Mmcv

New features

- Add seven detection3d related CUDA operators, which are `FurthestPointSample`, `KNN`, `GatherPoints`, `ThreeInterpolate`, `ThreeNN`, `AssignScoreWithK` and `RoIPointPool3d` (1337, 1354, 1338, 1355, 1357, 1358, )
- Add `Deformable Conv` CustomOp for onnxruntime (1343)

Improvements

- Switch the CUDA CI to the docker image provided by PyTorch (1376)
- Add code spell hook to pre-commit and fix typos (1384)
- Remove some obsolete code (1370)

Bug fixes

- Fix the weight initialization of deform conv (1262)
- Revise the unit test of `correlation` (1368)
- Fix the unrecognized symbol in `correlation.cpp` (1372)
- Fix a potential bug in `deprecated_api_warning` (1395)
- Add error handler when getting host failed (1374)
- Fix the infernence issue of `MultiScaleDeformableAttention` on CPU device (1223)
- Add a condition to disable invalid memory write by `furthest_sample_point` (1405)
- Reduce the test memory usage of `ms_deformable_attn` (1407)

Docs

- Add the Chinese documentation for `CNN`, `ONNX`, `TensorRT`, `Build mmcv-full from source` and `contributing` (1246, 1331, 1336, 1353, 1351)
- Add a note for `installing mmcv-full` (1320)

Notes

Since PyTorch 1.9 has been released, we plan to end the support for PyTorch 1.3 and 1.4 in future releases. We will still provide pre-compiled packages of these two versions for three months, but the reliability and technical support are not guaranteed.
In addition, starting from this version, we provide pre-compiled packages for **Python 3.9**.

Contributors

A total of 12 developers contributed to this release.

bofen97, grimoire, zhouzaida, ZwwWayne, MeowZheng, Fu0511, hhaAndroid, eugene123tw, DCNSW, del-zhenwu, ZhichengHuang, jshilong

New features

- Add two CUDA OPs `BallQuery` and `Correlation` (1332, 1361 )
- Support loading optical flow data from bytes (1362)
- Support empty tensor in MMSyncBN (1205)

Docs

- Replace the default theme of documentation with PyTorch Sphinx Theme (1321)
- Add the Chinese documentation for `registry.md`, `config.md`, `ops.md` and `visualization.md` (1204, 1349, 1350)

Notes

Since PyTorch 1.9 has been released, we plan to end the support for PyTorch 1.3 and 1.4 in future releases. We will still provide pre-compiled packages of these two versions for three months, but the reliability and technical support are not guaranteed.

Contributors

A total of 8 developers contributed to this release.

AluminiumOxide, MeowZheng, zhouzaida, grimoire, ZwwWayne, nbei, DCNSW, luopeichao

New features

- Add the CPU implementation of `DCN and Modulated DCN` (1278)
- Add the `revert_sync_batchnorm` function to convert SyncBN in any model to BN (1253)
- Add python3.9 in CI to verify the support for python3.9 (1291)

Improvements

- Reimplement `cc_attention` using pure PyTorch (1201)
- Enhance the output of exception (1309)
- Restrict the warning message (1267)

Bug fixes

- Fix missing `state_dict._metadata` when saving and loading checkpoints (1294)
- Fix typo about installation (1313)
- Fix sphinx version (1301)

Docs

- Add the Chinese documentation for `faq.md` (1252)

Notes

Since PyTorch 1.9 has been released, we plan to end the support for PyTorch 1.3 and 1.4 in future releases. We will still provide pre-compiled packages of these two versions for three months, but the reliability and technical support are not guaranteed.

Contributors

A total of 13 developers contributed to this release.

Leojc eugene123tw weiji14 mzr1996 zhouzaida yyz561 gaotongxiao EricKani MeowZheng ZwwWayne xvjiarui Junjun2016 grimoire

New features

- Add a `GradientCumulativeOptimizerHook` for training model with a larger batch size (1221)
- Add a `DefaultRunnerConstructor` for extending existing `Runner`(1296)
- Add `modulated_deform_conv` to support onnxruntime (1281)

Improvements

- Add mmcv namespace for custom onnx ops (1254)
- Compile mmcv with all available CPUs (1046)
- Support deprecation information in Config (1275)
- Add windows CI (1023)

Bug fixes

- Fix the nms implementation of CPU (1244)
- Remove the duplicate files in `mmcv/ops/csrc` (1284)
- Upgrade the version of ONNX Runtime (1290)

Docs

- Use myst-parser extension for sphinx and make the style more beautiful (1287)
- Add the Chinese documentation for `utils.md` and `runner.md` (1232, 1202)

Notes

Since PyTorch 1.9 has been released, we plan to end the support for PyTorch 1.3 and 1.4 in future releases. We will still provide pre-compiled packages of these two versions for three months, but the reliability and technical support are not guaranteed.

Contributors

A total of 11 developers contributed to this release.

innerlee, grimoire, zhouzaida, raoshenglong, jiang1997, ZhiyuanChen, mzr1996, PeiqiWang, hhaAndroid, ZwwWayne, densechen

Improvements

- `Refactor the directory of csrc` (1206)
- Add PR document (1209)
- Speed up mmcv import (1249)
- Do not depend on setuptools>=52 (1235)

Bug Fixes

- Fix the bug that the training log and evaluating log are mixed (1252)
- Fix dcn forward and backward when batch size is larger than im2col_step (1212)
- Fix potential bugs of basemodule when record the initilization information (1217)

Notes

Since PyTorch 1.9 has been released, we plan to end the support for PyTorch 1.3 and 1.4 in the future releases. We will still provide pre-compiled packages of these two versions for three months, but the reliability and technical support are not guaranteed.

Contributors

A total of 12 developers contributed to this release.

hellock, xvjiarui, jshilong, ZwwWayne, MeowZheng, innerlee, zhouzaida, kennymckormick, grimoire, nbei, dreamerlin, Junjun2016

New Features

- Register CrissCrossAttention into plugin layers (1189)
- Support custom hooks in IterBasedRunner (1193)
- Add ONNX export support to torch.roll (1194)
- Add `flat cosine lr updater` (1066)
- Add `logger for initialization of parameters` (1150)

Improvements

- Add `is_tracing` to wrap torch.jit.is_tracing in different versions of pytorch 1187
- Revise the interface of `upfirdn2d` function (1195)
- Improve digit_version (1185)
- Add TERMINOLOGY.md (895)

Bug Fixes

- Fix potential integer overflow in imequalize (1198)
- Fix floor/ceil error in deform_attn for windows (1037)
- Resolves the duplicate logs issue (1164)
- Fix example (1074)
- Prevent kernel_normalizer to change mask dtype (1210)

Notes

Since PyTorch 1.9 has been released, we plan to end the support for PyTorch 1.3 and 1.4 in the future releases. We will still provide pre-compiled packages of these two versions for three months, but the reliability and technical support are not guaranteed.

Contributors

A total of 19 developers contributed to this release.

yeliudev, 007gzs, wangg12, qizhuli, ddonatien, mzr1996, Junjun2016, grimoire, nbei, dreamerlin, jshilong, kennymckormick, zhouzaida, ZwwWayne, hellock, MeowZheng, innerlee, xvjiarui, RunningLeon