v0.0.18

Adds the Lion Optimizer

Full Changelog: v0.0.17...v0.0.18

v0.0.17

Bug fixes and compatibility with fastai 2.7.11

Full Changelog: v0.0.16...v0.0.17

v0.016

Decrease Adan memory usage and increase optimizer step speed
Fix EMA callbacks not applying to buffers
EMA callback accepts epochs or percent of total training

Full Changelog: v0.0.15...v0.0.16

v0.0.15

Drop support for Python 3.7. CI tests on Python 3.8, 3.9, & 3.10.

ProgressiveResize can be imported independent of other fastxtend features.

Bug fixes

Full Changelog: v0.0.14...v0.0.15

v0.0.14

0.2-0.25x speed improvement to EMACallback

v.0.0.13

Updated EMACallback and EMAWarmupCallback with new features, including a fast fused implementation, EMA start delay, and EMA warmup.

Added two new schedulers: fit_flat_warmup and fit_cos_anneal.

Bug fixes.

Full Changelog: v0.0.12...v.0.0.13

v.0.0.12

Support PyTorch 1.13 via fastai 2.7.10

v0.0.11

What's Changed

• Add fused ForEach and TorchScript Optimizers
• Add the Adan Optimizer
• Port to nbdev2
• Improved Callbacks & Bug Fixes: EMA, ProgressiveResize, SimpleProfiler
• XResNet compatibility with TorchScript and vision_learner
• BCEWithLogitLoss with 'batchmean' reduction

Full Changelog: v0.0.10...v0.0.11

v0.0.10

• CutMix and CutMixAugment support small batch sizes
• Bug fix in CutMix and CutMixAugment

0.0.9

• Add ProgressiveResize callback to implement automatic progressive resizing in fastai
• Add support for element-wise MixUp, CutMix and Augmentations in CutMixUp and CutMixUpAugment
• Add samples per second to simple profiler, logging and output improvements
• Add MixUp support to MultiLoss
• Add new batch augmentations

Thanks to @marii-moe for assistance in debugging and fixing ProgressiveResize memory allocation overflows.