This is an unofficial PyTorch implementation of Masked Autoencoders Are Scalable Vision Learners for self-supervised ViT. This repo is mainly based on moco-v3, pytorch-image-models, BEiT and MAE-pytorch.

• visualization of reconstruction image
• linear probing
• k-NN classification
• more results
• more datasets
• transfer learning for detection and segmentation
• multi-nodes training
• ...

We support two representations (repre.) for classification: GAP (Global Average Pooling) and Cls-token. According to paper, MAE works similarily well with both of them. In Cls-token mode, it is trained in encoder of MAE.

For k-NN evaluation, we use k=10 as default.

• We finetune models by 50 epochs as default. For 400 and 800 epochs pretraining, we use 50 / 100 / 150 epochs for fine-tuning (logs and weights provided under 50 epochs).
• BaiduNetdisk （2lt1）

• Following paper, we finetune models by 100 epochs as default.
• BaiduNetdisk (k2ef)

• Following paper, we finetune models by 50 epochs as default.
• BaiduNetdisk (825g)

The code has been tested with CUDA 11.4, PyTorch 1.8.2.

1. The batch size specified by -b is batch-size per card.
2. The learning rate specified by --lr is the base lr (corresponding to 256 batch-size), and is adjusted by the linear lr scaling rule.
3. In this repo, only multi-gpu, DistributedDataParallel training is supported; single-gpu or DataParallel training is not supported.
4. We support cls-token (token) and global averaging pooling (GAP) for classification. Please verify the correspondence of pretraining and finetuning/linear probing. For cls-token mode during pretraining, cls-token is trained in encoder.

Below is examples for MAE pre-training.

sh run_pretrain.sh \
	--config cfgs/pretrain/Vit-S_100E_GAP.yaml \
	--data_path /path/to/train/data

sh run_finetune.sh \
	--config cfgs/finetune/ViT-S_50E_GAP.yaml \
	--data_path /path/to/data \
	--finetune /path/to/pretrain/model

According to paper, we have two training modes: SGD + 4096 batch-size and LARS + 16384 batch-size.

sh run_lincls.sh \
	--config cfgs/lincls/ViT-S_SGD_GAP.yaml \
	--data_path /path/to/data \
	--finetune /path/to/pretrain/model

sh run_knn.sh \
	--config cfgs/finetune/ViT-S_50E_GAP.yaml \
	--data_path /path/to/data \
	--finetune /path/to/pretrain/model \
	--save_path /path/to/save/result

python tools/run_mae_vis.py \
	--config cfgs/pretrain/ViT-B_400E_Norm_GAP.yaml \
	--save_path output/restruct/ \
	--model_path /path/to/pretrain/model \
	--img_path /path/to/image

python tools/vit_explain.py
--config cfgs/finetune/ViT-S_50E_CLS-Token.yaml
--finetune /path/to/pretrain/model
--image_path /path/to/image
--head_fusion max
--discard_ratio 0.9

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

If you use the code of this repo, please cite the original paper and this repo:

@Article{he2021mae,
  author  = {Kaiming He* and Xinlei Chen* and Saining Xie and Yanghao Li and Piotr Dolla ́r and Ross Girshick},
  title   = {Masked Autoencoders Are Scalable Vision Learners},
  journal = {arXiv preprint arXiv:2111.06377},
  year    = {2021},
}

@misc{yang2021maepriv,
  author       = {Lu Yang* and Pu Cao* and Yang Nie and Qing Song},
  title        = {MAE-priv},
  howpublished = {\url{https://github.com/BUPT-PRIV/MAE-priv}},
  year         = {2021},
}