Aboleth

=============

This release focuses on better initialisation for the weights, and improves the
performance of feed-forward neural nets.

- Self-normalising neural net initialisation and dropout options.
- Noise contrastive prior layers for better uncertainty estimation away from
training data.
- TensorFlow Custom estimator interface demonstrated in the SARCOS demos.
- Simplifies interfaces for learning priors etc in the variational and kernel
layers.
- Remove "MAP" nomenclature from the non-variational layers, as these layers
have no regularisation by default now.
- Simplifies imputation layers interfaces.

=============

Refactor the user interface for more clarity and flexibility. Also a lot of
code maintenance and TensorBoard integration, specifically:

- Compatibility checked with TensorFlow up to r1.6.
- Convert the likelihoods to tensors away from distributions.
- Clarify what is being optimised in the layers (do not optimise priors by
default)
- Clean up the imputation module
- Make all Variables constructed within the layers view-able trough TensorBoard

=============

- Update to TensorFlow r1.4.
- Tutorials in the documentation on:

1. Interfacing with Keras
2. Saving/loading models
3. How to build a variety of regressors with Aboleth

- New prediction module with some convenience functions, including freezing the
weight samples during prediction.
- Bayesian convolutional layers with accompanying demo.
- Allow the number of samples drawn from a model to be varied by using
placeholders.
- Generalise the feature embedding layers to work on matrix inputs (instead of
just column vectors).
- Numerous numerical and usability fixes.

=============

Hotfix: Test batch shape of likelihoods to see if they are compatible with
models. Without this test the likelihoods may be broadcast, and result in poor
performance.

=============

Hotfix: Make a ab.MaskInputLayer for binary mask inputs when we don't want to
tile the inputs.

=============

- Make ab.InputLayer always make at least 1 sample of the networks for
consistency and simplicity.
- This also makes the quick start guide examples work.