V2.3.0-fix

Updated version number in setup.
Changelog see Release 2.3.0.

V2.3.0

• Datasets
  • Added sample_tokens to kitti and nuscenes implementations
  • Added 2d boxes to nuScenes (generated from 3D boxes)
  • Bugfix: NuImages can now load images and 2d boxes (semseg still untested)
• Model
  • Names in sequential layers can be named using kwargs
  • Added a parallel layer
  • Bugfix: Some convolution parameters were unsupported
• Core
  • Filelog level can be specified

V2.2.0

• Datasets
  • Renamed transformers to transforms in the datapipeline to avoid confusion with Transformers a component of neural networks originally used in NLP.
  • Refactoring/renaming of datasets
  • Added NuScenes and KITTI Dataset implementations.
  • Added a MultiDataset to allow merging datasets.
• Core
  • Improved Logging
  • Added a capture method for tensors, so that getting tensors from within a model is possible.
  • Removed Module and Loss Registry, as the DictLike fills that already and is much easier (to maintain).
• Bugfixes

Version 2.1.0

Changes:

• Losses no longer need parent argument in init. (Automatically taken care of via global variable, as only a logger was required from that.)
• Experiment requires DictLike, to reduce complexity.

Version 2.0.0

Changes:

• Significant rework of the experiment API.
  • Inheriting only required for very complex cases.
  • Most cases can just call the __init__ and run_training/run_inference.
• No more CLI tools, they were pretty useless anyways.
  • You can just create your own main
  • Or use a jupyter notebook

V1.1.5

• Added mode to experiment initializer to allow for different initializations in inference or testing.
• Minor bugfixes.

Version 1.1.4

• Support for indexed tensors
• Indexed tensors for object detection
• Batchsize > 1 FasterRCNN Demo
• Added manifest to include jsons in distributable

Version 1.1.3

leanai_remote --host=localhost does not use ssh anymore.

Version 1.1.2

• Added drawing of graph and printing summary at beginning of training
• Added missing layers and fixed bugs in existing ones
• Removed redundant vgg16bn implementation
• Improved debugging for json defined models

Version 1.1.1

Allowed parsers to be getters.