SparK

SparK: the first successful BERT/MAE-style pretraining on any convolutional networks  

This is the official implementation of ICLR paper Designing BERT for Convolutional Networks: Sparse and Hierarchical Masked Modeling, which can pretrain any CNN (e.g., ResNet) in a BERT-style self-supervised manner. We've tried our best to make the codebase clean, short, easy to read, state-of-the-art, and only rely on minimal dependencies.

https://user-images.githubusercontent.com/39692511/226858919-dd4ccf7e-a5ba-4a33-ab21-4785b8a7833c.mp4

News

• A brief introduction (in English) is available on our ICLR poster page! [Recorded Video, Poster, and Slides].
• On May. 11th another livestream on OpenMMLab & ReadPaper (bilibili)! [Recorded Video]
• On Apr. 27th (UTC+8 8pm) another livestream would be held at OpenMMLab (bilibili)!
• On Mar. 22nd (UTC+8 8pm) another livestream would be held at 极市平台 (bilibili)! [Recorded Video]
• The share on TechBeat (将门创投) is scheduled on Mar. 16th (UTC+8 8pm) too! [Recorded Video]
• We are honored to be invited by Synced ("机器之心机动组 视频号" on WeChat) to give a talk about SparK on Feb. 27th (UTC+0 11am, UTC+8 7pm), welcome! [Recorded Video]
• This work got accepted to ICLR 2023 as a Spotlight (notable-top-25%).
• Other articles: [Synced] [DeepAI] [TheGradient] [Bytedance] [CVers [QbitAI(量子位)] [BAAI(智源)] [机器之心机动组] [极市平台] [ReadPaper笔记]

️ Colab Visualization Demo

Check pretrain/viz_reconstruction.ipynb for visualizing the reconstruction of SparK pretrained models, like:

We also provide pretrain/viz_spconv.ipynb that shows the "mask pattern vanishing" issue of dense conv layers.

What's new here?

Pretrained CNN beats pretrained Swin-Transformer:

After SparK pretraining, smaller models can beat un-pretrained larger models:

All models can benefit, showing a scaling behavior:

Generative self-supervised pretraining surpasses contrastive learning:

See our paper for more analysis, discussions, and evaluations.

Todo list

• Pretraining code
• Pretraining toturial for customized CNN model (Tutorial for pretraining your own CNN model)
• Pretraining toturial for customized dataset (Tutorial for pretraining your own dataset)
• Pretraining Colab visualization playground (reconstruction, sparse conv)
• Finetuning code
• Weights & visualization playground in huggingface
• Weights in timm

Pretrained weights (self-supervised; w/o decoder; can be directly finetuned)

Note: for network definitions, we directly use timm.models.ResNet and official ConvNeXt.

reso.: the image resolution; acc@1: ImageNet-1K finetuned acc (top-1)

Installation & Running

We highly recommended you to use torch==1.10.0, torchvision==0.11.1, and timm==0.5.4 for reproduction. Check INSTALL.md to install all pip dependencies.

• Loading pretrained model weights in 3 lines

# download our weights `resnet50_1kpretrained_timm_style.pth` first
import torch, timm
res50, state = timm.create_model('resnet50'), torch.load('resnet50_1kpretrained_timm_style.pth', 'cpu')
res50.load_state_dict(state.get('module', state), strict=False)     # just in case the model weights are actually saved in state['module']

• Pretraining

  • any ResNet or ConvNeXt on ImageNet-1k:  see pretrain/
  • your own CNN model:  see pretrain/, especially pretrain/models/custom.py

• Finetuning

  • any ResNet or ConvNeXt on ImageNet-1k:  check downstream_imagenet/ for subsequent instructions.
  • ResNets on COCO:  see downstream_d2/
  • ConvNeXts on COCO:  see downstream_mmdet/

Acknowledgement

We referred to these useful codebases:

• BEiT, MAE, ConvNeXt
• timm, MoCoV2, Detectron2, MMDetection

License

This project is under the MIT license. See LICENSE for more details.

Citation

If you found this project useful, you can kindly give us a star ⭐, or cite us in your work :

@Article{tian2023designing,
  author  = {Keyu Tian and Yi Jiang and Qishuai Diao and Chen Lin and Liwei Wang and Zehuan Yuan},
  title   = {Designing BERT for Convolutional Networks: Sparse and Hierarchical Masked Modeling},
  journal = {arXiv:2301.03580},
  year    = {2023},
}