ConvNeXt V2

Co-designing and Scaling ConvNets with Masked Autoencoders

摘要

Driven by improved architectures and better representation learning frameworks, the field of visual recognition has enjoyed rapid modernization and performance boost in the early 2020s. For example, modern ConvNets, represented by ConvNeXt, have demonstrated strong performance in various scenarios. While these models were originally designed for supervised learning with ImageNet labels, they can also potentially benefit from self-supervised learning techniques such as masked autoencoders (MAE). However, we found that simply combining these two approaches leads to subpar performance. In this paper, we propose a fully convolutional masked autoencoder framework and a new Global Response Normalization (GRN) layer that can be added to the ConvNeXt architecture to enhance inter-channel feature competition. This co-design of self-supervised learning techniques and architectural improvement results in a new model family called ConvNeXt V2, which significantly improves the performance of pure ConvNets on various recognition benchmarks, including ImageNet classification, COCO detection, and ADE20K segmentation. We also provide pre-trained ConvNeXt V2 models of various sizes, ranging from an efficient 3.7M-parameter Atto model with 76.7% top-1 accuracy on ImageNet, to a 650M Huge model that achieves a state-of-the-art 88.9% accuracy using only public training data.

使用方式

Predict imageUse the modelTest Command

from mmpretrain import inference_model

predict = inference_model('convnext-v2-atto_fcmae-pre_3rdparty_in1k', 'demo/bird.JPEG')
print(predict['pred_class'])
print(predict['pred_score'])

Models and results

Pretrained models

模型	Params (M)	Flops (G)	配置文件	下载
convnext-v2-atto_3rdparty-fcmae_in1k*	3.71	0.55	config	model
convnext-v2-femto_3rdparty-fcmae_in1k*	5.23	0.78	config	model
convnext-v2-pico_3rdparty-fcmae_in1k*	9.07	1.37	config	model
convnext-v2-nano_3rdparty-fcmae_in1k*	15.62	2.45	config	model
convnext-v2-tiny_3rdparty-fcmae_in1k*	28.64	4.47	config	model
convnext-v2-base_3rdparty-fcmae_in1k*	88.72	15.38	config	model
convnext-v2-large_3rdparty-fcmae_in1k*	197.96	34.40	config	model
convnext-v2-huge_3rdparty-fcmae_in1k*	660.29	115.00	config	model

Models with * are converted from the official repo. The config files of these models are only for inference. We haven’t reproduce the training results.

Image Classification on ImageNet-1k

模型	预训练	Params (M)	Flops (G)	Top-1 (%)	Top-5 (%)	配置文件	下载
convnext-v2-atto_fcmae-pre_3rdparty_in1k*	FCMAE	3.71	0.55	76.64	93.04	config	model
convnext-v2-femto_fcmae-pre_3rdparty_in1k*	FCMAE	5.23	0.78	78.48	93.98	config	model
convnext-v2-pico_fcmae-pre_3rdparty_in1k*	FCMAE	9.07	1.37	80.31	95.08	config	model
convnext-v2-nano_fcmae-pre_3rdparty_in1k*	FCMAE	15.62	2.45	81.86	95.75	config	model
convnext-v2-nano_fcmae-in21k-pre_3rdparty_in1k*	FCMAE ImageNet-21k	15.62	2.45	82.04	96.16	config	model
convnext-v2-tiny_fcmae-pre_3rdparty_in1k*	FCMAE	28.64	4.47	82.94	96.29	config	model
convnext-v2-tiny_fcmae-in21k-pre_3rdparty_in1k*	FCMAE ImageNet-21k	28.64	4.47	83.89	96.96	config	model
convnext-v2-nano_fcmae-in21k-pre_3rdparty_in1k-384px*	FCMAE ImageNet-21k	15.62	7.21	83.36	96.75	config	model
convnext-v2-tiny_fcmae-in21k-pre_3rdparty_in1k-384px*	FCMAE ImageNet-21k	28.64	13.14	85.09	97.63	config	model
convnext-v2-base_fcmae-pre_3rdparty_in1k*	FCMAE	88.72	15.38	84.87	97.08	config	model
convnext-v2-base_fcmae-in21k-pre_3rdparty_in1k*	FCMAE ImageNet-21k	88.72	15.38	86.74	98.02	config	model
convnext-v2-large_fcmae-pre_3rdparty_in1k*	FCMAE	197.96	34.40	85.76	97.59	config	model
convnext-v2-large_fcmae-in21k-pre_3rdparty_in1k*	FCMAE ImageNet-21k	197.96	34.40	87.26	98.24	config	model
convnext-v2-base_fcmae-in21k-pre_3rdparty_in1k-384px*	FCMAE ImageNet-21k	88.72	45.21	87.63	98.42	config	model
convnext-v2-large_fcmae-in21k-pre_3rdparty_in1k-384px*	FCMAE ImageNet-21k	197.96	101.10	88.18	98.52	config	model
convnext-v2-huge_fcmae-pre_3rdparty_in1k*	FCMAE	660.29	115.00	86.25	97.75	config	model
convnext-v2-huge_fcmae-in21k-pre_3rdparty_in1k-384px*	FCMAE ImageNet-21k	660.29	337.96	88.68	98.73	config	model
convnext-v2-huge_fcmae-in21k-pre_3rdparty_in1k-512px*	FCMAE ImageNet-21k	660.29	600.81	88.86	98.74	config	model

Models with * are converted from the official repo. The config files of these models are only for inference. We haven’t reproduce the training results.

引用

@article{Woo2023ConvNeXtV2,
  title={ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders},
  author={Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon and Saining Xie},
  year={2023},
  journal={arXiv preprint arXiv:2301.00808},
}