Awesome

Peng Gao¹, Teli Ma¹, Hongsheng Li², Ziyi Lin², Jifeng Dai³, Yu Qiao¹,

¹ Shanghai AI Laboratory, ² MMLab, CUHK, ³ Sensetime Research.

* We change the project name from ConvMAE to MCMAE.

This repo is the official implementation of MCMAE: Masked Convolution Meets Masked Autoencoders. It currently concludes codes and models for the following tasks:

ImageNet Pretrain: See PRETRAIN.md.
ImageNet Finetune: See FINETUNE.md.
Object Detection: See DETECTION.md.
Semantic Segmentation: See SEGMENTATION.md.
Video Classification: See VideoConvMAE.

Updates

14/Mar/2023

MR-MCMAE (a.k.a. ConvMAE-v2) paper released: Mimic before Reconstruct: Enhancing Masked Autoencoders with Feature Mimicking.

15/Sep/2022

Paper accepted at NeurIPS 2022.

9/Sep/2022

ConvMAE-v2 pretrained checkpoints are released.

21/Aug/2022

Official-ConvMAE-Det which follows official ViTDet codebase is released.

08/Jun/2022

🚀FastConvMAE🚀: significantly accelerates the pretraining hours (4000 single GPU hours => 200 single GPU hours). The code is going to be released at FastConvMAE.

27/May/2022

1. The supported codes for ImageNet-1K pretraining.
2. The supported codes and models for semantic segmentation are provided.

20/May/2022

Update results on video classification.

16/May/2022

The supported codes and models for COCO object detection and instance segmentation are available.

11/May/2022

1. Pretrained models on ImageNet-1K for ConvMAE.
2. The supported codes and models for ImageNet-1K finetuning and linear probing are provided.

08/May/2022

The preprint version is public at arxiv.

Introduction

ConvMAE framework demonstrates that multi-scale hybrid convolution-transformer can learn more discriminative representations via the mask auto-encoding scheme.

• We present the strong and efficient self-supervised framework ConvMAE, which is easy to implement but show outstanding performances on downstream tasks.
• ConvMAE naturally generates hierarchical representations and exhibit promising performances on object detection and segmentation.
• ConvMAE-Base improves the ImageNet finetuning accuracy by 1.4% compared with MAE-Base. On object detection with Mask-RCNN, ConvMAE-Base achieves 53.2 box AP and 47.1 mask AP with a 25-epoch training schedule while MAE-Base attains 50.3 box AP and 44.9 mask AP with 100 training epochs. On ADE20K with UperNet, ConvMAE-Base surpasses MAE-Base by 3.6 mIoU (48.1 vs. 51.7).

Pretrain on ImageNet-1K

The following table provides pretrained checkpoints and logs used in the paper.

The following results are for ConvMAE-v2 (pretrained for 200 epochs on ImageNet-1k).

Main Results on ImageNet-1K

Main Results on COCO

Mask R-CNN

Main Results on ADE20K

UperNet

Main Results on Kinetics-400

Main Results on Something-Something V2

Getting Started

Prerequisites

• Linux
• Python 3.7+
• CUDA 10.2+
• GCC 5+

Training and evaluation

• See PRETRAIN.md for pretraining.
• See FINETUNE.md for pretrained model finetuning and linear probing.
• See DETECTION.md for using pretrained backbone on Mask RCNN.
• See SEGMENTATION.md for using pretrained backbone on UperNet.
• See VideoConvMAE for video classification.

Visualization

Acknowledgement

The pretraining and finetuning of our project are based on DeiT and MAE. The object detection and semantic segmentation parts are based on MIMDet and MMSegmentation respectively. Thanks for their wonderful work.

License

ConvMAE is released under the MIT License.

Citation

@article{gao2022convmae,
  title={ConvMAE: Masked Convolution Meets Masked Autoencoders},
  author={Gao, Peng and Ma, Teli and Li, Hongsheng and Dai, Jifeng and Qiao, Yu},
  journal={arXiv preprint arXiv:2205.03892},
  year={2022}
}