Integrate all of Glove's vocabulary into the model.

[benlabbe]

By default, tfner only stores the words seen during the learning phase. So we have plenty of OOV during the inference.

And to go to the end of this principle, it would even be necessary to experiment with a FastText embedding in tfner, making sure that we use the lib FastText (and not just the vector dictionary) .