Is it possible to calculate the caching (sub-) of a TensorFlow graph?

Question

Can TensorFlow automatically cache calculations if they involve multiple calls to the same calculation (sub) graph?

For example, I have a matrix F in which each record is a calculation based on the trained variables W My objective function multiplies this matrix several times by different vectors (each time with constant W).

Will TensorFlow reprogram, for example, F[1,2] whenever I access it, or will it cache this value?

In theory, one could precompute the matrix F for a fixed W , so that each entry in F is equal to tf.constant . But this will prevent the correct calculation of the gradients W

Answer 1

TensorFlow does a limited amount of caching, but probably does not apply to the case you are describing.

If you create tf.Session with the following parameters, permanent folding will be enabled:

 config = tf.ConfigProto(graph_options=tf.GraphOptions( optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L2))) sess = tf.Session(config=config) 

When you call sess.run() with this configuration, TensorFlow will evaluate the corresponding nodes to run, and then identify the subgraph of those nodes whose outputs are constant, evaluate them and cache the results. Therefore, it will avoid re-performing redundant calculations.

However, in your question, you note that F is a function of some trained variables. From the point of view of TensorFlow, these variables are volatile - they can change at any time and therefore do not cache values ​​obtained from these variables. If you want to reuse the same value for F several times, you can consider storing it in tf.constant() so that bending optimization is more useful.