v0.2 MultiGPU train

Adding multiGPU training with the integration of the horovod library, which handles the communication between the GPU or machines.

Adding a dockerfile for installing the horovod since it can be a hassle to install horovod with the correct dependencies

This version also denotes a compatibility checkpoint with the BioASQ2022

v0.1.9

Version 0.1.9:

Refactor of the validation callback, adding more deafults mode of operation

Integration of a HPO (hyperparameter optimizer) to the polus lib
    - Implementing optuna as the default backend for HPO
    - Implementing a prune callback
   
Adding add_lookup_data to the CachedDataloader, which makes a conversion to the CachedDataloaderwLookup instance

Adding deep_copy method to CachedDataloader

Adding tutorials folder with python scripts for examples
    - Creation of a classifier_example

Adding more unit tests

v0.1.8

Version 0.1.8:

• Adding a mask input to the CRF layer for masking impossible transitions (binary potentials)
  

• Adding complex json serializer and deserializer
  

• Adding split_bert_model that splits a TFBertModel into two keras models based on the given TFBertLayer index
  

• Adding PolusModel abstraction that implements the init_from_data function
  

Version 0.1.7:

• Adding tests for some data and models functions
  

• A lot of minor changes not individually discriminated
  

• Refactor and bug fixes
  

Version 0.1.6:

• Introduction of a information retrieval package with trainers focused in dense retrieval
  

• CacheDataLoaderWlookup:
  

•     - Extension of the CacheDataLoader but it will additionally save an python object associated with the cached dataset.
  

Version 0.1.5:

• - experimental.data moved to data
  

• CacheDataLoader:
  

•     - add clean method that deletes the index and associated files
  

•     - in case of error, automaticly cleans any file that was created
  

• training:
  

•     - Creation of the BaseTrainer abstraction
  

•     - Creation of a ClassifierTrainer instance that holds the logic to train models according to the classification problem
  

for_biocreative

This is the first pre-release and acts as a checkpoint that ensures backward compatibility with the biocreative 2021 tracks