Mmcv

New Features

- Add custom op `roialign` for `onnxruntime`. (724)
- Add `onnx-simplify`. (751)


Bug Fixes

- Fix `MMDistributedDataParallel` when pytorch>=1.7. (769)

Improvements

- Unify installation instruction. (765)


Bug Fixes

- Remove ``pytest`` from the runtime dependencies. (763)

New Features

- Add custom operators support for onnxruntime. (612)
- Support mpi launcher. (726)


Improvements

- Add 'iof' mode and 'aligned' mode for ``box_iou_rotated``. (726, 753)
- Refactor ``mmcv-full`` installation to fit ``pip>=20.3`` and follow PEP440. (748)
- Add model links (MobileNet-V3, ResNet18-V1c). (755, 757)
- Allow to replace nested tuple and list via options in ``Config``. (731)
- Support mmcv bricks in flops compuation. (715)

Bug Fixes

- Fix ``mmcls`` related models and links. (741, 745)
- Fix grad_output uncontiguous. (737)
- Avoid recursion error when using customized norm layer without ``_abbr_``. (732)
- Minor bug fixes, docs, unit tests related to nms. (728, 733)
- Fix checkpoint key error. (722)
- Fix comments of ``init_weights`` in ``ConvModule`` (730)

New Features

- Add clamp activation layer. (685)

Improvements

- Refactor flow_warp and remove its C implementation. (707)
- LossScaler: Allow the initial value to be set in dynamic mode. (692)
- Add installation instructions for ``mmcv-full`` build on Windows. (663)
- Allow list index keys in ``Config.merge_from_dict``. (696)

Bug Fixes

- Check MASTER_ADDR in the environment variable before assignment. (705)
- Fix ``mmcv-full`` installation using pip 20.3. (703)

New Features

- Add ``box_iou_rotated``, ``ml_nms_rotated`` and ``nms_rotated``. (625)
- Add wrappers for ``Conv3d``, ``MaxPool3d`` and ``ConvTranspose3d``. (652)
- Add onnx support to ``roi_align`` and ``roi_pool``. (593)
- Add more transforms: LUT(646) and CLAHE(647).

Improvements

- Better support of pytorch2onnx for detection models. (648)
- Add documentation for registry. (656)
- Refactor logger. (659)
- Enhance ``hsigmoid``: support more parameters. (657)

Bug Fixes

- Fix rotated NMS. (668)

New Features

- Support PyTorch 1.7. (631)
- Add dynamic scale that manages loss scaling in mixed precision training. (585)


Improvements

- Add hook message for checkpoint hook (635)


Bug Fixes

- Fix bug in logger hooks. (634, 637)
- Update ``lr_updater`` to correct period index computation. (574)
- Fix missing arguments when converting DCN to ONNX. (624)