Fastai

New Features

- Support py3.10 annotations ([3601](https://github.com/fastai/fastai/issues/3601))

Bugs Squashed

- Fix pin_memory=True breaking (batch) Transforms ([3606](https://github.com/fastai/fastai/pull/3606)), thanks to [johan12345](https://github.com/johan12345)
- Add Python 3.9 to `setup.py` for PyPI ([3604](https://github.com/fastai/fastai/pull/3604)), thanks to [nzw0301](https://github.com/nzw0301)
- removes add_vert from get_grid calls ([3593](https://github.com/fastai/fastai/pull/3593)), thanks to [kevinbird15](https://github.com/kevinbird15)
- Making `loss_not_reduced` work with DiceLoss ([3583](https://github.com/fastai/fastai/pull/3583)), thanks to [hiromis](https://github.com/hiromis)
- Fix bug in URLs.path() in 04_data.external ([3582](https://github.com/fastai/fastai/pull/3582)), thanks to [malligaraj](https://github.com/malligaraj)
- Custom name for metrics ([3573](https://github.com/fastai/fastai/pull/3573)), thanks to [bdsaglam](https://github.com/bdsaglam)
- Update import for show_install ([3568](https://github.com/fastai/fastai/pull/3568)), thanks to [fr1ll](https://github.com/fr1ll)
- Fix Classification Interpretation ([3563](https://github.com/fastai/fastai/pull/3563)), thanks to [warner-benjamin](https://github.com/warner-benjamin)
- Updates Interpretation class to be memory efficient ([3558](https://github.com/fastai/fastai/pull/3558)), thanks to [warner-benjamin](https://github.com/warner-benjamin)
- Learner.show_results uses passed dataloader via dl_idx or dl arguments ([3554](https://github.com/fastai/fastai/pull/3554)), thanks to [warner-benjamin](https://github.com/warner-benjamin)
- Fix learn.export pickle error with MixedPrecision Callback ([3544](https://github.com/fastai/fastai/pull/3544)), thanks to [warner-benjamin](https://github.com/warner-benjamin)
- Fix concurrent LRFinder instances overwriting each other by using tempfile ([3528](https://github.com/fastai/fastai/pull/3528)), thanks to [warner-benjamin](https://github.com/warner-benjamin)
- Fix _get_shapes to work with dictionaries ([3520](https://github.com/fastai/fastai/pull/3520)), thanks to [ohmeow](https://github.com/ohmeow)
- Fix torch version checks, remove clip_grad_norm check ([3518](https://github.com/fastai/fastai/pull/3518)), thanks to [warner-benjamin](https://github.com/warner-benjamin)
- Fix nested tensors predictions compatibility with fp16 ([3516](https://github.com/fastai/fastai/pull/3516)), thanks to [tcapelle](https://github.com/tcapelle)
- Learning rate passed via OptimWrapper not updated in Learner ([3337](https://github.com/fastai/fastai/issues/3337))
- Different results after running `lr_find()` at different times ([3295](https://github.com/fastai/fastai/issues/3295))
- lr_find() may fail if run in parallel from the same directory ([3240](https://github.com/fastai/fastai/issues/3240))

New Features

- add `at_end` feature to `SaveModelCallback` ([3296](https://github.com/fastai/fastai/pull/3296)), thanks to [tmabraham](https://github.com/tmabraham)

Bugs Squashed

- fix fp16 test ([3284](https://github.com/fastai/fastai/pull/3284)), thanks to [tmabraham](https://github.com/tmabraham)

- Import `download_url` from fastdownload

Breaking changes

- `config.yml` has been renamed to `config.ini`, and is now in `ConfigParser` format instead of YAML
- THe `_path` suffixes in `config.ini` have been removed

Bugs Squashed

- Training with `learn.to_fp16(`) fails with PyTorch 1.9 / Cuda 11.4 ([3438](https://github.com/fastai/fastai/issues/3438))
- pandas 1.3.0 breaks `add_elapsed_times` ([3431](https://github.com/fastai/fastai/issues/3431))

New Features

- add DiceLoss ([3386](https://github.com/fastai/fastai/pull/3386)), thanks to [tcapelle](https://github.com/tcapelle)
- TabularPandas data transform reproducibility ([2826](https://github.com/fastai/fastai/issues/2826))

Bugs Squashed

- Latest Pillow v8.3.0 breaks conversion Image to Tensor ([3416](https://github.com/fastai/fastai/issues/3416))

Breaking changes

- QRNN module removed, due to incompatibility with PyTorch 1.9, and lack of utilization of QRNN in the deep learning community. QRNN was our only module that wasn't pure Python, so with this change fastai is now a pure Python package.

New Features

- Support for PyTorch 1.9
- Improved LR Suggestions ([3377](https://github.com/fastai/fastai/pull/3377)), thanks to [muellerzr](https://github.com/muellerzr)
- SaveModelCallback every nth epoch ([3375](https://github.com/fastai/fastai/pull/3375)), thanks to [KeremTurgutlu](https://github.com/KeremTurgutlu)
- Send self.loss_func to device if it is an instance of nn.Module ([3395](https://github.com/fastai/fastai/pull/3395)), thanks to [arampacha](https://github.com/arampacha)
- Batch support for more than one image ([3339](https://github.com/fastai/fastai/issues/3339))
- Changable tfmdlists for TransformBlock, Datasets, DataBlock ([3327](https://github.com/fastai/fastai/issues/3327))

Bugs Squashed

- convert TensorBBox to TensorBase during compare ([3388](https://github.com/fastai/fastai/pull/3388)), thanks to [kevinbird15](https://github.com/kevinbird15)
- Check if normalize exists on `_add_norm` ([3371](https://github.com/fastai/fastai/pull/3371)), thanks to [renato145](https://github.com/renato145)