improve performance

[Tilana]

Dependent on the length of the documents the universal sentence encoder needs about 30-50 secs to find similar sentences in 10 documents.
For the CNN the lookup table is build by iterating through the fasttext vocabulary.
Let's see how to speed up the performance of the different algorithms.

[Tilana]

Furthermore, as the number of training samples the time required to trained the CNN increases

[Tilana]

Improve performance of USE
tensorflow/hub#40

[Tilana]

improves performance for finding similar sentences in a document of 2-3 pages from 14s to 10s

[Tilana]

build tensorflow graph in predict_one_model outside of the predictDoc function to build the graph only once and then compute prediction results for the different documents.
Improves predictDoc performance significantly (from about 2.30 min to 1 min10)