changelog.txt

Version 0.2.1: (master)
    
    Major refactor of the data.py API, now utilizes python generates and only makes a conversion to tf.data.Dataset when calling to_tfDataset()
    
    Changing the logging mechanism, now it utilizes a global library logger, instead of the BaseLogger Class (which was depricated)
    
    Change the logger default level to INFO and add an env variable that controls its value
    
    Adding a profiler, that traces the current tf train graph and also write its operations
    

Version 0.2.0:
    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
    

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
        
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