Exploring Long-Sequence Masked Autoencoders

This is the code release of the paper Exploring Long-Sequence Masked Autoencoders:

@Article{hu2022exploring,
  author  = {Ronghang Hu and Shoubhik Debnath and Saining Xie and Xinlei Chen},
  journal = {arXiv:2210.07224},
  title   = {Exploring Long-Sequence Masked Autoencoders},
  year    = {2022},
}

• This repo is a modification on the MAE repo, and supports long-sequence pretraining on both GPUs and TPUs using PyTorch.

• This repo is based on timm==0.4.12, which can be installed via pip3 install timm==0.4.12.

Fine-tuning with pre-trained checkpoints

The following table provides the pre-trained checkpoints used in the paper:

Model (pretrained w/ L=784, image size 448, patch size 16)	ViT-Base	ViT-Large
COCO (train2017 + unlabeled2017) 4000-epoch	download	download
ImageNet-1k 800-epoch	download	download
ImageNet-1k 1600-epoch	download	download

Using the codebase

• Follow PRETRAIN_LONG_SEQ_TPU.md for long-sequence pretraining on Google Cloud TPUs (which we used for our experiments).
• Follow PRETRAIN_LONG_SEQ_GPU.md for long-sequence pretraining on Nvidia GPUs.
• Follow FINETUNE_DETECTION.md to fine-tune on the object detection task using the ViTDet codebase from Detectron2.

In addition, this codebase is also compatible with the features in the original MAE repo. Follow README_MAE.md to use the features of the original MAE repo (such as fine-tuning on image classification).

License

This project is under the CC-BY-NC 4.0 license. See LICENSE for details.