CODEWITHSUNDEEP
tensorflow
windweller
windweller

Reputation: 2385

Tensorflow variable Reuse in rnn module

I'm very perplexed by TF variable reuse. For method rnn, I'm able to find this line of code:

   if time > 0: vs.get_variable_scope().reuse_variables()

However, for dynamic_rnn (the method I need to use), I do not find any reuse_variable line of code, or reuse=True.

All the RNN cells in rnn_cells module are initializing using _linear method, which does not check if a variable has been created or not, but in LSTM_cell, _get_concat_variable does nicely check if a variable's name exists in graph_key or not.

So does dynamic_rnn not reuse the variable? Should I write a method to explicitly check if a variable is created and return it if it is?

Upvotes: 0

Views: 1011

Answers (1)

Ian Chen
Ian Chen

Reputation: 335

The dynamic_rnn function has a parameter called scope. So you should create your own scope (using with tf.variable_scope('scope_name', reuse=True)) and set it when calling dynamic_rnn function.

You can checkout the implementation of model_with_buckets, which reuses the same model for each buckets.

Upvotes: 1

Related Questions