Mmediting Versions Save

Highlights

1. Support RealBasicVSR
2. Support Real-ESRGAN checkpoint

New Features

• Support video input and output in restoration demo (#622)
• Support RealBasicVSR (#632, #633, #647, #680)
• Support Real-ESRGAN checkpoint (#635)
• Support conversion to y-channel when loading images (643)
• Support random video compression during training (#646)
• Support crop sequence (#648)
• Support pixel_unshuffle (#684)

Bug Fixes

• Change 'target_size' for RandomResize from list to tuple (#617)
• Fix folder creation in preprocess_df2k_ost_dataset.py (#623)
• Change TDAN config path in README (#625)
• Change 'radius' to 'kernel_size' for UnsharpMasking in Real-ESRNet config (#626)
• Fix bug in MATLABLikeResize (#630)
• Fix 'flow_warp' comment (#655)
• Fix the error of Model Zoo and Datasets in docs (#664)
• Fix bug in 'random_degradations' (#673)
• Limit opencv-python version (#689)

Improvements

• Translate docs to Chinese (#576, #577, #578, #579, #581, #582, #584, #587, #588, #589, #590, #591, #592, #593, #594, #595, #596, #641, #647, #656, #665, #666)
• Add UNetDiscriminatorWithSpectralNorm (#605)
• Use PyTorch sphinx theme (#607, #608)
• Update mmcv (#609), mmflow (#621), mmfewshot (#634) and mmhuman3d (#649) in docs
• Convert minimum GCC version to 5.4 (#612)
• Add tiff in SRDataset IMG_EXTENSIONS (#614)
• Update metafile and update_model_index.py (#615)
• Update preprocess_df2k_ost_dataset.py (#624)
• Add Abstract to README (#628, #636)
• Align NIQE to MATLAB results (#631)
• Add official markdown lint hook (#639)
• Skip CI when some specific files were changed (#640)
• Update docs/conf.py (#644, #651)
• Try to create a symbolic link on windows (#645)
• Cancel previous runs that are not completed (#650)
• Update path of configs in demo.md and getting_started.md (#658, #659)
• Use mmcv root model registry (#660)
• Update README.md (#654, #663)
• Refactor the structure of documentation (#668)
• Add script to crop REDS images into sub-images for faster IO (#669)
• Capitalize the first letter of the task name in the metafile (#678)
• Update FixedCrop for cropping image sequence (#682)

Contributors @wangruohui @nbei @ckkelvinchan @Yshuo-Li @LeoXing1996 @RangiLyu @matrixgame2018 @huoshuai-dot @innerlee @okotaku @Adenialzz @kai422

Highlights

• GLEAN for blind face image restoration #530
• Real-ESRGAN model #546

New Features

• Exponential Moving Average Hook #542
• Support DF2K_OST dataset #566

Improvements

• Add MATLAB-like bicubic interpolation #507
• Support random degradations during training #504
• Support torchserve #568

Contributors @ckkelvinchan @Yshuo-Li @Adenialzz @kai422 @jiaqixuac @plygager @Ha0Tang @innerlee

Highlights

1. Support LIIF-RDN (CVPR'2021)
2. Support BasicVSR++ (NTIRE'2021)

New Features

• Support loading annotation from file for video SR datasets (#423)
• Support persistent worker (#426)
• Support LIIF-RDN (#428, #440)
• Support BasicVSR++ (#451, #467)
• Support mim (#455)

Bug Fixes

• Fix bug in stat.py (#420)
• Fix astype error in function tensor2img (#429)
• Fix device error caused by torch.new_tensor when pytorch >= 1.7 (#465)
• Fix _non_dist_train in .mmedit/apis/train.py (#473)
• Fix multi-node distributed test (#478)

Breaking Changes

• Refactor LIIF for pytorch2onnx (#425)

Improvements

• Update Chinese docs (#415, #416, #418, #421, #424, #431, #442)
• Add CI of pytorch 1.9.0 (#444)
• Refactor README.md of configs (#452)
• Avoid loading pretrained VGG in unittest (#466)
• Support specifying scales in preprocessing div2k dataset (#472)
• Support all formats in readthedocs (#479)
• Use version_info of mmcv (#480)
• Remove unnecessary codes in restoration_video_demo.py (#484)
• Change priority of DistEvalIterHook to 'LOW' (#489)
• Reset resource limit (#491)
• Update QQ QR code in README_CN.md (#494)
• Add myst_parser (#495)
• Add license header (#496)
• Fix typo of StyleGAN modules (#427)
• Fix typo in docs/demo.md (#453, #454)
• Fix typo in tools/data/super-resolution/reds/README.md (#469)

We thank all the contributors of this release:

@610265158, @AlexZou14, @Ha0Tang, @LiUzHiAn, @Yshuo-Li, @ckkelvinchan, @innerlee, @nbei, @orangeccc, @wileechou, @yivan-WYYGDSG

Thank you! ❤️

Highlights

1. Support DIC and DIC-GAN (CVPR'2020)
2. Support GLEAN Cat 8x (CVPR'2021)
3. Support TTSR-GAN (CVPR'2020)
4. Add colab tutorial for super-resolution

New Features

• Add DIC (#342, #345, #348, #350, #351, #357, #363, #365, #366)
• Add SRFolderMultipleGTDataset (#355)
• Add GLEAN Cat 8x (#367)
• Add SRFolderVideoDataset (#370)
• Add colab tutorial for super-resolution (#380)
• Add TTSR-GAN (#372, #381, #383, #398)
• Add DIC-GAN (#392, #393, #394)

Bug Fixes

• Fix bug in restoration_video_inference.py (#379)
• Fix Config of LIIF (#368)
• Change the path to pre-trained EDVR-M (#396)
• Fix normalization in restoration_video_inference (#406)
• Fix [brush_stroke_mask] error in unittest (#409)

Breaking Changes

• Change mmcv minimum version to v1.3 (#378)

Improvements

• Correct Typos in code (#371)
• Add Custom_hooks (#362)
• Refactor unittest folder structure (#386)
• Add documents and download link for Vid4 (#399)
• Update model zoo for documents (#400)
• Update metafile (407)

Highlights

1. Support GLEAN (CVPR'2021)
2. Support TTSR (CVPR'2020)
3. Support TDAN (CVPR'2020)

New Features

• Add GLEAN (#296, #332)
• Support PWD metafile (#298)
• Support CropLike in pipeline (#299)
• Add TTSR (#301, #304, #307, #311, #311, #312, #313, #314, #321, #326, #327)
• Add TDAN (#316, #334, #347)
• Add onnx2tensorrt (#317)
• Add tensorrt evaluation (#328)
• Add SRFacicalLandmarkDataset (#329)
• Add key point auxiliary model for DIC (#336, #341)
• Add demo for video super-resolution methods (#275)
• Add SR Folder Ref Dataset (#292)
• Support FLOPs calculation of video SR models (#309)

Bug Fixes

• Fix find_unused_parameters in PyTorch 1.8 for BasicVSR (#290)
• Fix error in publish_model.py for pt>=1.6 (#291)
• Fix PSNR when input is uint8 (#294)

Improvements

• Support backend in LoadImageFromFile (#293, #303)
• Update metric_average_mode of video SR dataset (#319)
• Add error message in restoration_demo.py (324)
• Minor correction in getting_started.md (#339)
• Update description for Vimeo90K (#349)
• Support start_index in GenerateSegmentIndices (#338)
• Support different filename templates in GenerateSegmentIndices (#325)
• Support resize by scale-factor (#295, #310)

TL;DR

1. Support BasicVSR (CVPR'2021)
2. Support IconVSR (CVPR'2021)
3. Support RDN (CVPR'2018)
4. Add onnx evaluation tool

New Features

• Add RDN (#233, #260, #280)
• Add MultipleGT datasets (#238)
• Add BasicVSR and IconVSR (#245, #252, #253, #254, #264, #274, #258, #252, #264)
• Add onnx evaluation tool (#279)

Bug Fixes

• Fix onnx conversion of maxunpool2d (#243)
• Fix inpainting in demo.md (#248)
• Tiny fix of config file of EDSR (#251)
• Fix link in README (#256)
• Fix restoration_inference key missing bug (#270)
• Fix the usage of channel_order in loading.py (#271)
• Fix the command of inpainting (#278)
• Fix preprocess_vimeo90k_dataset.py args name (#281)

Improvements

• Support empty_cache option in test.py (#261)
• Update projects in README (#249, #276)
• Support Y-channel PSNR and SSIM (#250)
• Add zh-CN README (#262)
• Update pytorch2onnx doc (#265)
• Remove extra quotation in English readme (#268)
• Change tags to comment (#269)
• List model zoo in README (#284, #285, #286)

Highlights

1. Support Local Implicit Image Function (LIIF)
2. Support exporting DIM and GCA from Pytorch to ONNX

New Features

• Add readthedocs config files and fix docstring (#92)
• Add github action file (#94)
• Support exporting DIM and GCA from Pytorch to ONNX (#105)
• Support concatenating datasets (#106)
• Support non_dist_train validation (#110)
• Add matting colab tutorial (#111)
• Support niqe metric (#114)
• Support PoolDataLoader for parrots (#134)
• Support collect-env (#137, #143)
• Support pt1.6 cpu/gpu in CI (#138)
• Support fp16 (139, #144)
• Support publishing to pypi (#149)
• Add modelzoo statistics (#171, #182, #186)
• Add doc of datasets (194)
• Support extended foreground option. (#195, #199, #200, #210)
• Support nn.MaxUnpool2d (#196)
• Add some FBA components (#203, #209, #215, #220)
• Support random down sampling in pipeline (#222)
• Support SR folder GT Dataset (#223)
• Support Local Implicit Image Function (LIIF) (#224, #226, #227, #234, #239)

Bug Fixes

• Fix _non_dist_train in train api (#104)
• Fix setup and CI (#109)
• Fix redundant loop bug in Normalize (#121)
• Fix get_hash in setup.py (#124)
• Fix tool/preprocess_reds_dataset.py (#148)
• Fix slurm train tutorial in getting_started.md (#162)
• Fix pip install bug (#173)
• Fix bug in config file (#185)
• Fix broken links of datasets (#236)
• Fix broken links of model zoo (#242)

Breaking Changes

• Refactor data loader configs (#201)

Improvements

• Updata requirements.txt (#95, #100)
• Update teaser (#96)
• Updata README (#93, #97, #98, #152)
• Updata model_zoo (#101)
• Fix typos (#102, #188, #191, #197, #208)
• Adopt adjust_gamma from skimage and reduce dependencies (#112)
• remove .gitlab-ci.yml (#113)
• Update import of first party (#115)
• Remove citation and contact (#122)
• Update version file (#136)
• Update download url (#141)
• Update setup.py (#150)
• Update the highest version of supported mmcv (#153, #154)
• modify Crop to handle a sequence of video frames (#164)
• Add links to other mm projects (#179, #180)
• Add config type (#181)
• Refector docs (#184)
• Add config link (#187)
• Update file structure (#192)
• Update config doc (#202)
• Update slurm_train.md script (#204)
• Improve code style (#206, #207)
• Use file_client in CompositeFg (#212)
• Replace random with numpy.random (#213)
• Refactor loader_cfg (#214)

New Features

• NIQE metric (#114)
• Support FP16 training (#139)

Improvements

• Update version file (#136)
• Update collect env function (#137)
• Update download urls (#141)
• Update docker file with pt1.6 (#144)