Mmcv

Bug Fixes

- Fix pybind exporting bug in fused_bias_leakyrelu and upfirdn2d (1005)

New Features

- Add Rotated ROI align op for pytorch and onnxruntime (cpu) (933)
- Support `Multi-Scale-DeformAttention` in `deformable-detr` (878)
- Support CPU training with MMDataParallel (972)
- Support `auto_fp16` using torch.cuda.amp when PyTorch >= 1.6.0 (951)
- Add PyTorch `Profiler` to analyze training process (937)

Improvements

- Remove redundant self.init_cfg (959)
- Refactor init cfg (958)
- Refactor readme.md (984)
- Use Sequential rather than nn.Sequential in build_model_from_cfg (992)
- Use len(result) as batch size in engine/test.py (987)
- Support clipping min_lr in StepLrUpdaterHook 995

Bug Fixes

- Fix compile without cuda (945)
- Rename init_weight to init_weights (971)
- Incorrect citation in transformer documentation (974)
- Fix data type in fused-bias-leakyrelu for apex fp16 training (981)
- Fix raise error bug in registering multiple names (949)

New Features

- Add OneCycleLrUpdaterHook (906)
- Add onnxruntime custom op grid_sample (916)
- Add `EvalHook` which will be used in downstream projects (739)
- Add `model registry` (760)

Improvements

- Add more detail of warning in basemodule (899)
- Remove deprecated onnx-simplify source code (927)
- Add total_steps attribute for OneCycleLrUpdaterHook (929)
- Add docstring for DeformConv (921)
- Revise init_weight in BaseModule (905)
- Add document for TensorRT and onnxruntime custom ops (920)
- Add init_cfg in transformer base classes (946)

Bug Fixes

- Fix initialization bug of override cfg (918)
- Fix typo in build_conv_layer docstring (928)
- Fix onnx interpolate conversion (917)
- Fix bug of `lr updater hook` (907)
- Fix wandb logger drop result bug (913)
- Revise override in init_cfg (930)
- Fix cuda compile error on visual studio 16.9 (891)
- Fix DeformConv2d bias error and add tests (940)
- Fix train error in EpochBasedRunner (942)

New Features

- Add cutout and lighting (909)
- Add `transformer base class` (892)
- Add Caffe2Xavier Initializer (902)
- Add cuda ops: UpFirDn2d and fused_bias_leakyrelu (900)
- Support to calculate FLOPs of GN, IN, LN (897)
- Add `Sequential` with init_weight (884)
- Add tifffile backend support for image io (868)
- Add `Deformable Conv2d` TensorRT Plugin (858)
- Support auto contrast (881)
- Add adjust sharpness (864)
- Add HTTP backend (860)
- Add NonMaxSuppression TensorRT Plugin (787)
- Add GELU activation function (843)
- Add readme of chinese version (846)

Improvements

- Revise function of layers and override keys in init_cfg (893)
- Revise function of is_init flag in BaseModule (898)
- Support revise_keys in load_checkpoint()(829)
- Support save last checkpoint (853)
- Update video/io.py (856)
- Add docstring for DistSamplerSeedHook (850)

Bug Fixes

- Fix bug of convmodule (889)
- Fix scatter in pytorch18 (882)
- Fix test unit of nms and batched_nms for tensorrt (872)
- Fix bugs in some photometric transforms (863)
- Fix pytorch2onnx failed for interpolate op with PyTorch==1.6.0 (848)
- Fix the iter error when the number of GPUs is different during resume (844)
- Fix typo in README (852)
- Fix document in cnn.md (841)

New Features

- Add initializers and BaseModule for unified parameter initialization (780, 825)
- Add ort nms op (803)

Improvements

- Add `register_itertimer_hook` function (838)
- Add more error info in `build_from_cfg` (834)
- Add `map_location` parameter for all checkpoint load functions (826)
- Add file mode parameter for logger to support customized mode (808)
- Polish CONTRIBUTING for all OpenMMLab projects (807)
- Add a short document for TensorRT plugins in mmcv (797)
- Refactor `_load_checkpoint` function (790)

Bug Fixes

- Fix `BaseRunner` hook registry bug (840)
- Fix no `torch.no_grad` decorator for val in `IterBasedRunner` (835)
- Fix weight shape mismatch bug of DCN when batch size larger than 64 (819)
- Fix `imshow_bboxes` with incontiguous img (823)
- Fix header file for std::iota(803) in nms op (820)
- Parse scales of list when PyTorch==1.7.0 (815)
- Fix misleading tutorial document in io.md (813)
- Hande exif orientation with `pillow` backend in `imread` (805)

New Features

- Add custom operator support for TensorRT. (786)
- Allow register multi-name for a module simultaneously. (775)
- Support load checkpoint from ceph. (778)
- Add common testing functions to OpenMMLab repos. (743)
- Add ScatterND TensorRT Plugin. (786)

Improvements

- Set installation of onnx simplify's related packages to optional. (799)
- Add ``torch.no_grad()`` decorator to the whole val workflow. (777)
- Add dataset classes name info to meta when saving checkpoints. (776)
- Support not to register lr scheduler hook. (774)

Bug Fixes

- Fix copy source file path for Windows. (796)