Upcoming Features

Documentation and Support Features

  • “Python API” overview of all components of the system
  • Improve CIFAR-10 example with a custom evaluation hook.

Model Features

  • More intuitive error messages when trying to assemble models.
  • Add “flatten” layers automatically before Dense layers.
  • Add “activation” layers automatically between layers.
  • Multi-model architectures, as used for GANs.

Container Features

  • Additional activation functions

Templating Features

  • Easily include model pieces from external sources (e.g., KurHub).

Loss/Optimizer Features

  • New optimizers
  • New loss functions

Data Features

  • More expressive data pipelines for creating derivative training sets, pre-processing data, and doing just-in-time manipulation of data.
  • Column mapping: assign a column of data from a file to a model container with a different name.

Log Features

  • Add a textual logger class.

Usage Features

  • Add “accuracy” statistic to the output during training/evaluation.
  • Allow command-line arguments to override the specification file.
  • Better detecting and handling unused keywords (which may have been typos).
  • Standalone “validate” command.

Internal Features

  • Review providers: should they always return dictionaries?