Alibaba EasyCV Versions Save

Highlights

• Support auto hyperparameter optimization of NNI (#211)
• Add DeiT III (#171)
• Add semantic segmentation model SegFormer (#191)
• Add 3d detection model BEVFormer (#203)

New Features

• Support semantic mask2former (#199)
• Support face 2d keypoint detection (#191)
• Support hand keypoints detection (#191)
• Support wholebody keypoint detection (#207)
• Support auto hyperparameter optimization of NNI (#211)
• Add DeiT III (#171)
• Add semantic segmentation model SegFormer (#191)
• Add 3d detection model BEVFormer (#203)

Improvements

• Optimze predcitor apis, support cpu and batch inference (#195)
• Speed up ViTDet model (#177)
• Support export jit model end2end for yolox (#215)

Bug Fixes

• Fix the bug of io.copytree copying multiple directories (#193)
• Move thirdparty into easycv (#216)

Highlights

• Release YOLOX-PAI which achieves SOTA results within 40~50 mAP (less than 1ms) (#154 #172 #174 )
• Add detection algo DINO (#144)
• Add mask2former algo (#115)
• Releases imagenet1k, imagenet22k, coco, lvis, voc2012 data with BaiduDisk to accelerate downloading (#145 )

New Features

• Add detection predictor which support model inference without exporting models(#158 )
• Add VitDet support for faster-rcnn (#155 )
• Release YOLOX-PAI which achieves SOTA results within 40~50 mAP (less than 1ms) (#154 #172 #174 )
• Support DINO algo (#144)
• Add mask2former algo (#115)

Improvements

• FCOS update torch_style (#170)
• Add algo tables to describe which algo EasyCV support (#157 )
• Refactor datasources api (#156 #140 )
• Add PR and Issule template (#150)
• Update Fast ConvMAE doc (#151)

Bug Fixes

• Fix YOLOXLrUpdaterHook conflict with mmdet ( #169 )
• Fix datasource cache problem(#153)
• Fix missing utils(#183 )

Highlights

• Release YOLOX-PAI which achieves SOTA results within 40~50 mAP (less than 1ms) (#154 #172 #174 )
• Add detection algo DINO (#144)
• Add mask2former algo (#115)
• Releases imagenet1k, imagenet22k, coco, lvis, voc2012 data with BaiduDisk to accelerate downloading (#145 )

New Features

• Add detection predictor which support model inference without exporting models(#158 )
• Add VitDet support for faster-rcnn (#155 )
• Release YOLOX-PAI which achieves SOTA results within 40~50 mAP (less than 1ms) (#154 #172 #174 )
• Support DINO algo (#144)
• Add mask2former algo (#115)

Improvements

• FCOS update torch_style (#170)
• Add algo tables to describe which algo EasyCV support (#157 )
• Refactor datasources api (#156 #140 )
• Add PR and Issule template (#150)
• Update Fast ConvMAE doc (#151)

Bug Fixes

• Fix YOLOXLrUpdaterHook conflict with mmdet ( #169 )
• Fix datasource cache problem(#153)

Highlights

• Self-Supervised support ConvMAE algorithm ((#101) (#121))
• Classification support EfficientFormer algorithm (#128)
• Detection support FCOS、DETR、DAB-DETR and DN-DETR algorithm ((#100) (#104) (#119))
• Segmentation support UperNet algorithm (#118)
• Support use torchacc to speed up training (#105)

New Features

• Support use analyze tools (#133)

Bug Fixes

• Update yolox config template and fix bugs (#134)
• Fix yolox detector prediction export error (#125)
• Fix common_io url error (#126)

Improvements

• Add ViTDet visualization (#102)
• Refactor detection pipline (#104)

Bug Fixes

• Fix yolox detector reading config file error (#125)

Highlights

• Add semantic segmentation modules, support FCN algorithm (#71)
• Expand classification model zoo (#55)
• Support export model with blade for yolox (#66)
• Support ViTDet algorithm (#35)
• Add sailfish for extensible fully sharded data parallel training (#97)
• Support run with mmdet models (#25)

New Features

• Set multiprocess env for speedup (#77)
• Add data hub, summarized various datasets in different fields (#70)

Bug Fixes

• Fix the inaccurate accuracy caused by missing the groundtruth_is_crowd field in CocoMaskEvaluator (#61)
• Unified the usage of pretrained parameter and fix load bugs（#79) (#85) (#95)

Improvements

• Update MAE pretrained models and benchmark (#50)
• Add detection benchmark for SwAV and MoCo-v2 (#58)
• Add moby swin-tiny pretrained model and benchmark (#72)
• Update prepare_data.md, add more details (#69)
• Optimize quantize code and support to export MNN model (#44)

Bug Fixes

• Fix the eval metric bug during the training caused by dataset.visualize (#67)

Highlights

• Support image visualization for tensorboard and wandb (#15)

New Features

• Update moby pretrained model to deit small (#10)
• Support image visualization for tensorboard and wandb (#15)
• Add mae vit-large benchmark and pretrained models (#24)

Bug Fixes

• Fix extract.py for benchmarks (#7)
• Fix inference error of classifier (#19)
• Fix multi-process reading of detection datasource and accelerate data preprocessing (#23)
• Fix torchvision transforms wrapper (#31)

Improvements

• Add chinese readme (#39)
• Add model compression tutorial (#20)
• Add notebook tutorials (#22)
• Uniform input and output format for transforms (#6)
• Update model zoo link (#8)
• Support readthedocs (#29)
• refine autorelease gitworkflow (#13)

EasyCV init release.

1. SOTA SSL Algorithms EasyCV provides state-of-the-art algorithms in self-supervised learning based on contrastive learning such as SimCLR, MoCO V2, Swav, DINO and also MAE based on masked image modeling. We also provides standard benchmark tools for ssl model evaluation.
2. Vision Transformers EasyCV aims to provide plenty vision transformer models trained either using supervised learning or self-supervised learning, such as ViT, Swin-Transformer and XCit. More models will be added in the future.
3. Functionality & Extensibility In addition to SSL, EasyCV also support image classification, object detection, metric learning, and more area will be supported in the future. Although convering different area, EasyCV decompose the framework into different componets such as dataset, model, running hook, making it easy to add new compoenets and combining it with existing modules. EasyCV provide simple and comprehensive interface for inference. Additionaly, all models are supported on PAI-EAS, which can be easily deployed as online service and support automatic scaling and service moniting.
4. Efficiency EasyCV support multi-gpu and multi worker training. EasyCV use DALI to accelerate data io and preprocessing process, and use fp16 to accelerate training process. For inference optimization, EasyCV export model using jit script, which can be optimized by PAI-Blade.