PAGI 0.1.2

• Introduction of new concepts in the training/validation/testing regime to clarify differences between "phases" (e.g. train, validate, test) and "batch types" (e.g. training, encoding)
• Consolidation of AE Stack and Visual Cortex components
• Added disk space check to avoid silent failure of TensorBoard writer (when disk runs out of space)
• Added missing summary checks in Sparse Conv Maxpool component
• Minor bugfixes and improvements

PAGI 0.1.1

• Added max_to_keep option to the tf.train.Saver to increase the number of checkpoints saved
• Combined leaky-relu and leaky_relu in the activation function selector as one condition, so either option should return tf.nn.leaky_relu
• Fixed a major bug where tf_do_training was always returning the training op, even at test time
• Added softmax nonlinearity to the activation function selector
• Major improvements to the Embedding base class
• Other bug fixes and minor improvements

PAGI 0.1.0

• Converted PAGI into an installable package with all the necessary tooling and dependencies
• Brought the components, workflows and utils up to date with the Project AGI internal codebase
• Improvements to the base Autoencoder component to share dual and summary related methods by inheriting from DualComponent & Summary Component
• Revamped the 'secondary' ops in Autoencoder; instead opting for a single architecture with inputs changing depending on the batch type provided
• Moved to off-graph summaries for 'secondary' ops to make it easier to build dynamic summaries, and avoid headaches with tf.data iterators while training
• Added an inhibition mechanism to the Sparse Convolutional Autoencoder, which can be enabled using hparams
• Added support for decoding middle layers using the secondary decoder method when working with composite components
• Small improvements and bugfixes, including: directory re-structuring and clean up, pylint warning fixes, code style fixes