0.6.1

• Set persistent_workers=True as default in train data loaders.
• Improved documentation.
• Removed deprecated fields in configs.

0.6.0

Major rework from 0.5.x

• Added YAML config file support.
• Added option to pass variables to jinja2 when parsing config files.
• Major rewrite of pipelines to make train.py unnecessary.
• Migrated docs to mkdocs + readthedocs.
• Implemented multi-tracker support. Added MLFlow tracking.
• Added support for multiple optimizers targeting different parameters.
• Improved object instantiation system in configs to support functions/lambdas, recursive instantiation.
• Better support for regional compilation with torch.compile + accelerate.

0.5.3

• Add trainer_loss_transforms to SupervisedPipeline.
• Add trainer_loss_transforms and trainer_disc_loss_transforms to GANPipeline.

0.5.2c

• Fix NaN metric checkpoints taking priority over non-NaN ones.
• Upgrade deprecated Github actions.

0.5.2b

• Fix bug in image logger when x, y or y_pred is not a tensor.
• Add short section on metrics to docs.

0.5.2

• Add better support for custom data loaders.
• Make "val" dataset optional in SupervisedPipeline and GANPipeline. Model validation is skipped when not provided.

0.5.1

• Add logging of learning rate to supervised and GAN pipelines.
• Forbid extra fields in configs.
• Add additional meta field to Config class to allow additional information.
• Fix wrong total_iters value when creating linear LR scheduler.

0.5.0

This release removes PyTorch Ignite as a dependency and replaces it with a custom engine/event system.

This change was made due to conflicts between Ignite and Accelerate.

Some features have not yet been ported over from 0.3.x/0.4.x.

0.4.0.dev2

• Fixed bug when loading checkpoint with --checkpoint-keys argument.

0.4.0.dev1

• Integrated Accelerate
• Added name generation to avoid reliance on wandb
• Misc. bug fixes