Hub: xling embedding extremely slow when not using feed dict / tf.Variable

[r-wheeler]

Importing the module

import tensorflow as tf 
import tensorflow_hub as hub 

xling_8_embed = hub.Module("https://tfhub.dev/google/universal-sentence-encoder-xling-many/1")

%time xling_8_embed(tf.placeholder(dtype=tf.string, shape=[None]))



md5-c5e9d354dee17fbbc095a40a7c060028



%time xling_8_embed(tf.constant(["dog", "Puppies are nice.", "I enjoy taking long walks along the beach with my dog."]))



md5-cdea590e8ca0ba7ccdbb697ab71d0410



def model_fn(features, labels):
    embeddings = xling_8_embed(features['text'])
...

[rmothukuru]

@r-wheeler ,
There is growth in the graph for each time you apply the "Module.__call__" and seeing slowness due to it.
The correct pattern with TF Graphs/Sessions is to create the nodes in the graph that describe the computation that you want to execute and then execute it multiple times with session.run(), with the right feed_dict arguments.

[tgsmith61591]

@rmothukuru would you happen to have an example of "the correct pattern"? This is pretty confusing

[r-wheeler]

@rmothukuru. Thanks for the reply!

In both of these examples I am only calling xling_8_embed 1 time, the only difference is call being abled to atf.Variable or a tf.Tensor

[andresusanopinto]

Hi r-wheeler you are right there is something wrong with the fact that feeding a constant takes minutes to create the graph and feeding a placeholder just takes a few seconds.

For the time being you can work around with tf.placeholder_with_default, to somehow trigger the code paths that are efficient. E.g. something like:

 inputs = tf.constant(["dog", "Puppies are nice.", "I enjoy taking long walks along the beach with my dog."])
 inputs = tf.placeholder_with_default(inputs, [None])
 xling_8_embed(inputs)

The root cause of the bad code paths is not clear yet.

PS: for the record, a copy-paste example ready to repro in colab.

[j-grimstad]

Closed due to lack of activity. Please reopen if issue persists.