The first release of the new backwards-incompatible API.
Major Changes
* The main difference is the new `protoflow.layers.prototypes` which makes the
distance layers obsolete. This makes implementing Siamese architectures much
easier. Distances are however still available as functions.
* The second big change is that `protoflow.applications` are now written by
subclassing `tf.keras.Model`. This makes the code a lot cleaner. However,
model serialization is no longer possible. The work around is to only save the
weights with `model.save_weights()`. Instead of deserializing the entire
model, simply create a new model and load the saved the weights.
* In order to highlight the breaking changes in the documentation, there is a
new logo.
New Features
* New benchmark and toy datasets are now available under `protoflow.datasets`.
Release 0.2.x
Bug fixes to the legacy API. No new features will be added to 0.2.x.