CODEWITHSUNDEEP
tensorflow
DVK
DVK

Reputation: 515

How to push back altered variables into a tensorflow graph?

I am trying to manipulate the weights of a multilayered LSTM RNN by first extracting the trainable variables in a session, like so:

    variables_names =[v.name for v in tf.trainable_variables()]
    values = session.run(variables_names)

Now the variable values is a list of all weights and biases in my tensorflow graph. After some arithmetic operations to the weights and biases, I would like to upload them back into the graph to continue training the RNN. Does anyone know how to do this? I thought if I use the numpy list values as a source for re-initializing the graph it would work but I have been unsuccessful. I've tried the following methods till now:

init = tf.constant(values)
tf.get_variables(variables_names, initializer = init).run()

and

init = tf.constant(values)
session.run(tf.variables_initializer(values))

In both these cases the code finishes executing abruptly by printing values I want to initialize back into the graph.

Upvotes: 0

Views: 153

Answers (1)

Patrick Coady
Patrick Coady

Reputation: 216

Variables have a .load() method. You can pass your updated values back to the graph this way.

Upvotes: 1

Related Questions