Tensorflow-lattice

Changes:
* (experimental) Parameterization option for Premade/Estimators that enables the use of both normal tfl.layers.Lattice layers ('all_vertices') and tfl.layers.KroneckerFactoredLattice layers ('kronecker_factored').
* (experimental) KroneckerFactoredLattice layer visualization support for Estimators.
* (experimental) KroneckerFactoredLattice bound constraints.
* 'input_keypoints_type' parameter for PWLCalibration layers that enables learned input keypoints ('learned_interior') or the original fixed keypoints ('fixed').
* General tutorial/code cleanup
* Typo fixes
* Bug fixes

PyPI Release:
* Generic package for py3 that should work for TF 1.15 or TF 2.x.

Changes:
* (experimental) KroneckerFactoredLattice initialization now sorts on kernel axis 1 such that we sort each term individually.
* (experimental) KroneckerFactoredLattice initialization defaults to [0.5, 1.5] instead of [0,1].
* (experimental) KroneckerFactoredLattice custom_reduce_prod in interpolation for faster gradient computations.
* Update bound and trust projection algorithms to compute violations for each unit separately.
* 'loss_fn' option for estimators to use custom loss without having to define a custom head.
* Enable calibrators to return a list of outputs per unit.
* Enable RTL layer to return non-averaged outputs.
* General tutorial/code cleanup
* Typo fixes
* Bug fixes

PyPI Release:
* Generic package for py3 that should work for TF 1.15 or TF 2.x.

TensorFlow is dropping py2 support, so we will be dropping support as well in our future releases. This is the last release that will support py2.

Changes:
* New (experimental) KroneckerFactoredLattice Layer, which introduces a new parameterization of our Lattice layer with linear space/time complexity.
* rtl_lib.py helper functions for RTL Layer.
* Utils module with useful helper functions for all layers.
* 'rtl_layer' option for CalibratedLatticeEnsemble Premade Models and Canned Estimators, which uses an RTL Layer for the underlying ensemble. Can potentially give a speed-boost for models with a large number of lattices.
* General code cleanup
* Typo fixes
* Bug fixes

PyPI release:
* Generic package for py2/py3 that should work for TF 1.15 or TF 2.x.

Changes:
* Simplex interpolation support for lattices: O(d log(d)) simplex interpolation compared to O(2^d) hypercube interpolation is 2-10x faster with similar or improved training loss.
* RTL layer performance optimization: 2-3x faster and scales much better with wider and deeper models with tens of thousands of lattices.
* Optimization of 2^D hypercube lattices: 10-15% speedup.
* PWL Calibration Sonnet Module (more to come in follow up releases)
* New aggregation function tutorial
* Linear combination support for canned ensemble models.
* Improvement and bug fixes for save/load functionality
* Bug fixes

PyPI release:
* Generic package for py2/py3 that should work for TF 1.15 or TF 2.x.

Changes:
* Save/load support for Keras models (HDF5/H5 format)
* RTL layer: An ensemble of Lattice layers that takes in a collection of monotonic and unconstrained features and randomly arranges them into lattices of a given rank.
* AggregateFunction Premade model and Aggregation layer: Applies monotonic function on set inputs passed in as ragged tensors.
* Crystals Lattice ensemble with Premade model
* Feature updates to Lattice layer
* Updates to tutorials
* Bug fixes

PyPI release:
* Generic package for py2/py3 that should work for TF 1.15 or TF 2.x.

Changes:
* Two new tutorials: premade models, shape constraints for ML fairness
* Improvements and additions to premade models
* New range dominance for Lattice and Linear layers
* Added 'peak' mode to unimodality constraint
* Updates to documentation
* Bug fixes

PyPI release:
* Generic package for py2/py3 that should work for TF 1.15 or TF 2.x.