Mmcv

New Features

- Support building MMCV with `ROCm` (1022)
- Support tensorrt custom plugin MMCVCornerPool (1179)

Improvements

- Use LooseVersion for version checking (1158)
- Refactor structure of documentation (1102)
- Build `Chinese documentation` (1073)

Bug Fixes

- Fix SyncBN in PyTorch 1.9 (1138)
- Fix saconv in PyTorch 1.9 (1147)
- Fix unittest in PyTorch 1.9 (1146)
- Fix missing check of directory in scandir (1110)
- `Change dict update order to fix bug of resumed checkpoint` (1108)

New Features

- Add TensorRT support to modulated deformable conv ops (1078)
- Support `variables in base files for configs` (1083)

Improvements

- Delete the warning report of BaseInit (1126)
- Add to_ntuple() helper function (1125)
- Support resize or rescale an image to multiple (1121)
- Refine the EvalHook to meet the demand for the downstream codebases (1076)
- Refine default hooks and custom hooks priority rank (1120)
- Refine the PyTorch CUDA implementation for Criss Cross Attention (1143)
- Support image reading while ignoring EXIF orientation info (1091)
- Support print hooks before running (1123)

Bug Fixes

- Fix the dimension of getOutputDimensions (1117)
- Fix the backward-compatible of empty tensor inference (1131)
- Fix the permission denied error on windows (1077)

Improvements

- Add a return value in TextLoggerHook.log() (1040)

Bug Fixes

- Avoid bc-breaking when importing MultiScaleDeformableAttention (1100)

New Features

- Add `DvcliveLoggerHook` (1075)
- Enable exporting to ONNX for `PointRend` (953)
- Support resuming loss scaler for FP16 training (1013)

Backwards Incompatible Changes

- In PR 978, `MultiScaleDeformableAttention` has been moved to mmcv/ops/multi_scale_deform_attn.py. So the import source has been changed from `mmcv.cnn.bricks.transformer` to `mmcv.ops.multi_scale_deform_attn`. See the example of migration in https://github.com/open-mmlab/mmdetection/pull/5338.

Improvements

- Add segmentation keys for `greater_keys` in EvalHook (1060)
- Improve windows support for `list_from_file()` (1043)
- Add runner tutorial (1082)
- Refactor the base class related to transformer (978)
- Add support for score_threshold and max_num for NMS operation when exporting to ONNX (957)
- Support using the name of the base classes in `init_cfg` (1057)

Bug Fixes

- Fix link of checkpoints for mmcls (1067)
- Fix the docstring for initializers (1071)

New Features

- Add `NeptuneLoggerHook` (1025)
- Add truncated normal weight init (935)
- Add cummax/cummin tensorrt plugin (1031)
- Add cuda version BorderAlign module (1021)
- Add TensorRT InstanceNormalization plugin (1034)

Improvements

- Support amp (pytorch >= 1.6.0) on DCN and DCNv2 (1029)

Bug Fixes

- Fix a bug when two or more loggers were set (1015)
- Fix generalized attention fp16 (1036)

New Features

- Support corner_pool related custom operators for onnxruntime (997)
- Support custom operators cummax and cummin for onnxruntime (1010)
- Add StepMomentumUpdaterHook (1011)
- Add pixel group and contour expand ops (993)
- Add `custom hook` by config file (970)

Improvements

- Add test util for checking stand-alone python scripts (1007)
- Add triangular policy in CyclicLrUpdaterHook (1001)
- Load hook_msgs when resume checkpoint (962)
- Update weight initialization in cnn.md (912)

Bug Fixes

- Fix `DCN fp16` (1014)
- Fix basemodule (1018)
- Fix softnms (1019)
- Fix wrong iter number and progress number in the logging during val/test time 914
- Read in rb mode then decode (983)