Collie

Added
- new model architectures ``CollaborativeMetricLearningModel``, ``MLPMatrixFactorizationModel``, and ``DeepFM``
Changed
- filename for ``HybridPretrainedModel`` to ``hybrid_pretrained_matrix_factorization.py``. The former model filepath is now deprecated and will be removed in future version ``0.6.0``
- ``collie.model.base`` is now split into its own directory with the same name
- reduced boilerplate docstrings required for models
- all ``model.freeze() -> model.eval()``
- bumped version of ``sphinx-rtd-theme`` to ``0.5.2``

Added
- ``CollieMinimalTrainer`` for a faster, simpler version of ``CollieTrainer``
- ``remove_duplicate_user_item_pairs`` argument to ``Interactions``
Changed
- renamed `BasePipeline.hparams.n_epochs_completed_ -> BasePipeline.hparams.num_epochs_completed`
Fixed
- a proper ``ValueError`` is now raised if no ``train`` data is passed into a model
- loss docstrings that incorrectly stated ``**kwargs`` would be accepted

Changed
- disable automated batching in ``ApproximateNegativeSamplingInteractionsDataLoader`` and ``HDF5InteractionsDataLoader``

Changed
- ``convert_to_implicit`` will now remove duplicate user/item pairs in DataFrame

Fixed
- duplicate user/item pairs in ``Interactions`` are now dropped from the COO matrix during instantiation

Added
- ability to run ``stratified_split`` without any ``joblib.Parallel`` parallelization
- data quality checks to ``Interactions.__init__`` to assert ``users`` and ``items`` and ``mat`` are not ``None`` and ``ratings`` does not contain any ``0``s (if so, those rows will now automatically be filtered out)
- increased test coverage
- header table to all Jupyter notebooks with links to Colab and GitHub
Changed
- default ``processes`` for ``stratified_split`` is now ``-1``
- default ``k`` value in ``mapk`` is now set to ``10``
- when GPU is available but not set, ``CollieTrainer`` now sets it to ``1`` rather than ``-1``
- all models now check that ``train_loader`` and ``val_loader`` attributes are consistent during initialization
- default ``unseen_items_only`` in ``BasePipeline.get_item_predictions`` method is now ``False``
- docs in ``get_recommendation_visualizations`` to be clearer
- ``get_recommendation_visualizations`` data quality checks have been moved to the beginning of the function to potentially fail faster
- ``create_ratings_matrix`` no longer raises ``ValueError`` if ``users`` and ``items`` do not start at ``0``
- refactored ``adaptive_hinge_loss``
Removed
- ``kwargs`` option for methods that did not explicitly need them
Fixed
- typo in ``cross_validation.py`` error message
- ``head`` and ``tail`` methods in ``interactions/datasets.py`` to no longer error with ``n < 1`` or large ``n`` values
- ``num_users`` and ``num_items`` are no longer incorrectly incremented when ``meta`` key is provided in ``HDF5Interactions``
- type hints for ``device`` now also include instances of ``torch.device``
- the type of metadata tensors sent to ``HybridPretrainedModel`` are now consistent across all input options
- removed ineffective quality checks in ``HybridPretrainedModel.save_model``
- no longer use deprecated ``nn.sigmoid`` in library
- a ``relu`` final activation layer now works in ``NeuralCollaborativeFiltering`` model
- ``df_to_html`` now outputs proper HTML when multiple ``html_tags`` options are specified
- tutorial notebooks now fully run on Colab without having to only rely on previously-saved data
- add value of ``1e-11`` to ``BasePipeline.get_item_predictions`` denominator to avoid potential ``NaN``s