SparK: BERT/MAE-style Pretraining on Any Convolutional Networks

Implementation of the paper Designing BERT for Convolutional Networks: Sparse and Hierarchical Masked Modeling.

   

🔥 News

• The share on TechBeat (将门创投) is scheduled on Mar. 16th (UTC+0 12am) 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]

📺 Video demo

https://user-images.githubusercontent.com/6366788/213662770-5f814de0-cbe8-48d9-8235-e8907fd81e0e.mp4

What's new here?

🔥 On ResNets, generative pre-training surpasses contrastive learning for the first time:

🔥 ConvNeXt gains more from pre-training than Swin-Transformer, up to +3.5 points:

🔥 Larger models benefit more from SparK pre-training, showing a scaling behavior:

🔥 Pre-trained model can make reasonable predictions:

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

Catalog

• Pre-training code
• Fine-tuning code
• Colab visualization playground
• Weights & visualization playground on Huggingface
• Weights in timm

ImageNet-1k results and pre-trained networks weights

Note: for network definitions, we directly use timm.models.ResNet and official ConvNeXt. reso.: the image resolution; acc@1: IN1k fine-tuned acc (top-1)

arch.	reso.	acc@1	#params	flops	weights on google drive
ResNet50	224	80.6	26M	4.1G	resnet50_1kpretrained_timm_style.pth
ResNet101	224	82.2	45M	7.9G	resnet101_1kpretrained_timm_style.pth
ResNet152	224	82.7	60M	11.6G	resnet152_1kpretrained_timm_style.pth
ResNet200	224	83.1	65M	15.1G	resnet200_1kpretrained_timm_style.pth
ConvNeXt-S	224	84.1	50M	8.7G	convnextS_1kpretrained_official_style.pth
ConvNeXt-B	224	84.8	89M	15.4G	convnextB_1kpretrained_official_style.pth
ConvNeXt-L	224	85.4	198M	34.4G	convnextL_1kpretrained_official_style.pth
ConvNeXt-L	384	86.0	198M	101.0G	convnextL_384_1kpretrained_official_style.pth
L-with-decoder	384	86.0	198M	101.0G	cnxL384_withdecoder_1kpretrained_spark_style.pth

Installation

For pre-training and fine-tuning on ImageNet-1k, we highly recommended you to use torch==1.10.0, torchvision==0.11.1, and timm==0.5.4.

Check INSTALL.md to install all dependencies for pre-training and ImageNet fine-tuning.

Pre-training

See pretrain/ to pre-train models on ImageNet-1k.

Fine-tuning

• All models on ImageNet: check downstream_imagenet/ for subsequent instructions.
• ResNets on COCO: see downstream_d2/ for details.
• ConvNeXts on COCO: see downstream_mmdet/ for details.

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 may consider staring ⭐, or citing us 📖:

@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},
}