- Made loading TensorFlow distribution objects cleaner.
- Added better support for sparse variables and minibatching parameters.
- Added the ability to run algorithms step by step. This allows custom storage of parameters, useful when the full chain does not fit into memory!
- Added new vignette to demonstrate step by step functionality – a Bayesian neural network model.
- Changed optimizer to TensorFlow SGDOptimizer for control variate methods.