Version 1.0.0b1

What's Changed

• Refactor the configs to support self-superised pretraining and add torchvision backbones by @tigist-d in #43
• Fixes EMA update in fully supervised mode by @tigist-d in #44
• Improve logging support by @tigist-d in #45
• Refactor datasets by @tigist-d in #46
• Enforce a naturally sorted build order for loss heads by @tigist-d in #47
• Add optional kNN evaluation support by @tigist-d in #48
• Improve data loading with mix of small and large datasets by @tigist-d in #49
• Enforce naming conventions throughout the codebase by @tigist-d in #50
• Fix support for unbounded iterable datasets by @tigist-d in #51
• Add a class-weighted (e.g. balanced) dataset wrapper by @tigist-d in #52
• Refactor dataset and loss head modules by @tigist-d in #53
• Various fixes and improvements resulting from testing by @tigist-d in #54
• Add unit test suite by @tigist-d in #55
• Update documentation and enable automated builds using readthedocs by @tigist-d in #56
• Upgrade Python version and include sesemi as a requirement in the readthedocs build by @tigist-d in #57

Full Changelog: v0.3.0...v1.0.0b1

Version 0.3.0

Release Notes

• Support new dataset formats (CIFAR-10/100, STL, and more).
• Support pseudo-label dataset generation.
• Add new SSL methods including consistency regularization, FixMatch, and noisy student.
• Fixes and improvements.

What's Changed

• Ensures that the unsup subset is included in the configs using imagewang by @tigist-d in #25
• Add support for consistency regularization (Pi Model) by @vuptran in #27
• Add EMA consistency regularization (Mean Teacher) by @vuptran in #28
• Move EMA config into model config and include logs by @tigist-d in #30
• Add CIFAR-10, CIFAR-100, and STL-10 datasets and baseline models by @tigist-d in #29
• Add FixMatch loss head by @tigist-d in #31
• Add support for pseudo-labeled dataset generation by @tigist-d in #32
• Add noisy student training configurations by @tigist-d in #33
• Fix minor issues with docs by @tigist-d in #34

Full Changelog: v0.2.0...v0.3.0

Version 0.2.0

Release Notes

• Add new pretext methods including entropy minimization and jigsaw prediction along with various different combinations.
• Add support for shared heads and structure configurations hierarchical for ease of reuse.
• Fixes and improvements.

Version 0.1.1

Release Notes

• Mitigate issues when using torch version 1.9.
• Improve documentation.

Version 0.1.0

Release Notes

• Supports both DP and DDP training using PyTorch Lightning.
• Supports Hydra for configuration management.
• Contains pretext rotation prediction self-supervised regularization.