MMAction2 V0.12.0 Release

Highlights

• Support TSM-MobileNetV2
• Support TANet
• Support GPU Normalize

New Features

• Support TSM-MobileNetV2 (#415)
• Support flip with label mapping (#591)
• Add seed option for sampler (#642)
• Support GPU Normalize (#586)
• Support TANet (#595)

Improvements

• Training custom classes of ava dataset (#555)
• Add CN README in homepage (#592, #594)
• Support soft label for CrossEntropyLoss (#625)
• Refactor config: Specify train_cfg and test_cfg in model (#629)
• Provide an alternative way to download older kinetics annotations (#597)
• Update FAQ for
  • 1). data pipeline about video and frames (#598)
  • 2). how to show results (#598)
  • 3). batch size setting for batchnorm (#657)
  • 4). how to fix stages of backbone when finetuning models (#658)
• Modify default value of save_best (#600)
• Use BibTex rather than latex in markdown (#607)
• Add warnings of uninstalling mmdet and supplementary documents (#624)
• Support soft label for CrossEntropyLoss (#625)

Bug and Typo Fixes

• Fix value of pem_low_temporal_iou_threshold in BSN (#556)
• Fix ActivityNet download script (#601)

ModelZoo

• Add TSM-MobileNetV2 for Kinetics400 (#415)
• Add deeper SlowFast models (#605)

MMAction2 V0.11.0 Release

Highlights

• Support imgaug
• Support spatial temporal demo
• Refactor EvalHook, config structure, unittest structure

New Features

• Support imgaug for augmentations in the data pipeline (#492)
• Support setting max_testing_views for extremely large models to save GPU memory used (#511)
• Add spatial temporal demo (#547, #566)

Improvements

• Refactor EvalHook (#395)
• Refactor AVA hook (#567)
• Add repo citation (#545)
• Add dataset size of Kinetics400 (#503)
• Add lazy operation docs (#504)
• Add class_weight for CrossEntropyLoss and BCELossWithLogits (#509)
• add some explanation about the resampling in slowfast (#502)
• Modify paper title in README.md (#512)
• Add alternative ways to download Kinetics (#521)
• Add OpenMMLab projects link in README (#530)
• Change default preprocessing to shortedge to 256 (#538)
• Add config tag in dataset README (#540)
• Add solution for markdownlint installation issue (#497)
• Add dataset overview in readthedocs (#548)
• Modify the trigger mode of the warnings of missing mmdet (583)
• Refactor config structure (#488, #572)
• Refactor unittest structure (#433)

Bug and Typo Fixes

• Fix a bug about ava dataset validation (#527)
• Fix a bug about ResNet pretrain weight initialization (#582)
• Fix a bug in CI due to MMCV index (#495)
• Remove invalid links of MiT and MMiT (#516)
• Fix frame rate bug for AVA preparation (#576)

MMAction2 V0.10.0 Release

Highlights

• Support Spatio-Temporal Action Detection (AVA)
• Support precise BN

New Features

• Support precise BN (#501)
• Support Spatio-Temporal Action Detection (AVA) (#351)
• Support to return feature maps in inference_recognizer (#458)

Improvements

• Add arg stride to long_video_demo.py, to make inference faster (#468)
• Support training and testing for Spatio-Temporal Action Detection (#351)
• Fix CI due to pip upgrade (#454)
• Add markdown lint in pre-commit hook (#255)
• Speed up confusion matrix calculation (#465)
• Use title case in modelzoo statistics (#456)
• Add FAQ documents for easy troubleshooting. (#413, #420, #439)
• Support Spatio-Temporal Action Detection with context (#471)
• Add class weight for CrossEntropyLoss and BCELossWithLogits (#509)
• Add Lazy OPs docs (#504)

Bug and Typo Fixes

• Fix typo in default argument of BaseHead (#446)
• Fix potential bug about output_config overwrite (#463)

ModelZoo

• Add SlowOnly, SlowFast for AVA2.1 (#351)

MMAction2 V0.9.0 Release

Highlights

• Support GradCAM utils for recognizers
• Support ResNet Audio model

New Features

• Automatically add modelzoo statistics to readthedocs (#327)
• Support GYM99 data preparation (#331)
• Add AudioOnly Pathway from AVSlowFast. (#355)
• Add GradCAM utils for recognizer (#324)
• Add print config script (#345)
• Add online motion vector decoder (#291)

Improvements

• Support PyTorch 1.7 in CI (#312)
• Support to predict different labels in a long video (#274)
• Update docs bout test crops (#359)
• Polish code format using pylint manually (#338)
• Update unittest coverage (#358, #322, #325)
• Add random seed for building filelists (#323)
• Update colab tutorial (#367)
• set default batch_size of evaluation and testing to 1 (#250)
• Rename the preparation docs to README.md (#388)
• Move docs about demo to demo/README.md (#329)
• Remove redundant code in tools/test.py (#310)
• Automatically calculate number of test clips for Recognizer2D (#359)

Bug and Typo Fixes

• Fix rename Kinetics classnames bug (#384)
• Fix a bug in BaseDataset when data_prefix is None (#314)
• Fix a bug about tmp_folder in OpenCVInit (#357)
• Fix get_thread_id when not using disk as backend (#354, #357)
• Fix the bug of HVU object num_classes from 1679 to 1678 (#307)
• Fix typo in export_model.md (#399)
• Fix OmniSource training configs (#321)
• Fix Issue #306: Bug of SampleAVAFrames (#317)

ModelZoo

• Add SlowOnly model for GYM99, both RGB and Flow (#336)
• Add auto modelzoo statistics in readthedocs (#327)
• Add TSN for HMDB51 pretrained on Kinetics400, Moments in Time and ImageNet. (#372)

MMAction2 V0.8.0 Release

v0.8.0 (31/10/2020)

Highlights

• Support OmniSource
• Support C3D
• Support video recognition with audio modality
• Support HVU
• Support X3D

New Features

• Support AVA dataset preparation (#266)
• Support the training of video recognition dataset with multiple tag categories (#235)
• Support joint training with multiple training datasets of multiple formats, including images, untrimmed videos, etc. (#242)
• Support to specify a start epoch to conduct evaluation (#216)
• Implement X3D models, support testing with model weights converted from SlowFast (#288)

Improvements

• Set default values of 'average_clips' in each config file so that there is no need to set it explicitly during testing in most cases (#232)
• Extend HVU datatools to generate individual file list for each tag category (#258)
• Support data preparation for Kinetics-600 and Kinetics-700 (#254)
• Add cfg-options in arguments to override some settings in the used config for convenience (#212)
• Rename the old evaluating protocol mean_average_precision as mmit_mean_average_precision since it is only used on MMIT and is not the mAP we usually talk about. Add mean_average_precision, which is the real mAP (#235)
• Add accurate setting (Three crop * 2 clip) and report corresponding performance for TSM model (#241)
• Add citations in each preparing_dataset.md in tools/data/dataset (#289)
• Update the performance of audio-visual fusion on Kinetics-400 (#281)
• Support data preparation of OmniSource web datasets, including GoogleImage, InsImage, InsVideo and KineticsRawVideo (#294)
• Use metric_options dict to provide metric args in evaluate (#286)

Bug Fixes

• Register FrameSelector in PIPELINES (#268)
• Fix the potential bug for default value in dataset_setting (#245)
• Fix the data preparation bug for something-something dataset (#278)
• Fix the invalid config url in slowonly README data benchmark (#249)
• Validate that the performance of models trained with videos have no significant difference comparing to the performance of models trained with rawframes (#256)
• Correct the img_norm_cfg used by TSN-3seg-R50 UCF-101 model, improve the Top-1 accuracy by 3% (#273)

ModelZoo

• Add Baselines for Kinetics-600 and Kinetics-700, including TSN-R50-8seg and SlowOnly-R50-8x8 (#259)
• Add OmniSource benchmark on MiniKineitcs (#296)
• Add Baselines for HVU, including TSN-R18-8seg on 6 tag categories of HVU (#287)
• Add X3D models ported from SlowFast (#288)

MMAction2 V0.7.0 Release

Highlights

• Support TPN
• Support JHMDB, UCF101-24, HVU dataset preparation
• support onnx model conversion

New Features

• Support the data pre-processing pipeline for the HVU Dataset (#277)
• Support real-time action recognition from web camera (#171)
• Support onnx (#160)
• Support UCF101-24 preparation (#219)
• Support evaluating mAP for ActivityNet with CUHK17_activitynet_pred (#176)
• Add the data pipeline for ActivityNet, including downloading videos, extracting RGB and Flow frames, finetuning TSN and extracting feature (#190)
• Support JHMDB preparation (#220)

ModelZoo

• Add finetuning setting for SlowOnly (#173)
• Add TSN and SlowOnly models trained with OmniSource, which achieve 75.7% Top-1 with TSN-R50-3seg and 80.4% Top-1 with SlowOnly-R101-8x8 (#215)

Improvements

• Support demo with video url (#165)
• Support multi-batch when testing (#184)
• Add tutorial for adding a new learning rate updater (#181)
• Add config name in meta info (#183)
• Remove git hash in __version__ (#189)
• Check mmcv version (#189)
• Update url with 'https://download.openmmlab.com' (#208)
• Update Docker file to support PyTorch 1.6 and update install.md (#209)
• Polish readsthedocs display (#217, #229)

Bug Fixes

• Fix the bug when using OpenCV to extract only RGB frames with original shape (#184)
• Fix the bug of sthv2 num_classes from 339 to 174 (#174, #207)

MMAction2 V0.6.0 Release

Highlights

• Support TIN, CSN, SSN, NonLocal
• Support FP16 training

New Features

• Support NonLocal module and provide ckpt in TSM and I3D (#41)
• Support SSN (#33, #37, #52, #55)
• Support CSN (#87)
• Support TIN (#53)
• Support HMDB51 dataset preparation (#60)
• Support encoding videos from frames (#84)
• Support FP16 training (#25)
• Enhance demo by supporting rawframe inference (#59), output video/gif (#72)

ModelZoo

• Update Slowfast modelzoo (#51)
• Update TSN, TSM video checkpoints (#50)
• Add data benchmark for TSN (#57)
• Add data benchmark for SlowOnly (#77)
• Add BSN/BMN performance results with feature extracted by our codebase (#99)

Improvements

• Polish data preparation codes (#70)
• Improve data preparation scripts (#58)
• Improve unittest coverage and minor fix (#62)
• Support PyTorch 1.6 in CI (#117)
• Support with_offset for rawframe dataset (#48)
• Support json annotation files (#119)
• Support multi-class in TSMHead (#104)
• Support using val_step() to validate data for each val workflow (#123)
• Use xxInit() method to get total_frames and make total_frames a required key (#90)
• Add paper introduction in model readme (#140)
• Adjust the directory structure of tools/ and rename some scripts files (#142)

Bug Fixes

• Fix configs for localization test (#67)
• Fix configs of SlowOnly by fixing lr to 8 gpus (#136)
• Fix the bug in analyze_log (#54)
• Fix the bug of generating HMDB51 class index file (#69)
• Fix the bug of using load_checkpoint() in ResNet (#93)
• Fix the bug of --work-dir when using slurm training script (#110)
• Correct the sthv1/sthv2 rawframes filelist generate command (#71)
• CosineAnnealing typo (#47)

MMAction2 V0.5.0 Release

The first release of MMAction2.