Releases · IDEA-Research/detrex (original) (raw)

detrex v0.5.0 Release

detrex v0.4.0 Release

Main updates

• Support CO-MOT aims for End-to-End Multi-Object Tracking by Feng Yan.
• Release DINO with optimized hyper-parameters which achieves 50.0 AP under 1x settings.
• Release pretrained DINO based on InternImage, ConvNeXt-1K pretrained backbones.
• Release Deformable-DETR-R50 pretrained weights.
• Release DETA and better H-DETR pretrained weights: achieving 50.2 AP and 49.1 AP respectively.

Pretrained Model

• DETA-R50-5scale-12ep bs=8: 50.0AP
• DETA-R50-5scale-12ep aligned hyper-param: 49.9AP
• DETA-R50-5scale-12ep with only freeze the stem of backbone: 50.2AP
• H-Deformable-DETR-two-stage-R50-12ep aligned optimizer hyper-params: 49.1AP
• DINO-R50-4scale-12ep aligned optimizer hyper-params: 49.4AP
• DINO-Focal-3level-4scale-36ep: 58.3AP

Benchmark ConvNeXt on DINO

• convnext-tiny-384: 52.4AP
• convnext-small-384: 54.2AP
• convnext-base-384: 55.1AP
• convnext-large-384: 55.5AP

Benchmark InternImage on DINO

• internimage-tiny: 52.3AP
• internimage-small: 53.5AP
• internimage-base: 54.7AP
• internimage-large: 57.0AP

Benchmark FocalNet on DINO

• focalnet-tiny
• focalnet-small
• focalnet-base

Other pre-trained weights

• Deformable-DETR R50: 44.9 AP (better than 44.5AP from original repo)
• Group-DETR-R50-12ep: 37.8AP

detrex v0.3.0 Release

What's New

New Algorithms

• Support Anchor-DETR
• Support DETA

More training techniques

• Support EMAHook during training by setting train.model_ema.enabled=True, which can further enhance the model performance.
• Fully support mixed precision training by setting train.amp.enabled=True, which can reduce 20% to 30% GPU memory usage.
• Support encoder-decoder checkpoint in DINO which may reduce 30% GPU memory usage. And for more details about the checkpoint usage, please refer to this PR: #200
• Support fast debugging by setting train.fast_dev_run=True.
• Support a great slurm training scripts by @rayleizhu , please check this issue #213 for more details.

Release more than 10+ pretrained checkpoints

Part of the pre-trained weights are converted from their official repo, and all the pre-trained weights can be downloaded in detrex Model Zoo

MaskDINO Release

detrex v0.2.1 Release

Highlights

• DINO has been accepted to ICLR 2023!
• Thanks a lot for @powermano provides us a detailed usage about onnx export in detrex. Please see this issue #192

What's New

New Algorithm

• MaskDINO COCO instance-seg/panoptic-seg pre-release #154

New Features

• New baselines for Res/Swin-DINO-5scale, ViTDet-DINO, FocalNet-DINO, etc. #138, #155
• Support FocalNet backbone #145
• Support Swin-V2 backbone #152

Documentation

• Add ViTDet / FocalNet download links and usage example, please refer to Download Pretrained Weights.
• Add tutorial on how to verify the correct installation of detrex. #194

Bug Fixes

• Fix demo confidence filter not to remove mask predictions #156

Code Refinement

• Make more readable logging info for criterion and matcher #151
• Modified learning rate scheduler config usage, add fundamental scheduler configuration #191

New Pretrained Models

All the pretrained weights can be downloaded in detrex Model Zoo

DINO

detrex v0.2.0 Release

What's New

• Rebuild cleaner config files for projects
• Support H-Deformable-DETR, thanks a lot for @JiaDingCN
• Release H-Deformable-DETR pretrained weights including H-Deformable-DETR-R50, H-Deformable-DETR-Swin-Tiny, H-Deformable-DETR-Swin-Large.
• Add demo for visualizing customized input images or videos using pretrained weights in demo, please check our documentation about the usage.
• Release new baselines for DINO-Swin-Large-36ep, DAB-Deformable-DETR-R50-50ep, DAB-Deformable-DETR-Two-Stage-50ep.

New Pretrained Models

All the pretrained weights can be downloaded in detrex Model Zoo

H-Deformable-DETR

DINO

DAB-Deformable-DETR

detrex v0.1.1 Release

What's New

• Add model analysis tools in tools.
• Support visualization on COCO eval results and annotations in tools.
• Support Group-DETR.
• Release more DINO training results in DINO.
• Release better Deformable-DETR baselines in Deformable-DETR.
• Fix ConvNeXt bugs.
• Add pretrained weights download links and usage in documentation, see Download Pretrained Backbone Weights.
• Add documentation for tools, see Practical Tools and Scripts.

New Pretrained Models

All the pretrained weights can be downloaded in detrex Model Zoo.

DINO

Deformable-DETR