Get Gate Activation of GRU-Cells in TensorFlow

I implemented a standard RNN in Tensorflow (0.12) using the classes

  • tf.python.ops.rnn_cell.GRUCell
  • tf.nn.dynamic_rnn

I am interested in the gates and which values they take during. Unfortunately the function dynamic_rnn doesn't support this and just gives the 'output' and 'state' (The final hidden state). In the code of the GRUCell the gates are called 'r' and 'u' for reset and update.

Is there a clever way to save the values of the gates over time or do I have to write my own GRU cell?

Thanks!

Topic tensorflow

Category Data Science

1
answered at

I am trying to figure out the same problem as well. I don't have the full solution but maybe we can help each other out.

If you do: names =[n.name for n in tf.get_default_graph().as_graph_def().node]

This will give you a list of all the names of the nodes in the graph. For Rnn the out put for names should look some what like this

.
'RNN/Shape',
'RNN/strided_slice/stack',
'RNN/strided_slice/stack_1',
.
. 
.
'RNN/transpose'
.

Then you can use get_operation_by_name to get the operation you want to look at. 

 rnn_transpose=tf.get_default_graph().get_operation_by_name('RNN/transpose').outputs[0]

From that you can run the eval() with your regular sess and feed_dict

 rnn_transpose.eval(session=sess,feed_dict =feed_dict)

The problem here is now we have to find which operations in the graph correspond to the proper gate operations.

About

Geeks Mental is a community that publishes articles and tutorials about Web, Android, Data Science, new techniques and Linux security.