Is there anyway to judge bewteen data_input variable nodes and trainable nodes?

TensorFlow General Discussion
, ,
1

I want to implement a grappler graph optimization pass. I want to do split for data input node only. For example, a node handling y = wx + b may have 3 input nodes, w , x and b.

I only want to do tf.split(x), not touch w and b.

So,

  1. is there any way to get x out from a list of nodes? (for example, get x from node->inputs()).
  2. Similarly, if the above w is trainable , and b is not trainable, how could I get this information from graph?

Thanks in advance

3

Hi @Stonepia

Thank you for using Tensorflow
I implemented the optimization pass in this notebook please look into it and make necessary changes to add it in your code.
Thank you.