Skip to content

Latest commit

 

History

History
95 lines (68 loc) · 2.79 KB

RELEASE.rst

File metadata and controls

95 lines (68 loc) · 2.79 KB
Raw

Release 0.9.0

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.

Release 0.8.0

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

Release 0.7.0

  • 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.

Release 0.6.5

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.

Release 0.6.4

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

Release 0.6.3

  • 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.

Release 0.6.2

Hotfix: Fix the dropout noise shape so we get samples of the latent function of the layer (rather than the observations). Also some doco tweaks.

Release 0.6.1

Hotfix: Fix regression whereby setting the random seed was not working with the new distribution objects from TensorFlow (tf.distributions).

Release 0.6.0

Some moderate changes to the API from:

  • Using TensorFlow's tf.distributions to replace Aboleth's likelihoods
  • Using TensorFlow's tf.distributions to replace Aboleth's distributions

Release 0.5.0

Initial release of Aboleth