Fastai

Bugs Squashed

- tensorboard and wandb can not access `smooth_loss` ([3131](https://github.com/fastai/fastai/issues/3131))

Breaking Changes

- Promote `NativeMixedPrecision` to default `MixedPrecision` (and similar for `Learner.to_fp16`); old `MixedPrecision` is now called `NonNativeMixedPrecision` ([3127](https://github.com/fastai/fastai/issues/3127))
- Use the new `GradientClip` callback instead of the `clip` parameter to use gradient clipping
- Adding a `Callback` which has the same name as an attribute no longer raises an exception ([3109](https://github.com/fastai/fastai/issues/3109))
- RNN training now requires `RNNCallback`, but does not require `RNNRegularizer`; `out` and `raw_out` have moved to `RNNRegularizer` ([3108](https://github.com/fastai/fastai/issues/3108))
- Call `rnn_cbs` to get all callbacks needed for RNN training, optionally with regularization
- replace callback `run_after` with `order`; do not run `after` cbs on exception ([3101](https://github.com/fastai/fastai/issues/3101))

New Features

- Add `GradientClip` callback ([3107](https://github.com/fastai/fastai/issues/3107))
- Make `Flatten` cast to `TensorBase` to simplify type compatibility ([3106](https://github.com/fastai/fastai/issues/3106))
- make flattened metrics compatible with all tensor subclasses ([3105](https://github.com/fastai/fastai/issues/3105))
- New class method `TensorBase.register_func` to register types for `__torch_function__` ([3097](https://github.com/fastai/fastai/issues/3097))
- new `dynamic` flag for controlling dynamic loss scaling in `NativeMixedPrecision` ([3096](https://github.com/fastai/fastai/issues/3096))
- remove need to call `to_native_fp32` before `predict`; set `skipped` in NativeMixedPrecision after NaN from dynamic loss scaling ([3095](https://github.com/fastai/fastai/issues/3095))
- make native fp16 extensible with callbacks ([3094](https://github.com/fastai/fastai/issues/3094))
- Calculate correct `nf` in `create_head` based on `concat_pool` ([3115](https://github.com/fastai/fastai/pull/3115)) thanks to [muellerzr](https://github.com/muellerzr)

New Features

- Small DICOM segmentation dataset ([3034](https://github.com/fastai/fastai/pull/3034)), thanks to [moritzschwyzer](https://github.com/moritzschwyzer)

Bugs Squashed

- `NoneType object has no attribute append` in fastbook chapter 6 BIWI example ([3091](https://github.com/fastai/fastai/issues/3091))

New Features

- Refactor MixUp and CutMix into MixHandler ([3037](https://github.com/fastai/fastai/pull/3037)), thanks to [muellerzr](https://github.com/muellerzr)
- Refactors into a general MixHandler class, with MixUp and CutMix simply implementing a `before_batch` to perform the data augmentation. See `fastai.callback.mixup`

Bugs Squashed

- Gradient Accumulation + Mixed Precision shows artificially high training loss ([3048](https://github.com/fastai/fastai/issues/3048))

New Features

Bugs Squashed

- Update for fastcore `negate_func`->`not_`
- LR too high for gradient accumulation ([3040](https://github.com/fastai/fastai/pull/3040)), thanks to [marii-moe](https://github.com/marii-moe)
- Torchscript transforms incompatibility with nn.Sequential ([2920](https://github.com/fastai/fastai/issues/2920))

New Features

- Pytorch 1.7 subclassing support ([2769](https://github.com/fastai/fastai/issues/2769))

Bugs Squashed

- unsupported operand type(s) for +=: 'TensorCategory' and 'TensorText' when using AWD_LSTM for text classification ([3027](https://github.com/fastai/fastai/issues/3027))
- UserWarning when using SaveModelCallback() on after_epoch ([3025](https://github.com/fastai/fastai/issues/3025))
- Segmentation error: no implementation found for 'torch.nn.functional.cross_entropy' on types that implement torch_function ([3022](https://github.com/fastai/fastai/issues/3022))
- `TextDataLoaders.from_df()` returns `TypeError: 'float' object is not iterable` ([2978](https://github.com/fastai/fastai/issues/2978))
- Internal assert error in awd_qrnn ([2967](https://github.com/fastai/fastai/issues/2967))