Skip to content

MMCV Releases V2.0.0
Compare
Choose a tag to compare
@zhouzaida zhouzaida released this 06 Apr 03:33
· 1 commit to 2.x since this release
v2.0.0
1db3967
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.

Highlights

Introducing mmcv v2.0.0, we have made significant updates to the package. Training process-related modules, such as Runner, Hook, and Parallel, have been removed and are now provided by MMEngine. With this release, we have added numerous data transformations, including CenterCrop. Additionally, starting from mmcv v2.0.0, we are providing pre-built packages for PyTorch v2.0.0.

Features

  • Introduced a variety of data transformations
  • Added StyleGAN3 Ops (#2290)
  • Implemented BezierAlign CUDA op (#2393)
  • Added box_iou_quadri & nms_quadri (#2277)
  • Introduced stack ball query and stack group points ops (#2292)
  • Supported LayerScale in FFN (#2451)
  • Added C++ implementation for bbox_overlaps (#2477)

Enhancements

  • Adopted the new fileio interface from mmengine (#2468)
  • Enabled MultiScaleDeformableAttention with AMP (#2541)
  • Registered Compose to TRANSFORMS (#2600)
  • Supported value_proj_ratio in MultiScaleDeformableAttention (#2452)

Backwards Incompatible Changes

  • Removed mmcv.fileio (#2179). The new fileio interfaces can be found at mmengine.fileio
  • Removed mmcv.runner, mmcv.parallel, mmcv.engine and mmcv.device (#2216)
  • Removed all classes in mmcv.utils (eg Config and Registry) and many functions. Only a few functions related to mmcv are reserved (#2217). The new Config and Registry can be found at mmengine.registry and mmengine.config
  • Removed mmcv.onnex, mmcv.tensorrt modules, and related functions (#2225)
  • Dropped support for Python versions < 3.7 (#2525)
  • Dropped support for PyTorch versions < 1.8 and CUDA versions < 10.1
  • Renamed the package name mmcv to mmcv-lite and mmcv-full to mmcv. Also, changed the default value of the environment variable MMCV_WITH_OPS from 0 to 1 (#2235)
MMCV < 2.0 MMCV >= 2.0 
pip install openmim

# Contains ops, because the highest version of mmcv-full
# is less than 2.0.0, so there is no need to add version
# restrictions
mim install mmcv-full

# do not contain ops
mim install "mmcv < 2.0.0"
 
pip install openmim

# Contains ops
mim install mmcv

# Ops are not included, because the starting version of
# mmcv-lite is 2.0.0, so there is no need to add version
# restrictions
mim install mmcv-lite

Bug Fixes

  • Improved compatibility of training stylegan2 (#2694)
  • Replace torch/extension.h with torch/types.h to fix compilation error on Windows (#2698)
  • Fixed torch2.0 dcn/mdcn symbolic (#2695)
  • Skipped fused_lrelu op when gcc is less than 6.0 or CUDA is less than 10.2 (#2671)
  • Skipped filtered_lrelu ut when CUDA is less than 10.2 (#2677)
  • Called operator __half explicitly (#2608)
  • Ensured fileio related interfaces maintain backward compatibility (#2539)
  • Fixed CI failure due to protobuf version (#2529)
  • Corrected _flip_keypoint result of RandomFlip (#2527)
  • Fixed Windows OS version to 2019 (#2503)
  • Resolved optimizer 'ValueError' when using PReLU activation (#2444)
  • Fixed improper assertion in RandomChoiceResize (#2450)
  • Fixed Correlation op (#2274)
  • Fixed roi align symbolic for torch>=1.13 (#2443)
  • Added swap_label_pairs in RandomFlip (#2332)
  • Fixed MPS ops for torch1.12+ (#2302)
  • Corrected calculation error of out_w in MaskedConv2dFunction (#2264)
  • Cast mask type to enable training with amp (#2220)

Documentations

  • Fixed torch2.0 dcn/mdcn symbolic (#2695)
  • Skipped fused_lrelu op when gcc is less than 6.0 or CUDA is less than 10.2 (#2671)
  • Skipped filtered_lrelu ut when CUDA is less than 10.2 (#2677)
  • Called operator __half explicitly (#2608)
  • Used official isort (#2575)
  • Ensured fileio related interfaces maintain backward compatibility (#2539)
  • Fixed CI failure due to protobuf version (#2529)
  • Corrected _flip_keypoint result of RandomFlip (#2527)
  • Fixed Windows OS version to 2019 (#2503)
  • Resolved optimizer 'ValueError' when using PReLU activation (#2444)
  • Fixed improper assertion in RandomChoiceResize (#2450)
  • Fixed Correlation op (#2274)
  • Fixed roi align symbolic for torch>=1.13 (#2443)
  • Added swap_label_pairs in RandomFlip (#2332)
  • Fixed MPS ops for torch1.12+ (#2302)
  • Corrected calculation error of out_w in MaskedConv2dFunction (#2264)
  • Cast mask type to enable training with amp (#2220)

Contributors

@HAOCHENYE @grimoire @zhouzaida @nijkah @BainOuO @enemy1205 @C1rN09 @triple-Mu @okotaku @vansinhu @Harold-lkk @KeiChiTse @xiexinch @xin-li-67 @VVsssssk @LYMDLUT @MeowZheng @yxzhao2022 @mzr1996 @ZwwWayne @YuanLiuuuuuu @ly015 @plyfager @GT9505 @zytx121 @hhaAndroid @fangyixiao18 @JingweiZhang12 @xcnick @zengyh1900

Contributors

xcnick, grimoire, and 28 other contributors
Assets 2
Loading