Adding new functionality

As S4TF has matured, we have adopted relatively careful Backwards Compatibility rules. In order to ensure APIs are well considered, most new functionality is added in the context of a representative model in the swift-models repository. After we have sufficient confidence in the API, it graduates into swift-apis to be incorporated in the next release.

There are a few exceptions to this guideline, such as:

• Changes to the implementation under Tensor, such as new runtimes.
• Adding standard neural network layers or mathematical/loss functions, such as those included with Keras, Flax, or Sonnet.