),tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, 'float'))
predict=tf.argmax(y,1)

这是我的 session

train_accuracies = []
validation_accuracies = []
x_range = []

num_examples=train_images.shape[0]
init=tf.global_variables_initializer()
minibatches=random_mini_batches(train_images,train_labels,
                            mini_batch_size = BATCH_SIZE)
display_step=1
init = tf.initialize_all_variables()
with tf.Session().as_default() as sess:
sess.run(init)
for epoch in range(TRAINING_ITERATIONS):
    for minibatch in minibatches:
        (minibatch_X,minibatch_Y)=minibatch
        if epoch%display_step == 0 or (epoch+1) == TRAINING_ITERATIONS:

            train_accuracy = accuracy.eval(session=sess,feed_dict={x:minibatch_X, 
                                                      y: minibatch_Y, 
                                                      keep_prob: 1.0})       
        if(VALIDATION_SIZE):
            validation_accuracy = accuracy.eval(session=sess,feed_dict={ x: validation_images[0:BATCH_SIZE], 
                                                            y: validation_labels[0:BATCH_SIZE], 
                                                            keep_prob: 1.0})                                  
            print('training_accuracy / validation_accuracy => %.2f / %.2f for step %d'%(train_accuracy, validation_accuracy, epoch))

            validation_accuracies.append(validation_accuracy)

        else:
             print('training_accuracy => %.4f for step %d'%(train_accuracy, epoch))
        train_accuracies.append(train_accuracy)
        x_range.append(epoch)

        # increase display_step
        if epoch%(display_step*10) == 0 and epoch:
            display_step *= 10
    # train on batch
    sess.run(train_step, feed_dict={x: minibatch_X, y:minibatch_Y, keep_prob: DROPOUT})

并生成以下错误

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-63-910bbc0840b2> in <module>
      18                 train_accuracy = accuracy.eval(session=sess,feed_dict={x:minibatch_X, 
      19                                                           y: minibatch_Y,
 ---> 20                                                           keep_prob: 1.0})       
      21             if(VALIDATION_SIZE):
      22                 validation_accuracy = accuracy.eval(session=sess,feed_dict={ x: 
      validation_images[0:BATCH_SIZE], 

      /opt/conda/lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py in eval(self, 
     feed_dict, session)
      788 
      789     """
  --> 790     return _eval_using_default_session(self, feed_dict, self.graph, session)
      791 
      792   def experimental_ref(self):

      /opt/conda/lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py in 
     _eval_using_default_session(tensors, feed_dict, graph, session)
      5307   else:
      5308     if session.graph is not graph:
   -> 5309       raise ValueError("Cannot use the given session to evaluate tensor: "
      5310                        "the tensor's graph is different from the session's "
      5311                        "graph.")

      ValueError: Cannot use the given session to evaluate tensor: the tensor's graph is different 
      from the session's graph.

请建议如何使用两个 session 以及如何解决此问题。主要问题是我尝试将 session 作为 eval(session=sess) 传递，但它不起作用。意思是我用的计算图和accuracy tensor的图不一样

Answer 1

我已经重新创建了由于可能的方式导致的错误，并且还提供了修复。

在代码中提供了更多注释，以便更清楚地了解错误和修复。

注意 - 我使用了相同的代码并稍作调整来重新创建错误原因的可能性并修复相同的错误。

最佳修复代码出现在这个答案的末尾。

错误代码 1 -默认 session 和使用在另一个图中创建的变量时出错

%tensorflow_version 1.x
import tensorflow as tf

g = tf.Graph()
with g.as_default():
  x = tf.constant(1.0)  # x is created in graph g

with tf.Session().as_default() as sess:
  y = tf.constant(2.0) # y is created in TensorFlow's default graph!!!
  print(y.eval(session=sess)) # y was created in TF's default graph, and is evaluated in
                  # default session, so everything is ok.  
  print(x.eval(session=sess)) # x was created in graph g and it is evaluated in session s
                  # which is tied to graph g, but it is evaluated in
                  # session s which is tied to graph g => ERROR

输出-

2.0
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-5-f35cb204cf59> in <module>()
     10   print(y.eval(session=sess)) # y was created in TF's default graph, and is evaluated in
     11                   # default session, so everything is ok.
---> 12   print(x.eval(session=sess)) # x was created in graph g and it is evaluated in session s
     13                   # which is tied to graph g, but it is evaluated in
     14                   # session s which is tied to graph g => ERROR

1 frames
/tensorflow-1.15.2/python3.6/tensorflow_core/python/framework/ops.py in _eval_using_default_session(tensors, feed_dict, graph, session)
   5402   else:
   5403     if session.graph is not graph:
-> 5404       raise ValueError("Cannot use the given session to evaluate tensor: "
   5405                        "the tensor's graph is different from the session's "
   5406                        "graph.")

ValueError: Cannot use the given session to evaluate tensor: the tensor's graph is different from the session's graph.

错误代码 2 -默认图形 session 和使用在默认图形中创建的变量时出错

%tensorflow_version 1.x
import tensorflow as tf

g = tf.Graph()
with g.as_default():
  x = tf.constant(1.0)  # x is created in graph g

with tf.Session(graph=g).as_default() as sess:
  print(x.eval(session=sess)) # x was created in graph g and it is evaluated in session s
                         # which is tied to graph g, so everything is ok.
  y = tf.constant(2.0) # y is created in TensorFlow's default graph!!!
  print(y.eval()) # y was created in TF's default graph, but it is evaluated in
                  # session s which is tied to graph g => ERROR

输出-

1.0
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-15-6b8b687c5178> in <module>()
     10                          # which is tied to graph g, so everything is ok.
     11   y = tf.constant(2.0) # y is created in TensorFlow's default graph!!!
---> 12   print(y.eval()) # y was created in TF's default graph, but it is evaluated in
     13                   # session s which is tied to graph g => ERROR

1 frames
/tensorflow-1.15.2/python3.6/tensorflow_core/python/framework/ops.py in _eval_using_default_session(tensors, feed_dict, graph, session)
   5396                        "`eval(session=sess)`")
   5397     if session.graph is not graph:
-> 5398       raise ValueError("Cannot use the default session to evaluate tensor: "
   5399                        "the tensor's graph is different from the session's "
   5400                        "graph. Pass an explicit session to "

ValueError: Cannot use the default session to evaluate tensor: the tensor's graph is different from the session's graph. Pass an explicit session to `eval(session=sess)`.

错误代码 3 - 按照 错误代码 2 - 输出中的建议，将显式 session 传递给 eval(session=sess)。让我们试试这个。

%tensorflow_version 1.x
import tensorflow as tf

g = tf.Graph()
with g.as_default():
  x = tf.constant(1.0)  # x is created in graph g

with tf.Session(graph=g).as_default() as sess:
  print(x.eval(session=sess)) # x was created in graph g and it is evaluated in session s
                         # which is tied to graph g, so everything is ok.
  y = tf.constant(2.0) # y is created in TensorFlow's default graph!!!
  print(y.eval(session=sess)) # y was created in TF's default graph, but it is evaluated in
                  # session s which is tied to graph g => ERROR

输出-

1.0
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-16-83809aa4e485> in <module>()
     10                          # which is tied to graph g, so everything is ok.
     11   y = tf.constant(2.0) # y is created in TensorFlow's default graph!!!
---> 12   print(y.eval(session=sess)) # y was created in TF's default graph, but it is evaluated in
     13                   # session s which is tied to graph g => ERROR

1 frames
/tensorflow-1.15.2/python3.6/tensorflow_core/python/framework/ops.py in _eval_using_default_session(tensors, feed_dict, graph, session)
   5402   else:
   5403     if session.graph is not graph:
-> 5404       raise ValueError("Cannot use the given session to evaluate tensor: "
   5405                        "the tensor's graph is different from the session's "
   5406                        "graph.")

ValueError: Cannot use the given session to evaluate tensor: the tensor's graph is different from the session's graph.

修复 1 - 修复默认 session 和未分配给任何图形的变量

%tensorflow_version 1.x
import tensorflow as tf

x = tf.constant(1.0)  # x is in not assigned to any graph

with tf.Session().as_default() as sess:
  y = tf.constant(2.0) # y is created in TensorFlow's default graph!!!
  print(y.eval(session=sess)) # y was created in TF's default graph, and is evaluated in
                  # default session, so everything is ok.  
  print(x.eval(session=sess)) # x not assigned to any graph, and is evaluated in
                  # default session, so everything is ok.  

输出-

2.0
1.0

修复 2 - 最佳修复 是将构造阶段和执行阶段完全分开。

import tensorflow as tf

g = tf.Graph()
with g.as_default():
  x = tf.constant(1.0)  # x is created in graph g
  y = tf.constant(2.0) # y is created in graph g

with tf.Session(graph=g).as_default() as sess:
  print(x.eval()) # x was created in graph g and it is evaluated in session s
                         # which is tied to graph g, so everything is ok.
  print(y.eval()) # y was created in graph g and it is evaluated in session s
                         # which is tied to graph g, so everything is ok.

输出-

1.0
2.0