TensorFlow: AttributeError: 'Session' object has no attribute '_session'

Question

I'm trying to run my code using TensorFlow.

init = tf.global_variables_initializer()

loaded_graph = tf.Graph()
saver = tf.train.Saver()

with tf.Session(loaded_graph) as sess:
    sess.run(init)
    ...

But I got this error.

  File "C:\Users\K451LN\My Documents\LiClipse Workspace\neuralnet\FFNN.py", line 68, in <module>
with tf.Session(loaded_graph) as sess:
AttributeError: 'Session' object has no attribute '_session'

Is there anything wrong with the tf.Graph()?

Here is my code:

for i in range(num_networks):
     print("Neural network: {0}".format(i))

     X = tf.placeholder(tf.float32)
     Y = tf.placeholder(tf.float32)

     W1 = tf.Variable(tf.random_uniform([n_input, n_hidden], -1.0, 1.0), name = 'W1')
     W2 = tf.Variable(tf.random_uniform([n_hidden, n_output], -1.0, 1.0), name = 'W2')

     b1 = tf.Variable(tf.zeros([n_hidden]), name="Bias1")
     b2 = tf.Variable(tf.zeros([n_output]), name="Bias2")

     L2 = tf.sigmoid(tf.matmul(X, W1) + b1)
     hy = tf.sigmoid(tf.matmul(L2, W2) + b2, name="op_to_restore")

     cost = tf.reduce_mean(-Y*tf.log(hy) - (1-Y)*tf.log(1-hy))
     optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)

    init = tf.global_variables_initializer()

    loaded_graph = tf.Graph()
    saver = tf.train.Saver()

    with tf.Session(loaded_graph) as sess:
        sess.run(init)
    ...

I'm adding thistf.Graph() to solve the error of ValueError: At least two variables have the same name: Bias2.

Burton2000 · Accepted Answer · 2018-03-16 11:21:10Z

1

Passing loaded_graph to the tf.Session() means you can only run ops created in that graph. As all you do is create a graph called loaded_graph but dont add anything to it then you get this error when trying to do sess.run(init) as init op is not in loaded_graph's graph.

I guess the reason for your original error with Bias2 is the for loop. If you remove the for loop and don't create/pass loaded_graph you wont have any errors.

If you wish to have the for loop then you may need to on each loop create a new graph using

g_1 = tf.Graph() with g_1.as_default(): ...

so your code will be like:

for i in range(num_networks):
     g_1 = tf.Graph()
     with g_1.as_default():

         print("Neural network: {0}".format(i))

         X = tf.placeholder(tf.float32)
         Y = tf.placeholder(tf.float32)

         W1 = tf.Variable(tf.random_uniform([n_input, n_hidden], -1.0, 1.0), name = 'W1')
         W2 = tf.Variable(tf.random_uniform([n_hidden, n_output], -1.0, 1.0), name = 'W2')

         b1 = tf.Variable(tf.zeros([n_hidden]), name="Bias1")
         b2 = tf.Variable(tf.zeros([n_output]), name="Bias2")

         L2 = tf.sigmoid(tf.matmul(X, W1) + b1)
         hy = tf.sigmoid(tf.matmul(L2, W2) + b2, name="op_to_restore")

         cost = tf.reduce_mean(-Y*tf.log(hy) - (1-Y)*tf.log(1-hy))
         optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)

         init = tf.global_variables_initializer()


         saver = tf.train.Saver()

         with tf.Session(graph=g_1) as sess:
             sess.run(init)
             ...

edited Mar 16, 2018 at 11:21

answered Mar 16, 2018 at 10:40

Burton2000

2,11218 silver badges28 bronze badges

Sign up to request clarification or add additional context in comments.

2 Comments

gonzalloe Over a year ago

FYI, the for loop is added because i wish to create an ensemble NN. I've try your suggestion with g_1.as_default():, but seems doesn't work for me too. Any other suggestions?

Burton2000 Over a year ago

Sorry forgot you need to put tf.Session(graph = g_1) as first argument to tf.Session() is target and not graph. Updated my answer.

Collectives™ on Stack Overflow

TensorFlow: AttributeError: 'Session' object has no attribute '_session'

1 Answer 1

2 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

2 Comments

Your Answer

Sign up or log in

Post as a guest

Related