MMEngine Release V0.8.3

v0.8.3 (31/07/2023)

Highlights

• Support enabling efficient_conv_bn_eval for efficient convolution and batch normalization. See save memory on gpu for more details
• Add Llama2 finetune example
• Support multi-node distributed training with MLU backend

New Features & Enhancements

• Enable efficient_conv_bn_eval for memory saving convolution and batch normalization by @youkaichao in #1202, #1251 and #1259
• Add Llama2 example by @HAOCHENYE in #1264
• Compare the difference of two configs by @gachiemchiep in #1260
• Enable explicit error for deepspeed not installed by @Li-Qingyun in #1240
• Support skipping initialization in BaseModule by @HAOCHENYE in #1263
• Add parameter save_begin to control when to save checkpoints by @KerwinKai in #1271
• Support multi-node distributed training with MLU backend by @josh6688 in #1266
• Enhance error message thrown by Config, build function and ConfigDict.items by @HAOCHENYE in #1272, #1270 and #1088
• Add the loop_stage runtime information in message_hub by @zhouzaida in #1277
• Fix Visualizer that built vis_backends will not be used when save_dir is None by @Xinyu302 in #1275

Bug fixes

• Fix scalar check in RuntimeInfoHook by @i-aki-y in #1250
• Move data preprocessor to target device in FSDPStrategy by @HAOCHENYE in #1261

Docs

• Add ecosystem in README by @zhouzaida in #1247
• Add short explanation about registry scope by @mmeendez8 in #1114
• Add the data flow of Runner in README by @zhouzaida in #1257
• Introduce how to customize distributed training settings @zhouzaida in #1279

New Contributors

• @youkaichao made their first contribution in #1202
• @mmeendez8 made their first contribution in #1114
• @Xinyu302 made their first contribution in #1275

Full Changelog: v0.8.2...v0.8.3

MMEngine Release V0.8.2

Bug fixes

• Fix pickling the Python style config by @HAOCHENYE in #1241
• Fix the logic of setting lazy_import by @Li-Qingyun in #1239

New Contributors

• @Li-Qingyun made their first contribution in #1239

Full Changelog: v0.8.1...v0.8.2

MMEngine Release V0.8.1

New Features & Enhancements

• Accelerate Config.dump and support converting Lazyxxx to string in ConfigDict.to_dictby @HAOCHENYE in #1232

Bug fixes

• FSDP should call _get_ignored_modules by @HAOCHENYE in #1235

Docs

• Add a document to introduce how to train a large model by @zhouzaida in #1228

Full Changelog: v0.8.0...v0.8.1

MMEngine Release V0.8.0

v0.8.0 (07/03/2023)

Highlights

• Support training with FSDP and DeepSpeed. Refer to the example for more detailed usages.

• Introduce the pure Python style configuration file:

  • Support navigating to base configuration file in IDE
  • Support navigating to base variable in IDE
  • Support navigating to source code of class in IDE
  • Support inheriting two configuration files containing the same field
  • Load the configuration file without other third-party requirements

  Refer to the tutorial for more detailed usages.

  

New Features & Enhancements

• Support training with FSDP by @HAOCHENYE in #1213
• Add FlexibleRunner and Strategies, and support training with DeepSpeed by @zhouzaida in #1183
• Support pure Python style configuration file by @HAOCHENYE in #1071
• Learning rate in log can show the base learning rate of optimizer by @AkideLiu in #1019
• Refine the error message when auto_scale_lr is not set correctly by @alexander-soare in #1181
• WandbVisBackend supports updating config by @zgzhengSEU in #977

Bug fixes

• CheckpointHook should check whether file exists before removing it by @zhouzaida in #1198
• Fix undefined variable error in Runner by @HAOCHENYE in #1219

Docs

• Add a document to introduce how to debug with vscode by @zhouzaida in #1212
• Update English introduction by @evdcush in #1189
• Fix parameter typing error in document by @syo093c in #1201
• Fix gpu collection during evaluation by @edkair in #1208
• Fix a comment in runner tutorial by @joihn in #1210

New Contributors

• @alexander-soare made their first contribution in #1181
• @zgzhengSEU made their first contribution in #977
• @AkideLiu made their first contribution in #1019
• @syo093c made their first contribution in #1201
• @edkair made their first contribution in #1208
• @joihn made their first contribution in #1210

Full Changelog: v0.7.4...v0.8.0

MMEngine Release V0.7.4

v0.7.4 (06/03/2023)

Highlights

• Support using ClearML to record experiment data
• Add Sophia optimizers

New Features & Enhancements

• Add visualize backend for clearml by @gachiemchiep in #1091
• Support Sophia optimizers by @zhouzaida in #1170
• Refactor unittest syncbuffer by @HAOCHENYE in #813
• Allow ann_file, data_root is None for BaseDataset by @HAOCHENYE in #850
• Enable full precision training on Ascend NPU by @Ginray in #1109
• Creating a text classification example by @TankNee in #1122
• Add option to log selected config only by @KickCellarDoor in #1159
• Add an option to control whether to show progress bar in BaseInference by @W-ZN in #1135
• Support dipu device by @CokeDong in #1127
• Let unit tests not affect each other by @zhouzaida in #1169
• Add support for full wandb's define_metric arguments by @i-aki-y in #1099

Bug fixes

• Fix the incorrect device of inputs in get_model_complexity_info by @CescMessi in #1130
• Correctly saves _metadata of state_dict when saving checkpoints by @Bomsw in #1131
• Correctly record random seed in log by @Shiyang980713 in #1152
• Close MLflowVisBackend only if active by @zimonitrome in #1151
• Fix ProfileHook cannot profile ddp-training by @HAOCHENYE in #1140
• Handle the case for Multi-Instance GPUs when using cuda_visible_devices by @adrianjoshua-strutt in #1164
• Fix attribute error when parsing CUDA_VISIBLE_DEVICES in logger @Xiangxu-0103 in #1172

Docs

• Translate infer.md by @Hongru-Xiao in #1121
• Fix a missing comma in tutorials/runner.md by @gy-7 in #1146
• Fix typo in comment by @YQisme in #1154
• Translate data_element.md by @xin-li-67 in #1067
• Add the usage of clearml by @zhouzaida in #1180

New Contributors

• @CescMessi made their first contribution in #1130
• @Bomsw made their first contribution in #1131
• @Hongru-Xiao made their first contribution in #1121
• @TankNee made their first contribution in #1122
• @W-ZN made their first contribution in #1135
• @gy-7 made their first contribution in #1146
• @YQisme made their first contribution in #1154
• @Shiyang980713 made their first contribution in #1152
• @KickCellarDoor made their first contribution in #1159
• @CokeDong made their first contribution in #1127
• @zimonitrome made their first contribution in #1151
• @adrianjoshua-strutt made their first contribution in #1164
• @gachiemchiep made their first contribution in #1091
• @i-aki-y made their first contribution in #1099

Full Changelog: v0.7.3...v0.7.4

MMEngine Release V0.7.3

What's Changed

New Features & Enhancements

• Add MLflowVisBackend by @sh0622-kim in #878
• Support customizing worker_init_fn in dataloader config by @shufanwu in #1038
• Make the parameters of get_model_complexity_info() friendly by @sjiang95 in #1056
• Add torch_npu optimizer by @luomaoling in #1079
• Support registering callable objects @C1rN09 in #595
• Complement type hint of get_model_complexity_info() by @sjiang95 in #1064
• MessageHub.get_info() supports returning a default value by @enkilee in #991
• Refactor logger hook unit test by @HAOCHENYE in #797
• Support BoolTensor and LongTensor on Ascend NPU by @Ginray in #1011
• Remove useless variable declaration by @HAOCHENYE in #1052
• Enhance the support for MLU device by @josh6688 in #1075
• Support configuring synchronization directory for BaseMetric by @HAOCHENYE in #1074
• Support accepting multiple input_shape for get_model_complexity_info by @sjiang95 in #1065
• Enhance docstring and error catching in MessageHub by @HAOCHENYE in #1098
• Enhance the efficiency of Visualizer.show by @HAOCHENYE in #1015
• Update repo list by @HAOCHENYE in #1108
• Enhance error message during custom import by @HAOCHENYE in #1102
• Support _load_state_dict_post_hooks in load_state_dict by @mzr1996 in #1103

Bug fixes

• Fix publishing multiple checkpoints when using multiple GPUs by @JunweiZheng93 in #1070
• Fix error when log_with_hierarchy is True by @HAOCHENYE in #1085
• Call SyncBufferHook before validation in IterBasedTrainLoop by @Luo-Yihang in #982
• Fix the resuming error caused by HistoryBuffer by @HAOCHENYE in #1078
• Failed to remove the previous best checkpoints by @HAOCHENYE in #1086
• Fix using incorrect local rank by @C1rN09 in #973
• No training log when the num of iterations is smaller than the default interval by @shufanwu in #1046
• collate_fn could not be a function object by @zhouzaida in #1093
• Fix optimizer.state could be saved in cuda:0 by @HAOCHENYE in #966
• Fix building unnecessary loop during train/test/val by @HAOCHENYE in #1107

Docs

• Introduce the use of wandb and tensorboard by @zhouzaida in #912
• Translate tutorials/evaluation.md by @LEFTeyex in #1053
• Translate design/evaluation.md by @zccjjj in #1062
• Fix three typos in runner by @jsrdcht in #1068
• Translate migration/hook.md to English by @SheffieldCao in #1054
• Replace MMCls with MMPretrain in docs by @zhouzaida in #1096

New Contributors

• @sh0622-kim made their first contribution in #878
• @Ginray made their first contribution in #1011
• @shufanwu made their first contribution in #1038
• @sjiang95 made their first contribution in #1056
• @JunweiZheng93 made their first contribution in #1070
• @SheffieldCao made their first contribution in #1054
• @jsrdcht made their first contribution in #1068
• @josh6688 made their first contribution in #1075
• @Luo-Yihang made their first contribution in #982
• @zccjjj made their first contribution in #1062

Full Changelog: v0.7.2...v0.7.3

MMEngine Release V0.7.2

v0.7.2 (04/06/2023)

Bug fixes

• Align the evaluation result in log by @kitecats in #1034
• Update the logic to calculate the repeat_factors in ClassBalancedDataset by @BIGWangYuDong in #1048
• Initialize sub-modules in DistributedDataParallel that define init_weights during initialization by @HAOCHENYE in #1045
• Refactor checkpointhook unittest by @HAOCHENYE in #789

New Contributors

• @kitecats made their first contribution in #1034

Full Changelog: v0.7.1...v0.7.2

MMEngine Release V0.7.1

v0.7.1 (04/03/2023)

Highlights

• Support compiling the model and enabling mixed-precision training at the same time
• Fix the bug where the logs cannot be properly saved to the log file after calling torch.compile

New Features & Enhancements

• Add mmpretrain to the MODULE2PACKAGE. by @mzr1996 in #1002
• Support using get_device in the compiled model by @C1rN09 in #1004
• Make sure the FileHandler still alive after torch.compile by @HAOCHENYE in #1021
• Unify the use of print_log and logger.info(warning) by @LEFTeyex in #997
• Publish models after training if published_keys is set in CheckpointHook by @KerwinKai in #987
• Enhance the error catching in registry by @HAOCHENYE in #1010
• Do not print config if it is empty by @zhouzaida in #1028

Bug fixes

• Fix there is no space between data_time and metric in logs by @HAOCHENYE in #1025

Docs

• Minor fixes in EN docs to remove or replace unicode chars with ascii by @evdcush in #1018

New Contributors

• @evdcush made their first contribution in #1018
• @KerwinKai made their first contribution in #987

Full Changelog: v0.7.0...v0.7.1

MMEngine Release V0.7.0

v0.7.0 (03/16/2023)

Highlights

• Support PyTorch 2.0! Accelerate training by compiling models. See the tutorial Model Compilation for details
• Add EarlyStoppingHook to stop training when the metric does not improve

New Features & Enhancements

• Add configurations to support torch.compile in Runner by @C1rN09 in #976
• Support EarlyStoppingHook by @nijkah in #739
• Disable duplicated warning during distributed training by @HAOCHENYE in #961
• Add FUNCTIONS root Registry by @HAOCHENYE in #983
• Save the "memory" field to visualization backends by @enkilee in #974
• Enable bf16 in AmpOptimWrapper by @C1rN09 in #960
• Support writing data to vis_backend with prefix by @HAOCHENYE in #972
• Support exporting logs of different ranks in debug mode by @HAOCHENYE in #968
• Silence error when ManagerMixin built instance with duplicate name. by @HAOCHENYE in #990

Bug fixes

• Fix optim_wrapper unittest for pytorch < 1.10.0 by @C1rN09 in #975
• Support calculating the flops of matmul with single dimension matrix by @HAOCHENYE in #970
• Fix repeated warning by @HAOCHENYE in #992
• Fix lint by @zhouzaida in #993
• Fix AMP in Ascend and support using NPUJITCompile environment by @luomaoling in #994
• Fix inferencer gets wrong configs path by @HAOCHENYE in #996

Docs

• Translate "Debug Tricks" to English by @enkilee in #953
• Translate "Model Analysis" document to English by @enkilee in #956
• Translate "Model Complexity Analysis" to Chinese. by @VoyagerXvoyagerx in #969
• Add a document about setting interval by @YuetianW in #964
• Translate "how to set random seed" by @xin-li-67 in #930
• Fix typo by @zhouzaida in #965
• Fix typo in hook document by @acdart in #980
• Fix changelog date by @HAOCHENYE in #986

New Contributors

• @YuetianW made their first contribution in #964
• @enkilee made their first contribution in #953
• @acdart made their first contribution in #980
• @VoyagerXvoyagerx made their first contribution in #969

Full Changelog: v0.6.0...v0.7.0

MMEngine Release V0.6.0

v0.6.0 (02/24/2023)

Highlights

• Support Apex with ApexOptimWrapper
• Support analyzing model complexity.
• Add Lion optimizer.
• Support using environment variable in the config file.

New Features & Enhancements

• Support model complexity computation by @tonysy in #779
• Add Lion optimizer by @zhouzaida in #952
• Support using environment variable in config file. by @jbwang1997 in #744
• Improve registry infer_scope by @zhouzaida in #334
• Support configuring timeout in dist configuration by @apacha in #877
• Beautify the print result of the registry by @Eiuyc in #922
• Refine the style of table by @zhouzaida in #941
• Refine the repr of Registry by @zhouzaida in #942
• Feature NPUProfilerHook by @luomaoling in #925
• Refactor hooks unittest by @HAOCHENYE in #946
• Temporarily fix collect_env raise errors and stops programs by @C1rN09 in #944
• Make sure Tensors to broadcast is contiguous by @XWHtorrentx in #948
• Clean the UT warning caused by pytest by @zhouzaida in #947

Bug fixes

• Backend_args should not be modified by get_file_backend by @zhouzaida in #897
• Support update np.ScalarType data in message_hub by @HAOCHENYE in #898
• Support rendering Chinese character in Visualizer by @KevinNuNu in #887
• Support Apex with ApexOptimWrapper by @xcnick in #742
• Fix the bug of DefaultOptimWrapperConstructor when the shared parameters do not require the grad by @HIT-cwh in #903
• Support model complexity computation by @tonysy in #779

Docs

• Add the document for the transition between IterBasedTraining and EpochBasedTraining by @HAOCHENYE in #926
• Introduce how to set random seed by @zhouzaida in #914
• Count FLOPs and parameters by @zhouzaida in #939
• Enhance README by @Xiangxu-0103 in #835
• Add a document about debug tricks by @zhouzaida in #938
• Refine the format of changelog and visualization document by @zhouzaida in #906
• Move examples to a new directory by @zhouzaida in #911
• Resolve warnings in sphinx build by @C1rN09 in #915
• Fix docstring by @zhouzaida in #913
• How to set the interval parameter by @zhouzaida in #917
• Temporarily skip errors in building pdf docs at readthedocs by @C1rN09 in #928
• Add the links of twitter, discord, medium, and youtube by @vansin in #924
• Fix typo shedule by @Dai-Wenxun in #936
• Fix failed URL by @zhouzaida in #943

New Contributors

• @apacha made their first contribution in #877
• @KevinNuNu made their first contribution in #887
• @xcnick made their first contribution in #742
• @Eiuyc made their first contribution in #922
• @tonysy made their first contribution in #779
• @luomaoling made their first contribution in #925
• @XWHtorrentx made their first contribution in #948

Full Changelog: v0.5.0...v0.6.0