Simple-elmo

Changes
=======
Fixed loading models without vocabularies from ZIP archives

Changes
=======
`Simple_elmo` is now compatible with the up-to-date versions of the [smart_open](https://github.com/RaRe-Technologies/smart_open/) library.

Changes
=======
* The `get_elmo_vectors()` and `get_elmo_vector_average()` methods can now re-use existing TensorFlow sessions. This is extremely useful, for example, when using ELMo models as embedding generators for downstream models. See an example in the README.
* Further speed improvements for the new experimental `get_elmo_substitutes()` method for language modeling

Changes
=======

* The new `get_elmo_substitutes()` method for language modeling now uses batching and is significantly faster.
* Language mode predictions yield lists instead of numpy arrays, for easier serializing.
* Language model predictions yield raw logit scores instead of probabilities.

Changes
=======

* A new `get_elmo_substitutes()` method for language modeling. Requires full model checkpoint with the softmax weights (should be loaded with `full=True`). Still experimental.
* The number of deprecation warnings is reduced.

Changes
=======

* Whether to use the top layer or the average of all layers is now the `layers` argument of the `get_elmo_vectors()` and `get_elmo_vectors_average()` methods. It is no longer necessary to state this when loading a model.
* The `layers` argument can take the `all` value. In this case, raw weights from all 3 ELMo layers are returned.
* Small fixes