Deeplay

This release introduces several new features and improvements to enhance the functionality and performance of the project. New models include a revised Attention UNet and a Vision Transformer.

Applications can now be loaded directly from lightning checkpoints, using the class method `.load_from_checkpoint()`.

Blocks have been significantly revised, including the implementation of `Conv2dBlock`, `LinearBlock`, `SequenceBlock`, and more. These have special methods to apply specific configurations, like`.normalized()` and `.pooled()`.

Several bug fixes are included in this release, addressing various issues in GANs and resolving a problem with multi-configuration handling. Optimizers now have a more stable and easy-to-use creation process internally.

This release introduces enhancements including dictionary-based data loaders for better application compatibility, active learning functionalities, new transformer model features like LayerDropoutSkipNormalization and Positional Embedding Layers, deeptrack integration methods and classes, updates to the AdversarialStrategy class weights, default values for RecurrentModel, and a RichProgressBarTheme for progress bars, as well as various bug fixes.

Note that this release of deeplay is an initial release and may not be fully tested or documented. We welcome feedback and contributions to improve future versions.