AttributeError: module 'tensorflow' has no attribute 'assign'

Question

I am trying to migrate my previous tf1 code to tf2. Unfortunately my code was not on eager mode so I am having more difficulties. I did the following code (not yet training) and I got into the error message:

Traceback (most recent call last):
    training_op = tf.assign(W, W - learning_rate * gradients)
AttributeError: module 'tensorflow' has no attribute 'assign'

This is my minimum code example PS: it has to work with complex numbers!

# Data pre-processing
    m = 50
    n = 20
    x_train, y_train, x_test, y_test = get_my_data(x, y, m, n) # data x of size mxn

    # Network Declaration
    input_size = n
    output_size = 1
    learning_rate = 0.001  # The optimization learning rate
    # Create weight matrix initialized randomely from N~(0, 0.01)
    W = tf.Variable(tf.complex(np.random.rand(input_size, output_size),
                               np.random.rand(input_size, output_size)), name="weights")

    with tf.GradientTape() as gtape:
        y_out = tf.matmul(x_train, W, name="out")
        error = y - y_out
        loss = tf.reduce_mean(tf.square(tf.abs(error)), name="mse")
        gradients = gtape.gradient(loss, [W])[0]
        training_op = tf.assign(W, W - learning_rate * gradients)

I do this manually because unless they changed that, optimizers are not supported for complex numbers so I do it "by hand".

In TensorFlow 2.0 assign is no more a function of the TensorFlow package but it is a method of the tf.Variable objects. So instead of tf.assign(x, new_value), you can use x.assign(new_value). — Dr Nisha Arora
– Dr Nisha Arora, Commented Sep 6, 2020 at 5:04

Arcones · Accepted Answer · 2019-11-08 15:33:19Z

8

Try tf.compat.v1.assign instead. It worked for me.

edited Nov 8, 2019 at 15:33

Arcones

4,8328 gold badges29 silver badges52 bronze badges

answered Nov 8, 2019 at 15:10

Venkateswara Sarma Krishnamoor

811 bronze badge

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

Comments

Sergei Lebedev · Accepted Answer · 2019-10-04 11:34:55Z

5

tf.assign* functions are available as methods on tf.Variable in TF 2.0. So, your example could be rewritten as

with tf.GradientTape() as gtape:
    ...
    W.assign_sub(learning_rate * gradients)

Note that unlike tf.assign in TF 1.X, tf.Variable.assing_sub will execute the assignment eagerly.

answered Oct 4, 2019 at 11:34

Sergei Lebedev

2,67920 silver badges23 bronze badges

1 Comment

J Agustin Barrachina Over a year ago

Ok, so this gradient and assign I should do it in the training iteration then right?

Collectives™ on Stack Overflow

AttributeError: module 'tensorflow' has no attribute 'assign'

2 Answers 2

Comments

1 Comment

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

Comments

1 Comment

Your Answer

Sign up or log in

Post as a guest

Related