v0.4.0

Merge pull request #26 from andrei-papou/rc-v0.4.0

RC v0.4.0

v0.3.0

Changes:

• fixed memory leak when using iter submodule.
• WAndBLogger enhancements.
• Moved criteria of the metric to the metric class itself.
• Introduced DatasetKind enum.
• Added custom sampler option to the FullCycleTrainer.
• Code quality is now validated by mypy and flake8.

v0.2.1

Changes:

• Fixed FP16 implementation
• Tensors are not detach-ed before being passed to metrics.

v0.2.0

Changes:

• Fixed #12 by creating a separate collator for labeled dataset item.
• Added AWP and FGM.
• Added hook system for StandardIterationTrainer.
• Added BCEWithLogitsLoss.
• Added MulticlassMajorityVotingEnsemblingStrategy.
• Minor fixes and improvements.

v0.1.8

Changes

• Added standard initialization for torch.nn.Module and torch.nn.Embedding.

v0.1.7

Changes:

• _before_optimizer_step invocation of StandardIterationTrainer is now aligned with gradient accumulation.

v0.1.6

Changes:

• Added x and y args to _before_optimizer_step of StandardIterationTrainer

v0.1.5

Changes:

• StandardIterationTrainer now provides _before_optimizer_step hook
• StandardIterationTrainer now supports max_grad_norm argument for gradient clipping.

v0.1.4

Changes:

• added FuncList feature generator.

v0.1.3

Changes:

• removed all references of t.TypeAlias