Mugs: A Multi-Granular Self-Supervised Learning Framework

This is a PyTorch implementation of Mugs proposed by our paper "Mugs: A Multi-Granular Self-Supervised Learning Framework".

Rank 1 Linear Probing and KNN Classification Results on ImageNet Without Extra Training Data

Fig 1. Overall framework of Mugs. In (a), for each image, two random crops of one image are fed into backbones of student and teacher. Three granular supervisions: 1) instance discrimination supervision, 2) local-group discrimination supervision, and 3) group discrimination supervision, are adopted to learn multi-granular representation. In (b), local-group modules in student/teacher averages all patch tokens, and finds top-k neighbors from memory buffer to aggregate them with the average for obtaining a local-group feature.

Pretrained models on ImageNet-1K

You can choose to download only the weights of the pretrained backbone used for downstream tasks, or the full checkpoint which contains backbone and projection head weights for both student and teacher networks.

Table 1. KNN and linear probing performance with their corresponding hyper-parameters, logs and model weights.

arch	params	pretraining epochs	k-nn	linear	download
ViT-S/16	21M	100	72.3%	76.4%	backbone only	full ckpt	args	logs	eval logs
ViT-S/16	21M	300	74.8%	78.2%	backbone only	full ckpt	args	logs	eval logs
ViT-S/16	21M	800	75.6%	78.9%	backbone only	full ckpt	args	logs	eval logs
ViT-B/16	85M	400	78.0%	80.6%	backbone only	full ckpt	args	logs	eval logs
ViT-L/16	307M	250	80.3%	82.1%	backbone only	full ckpt	args	logs	eval logs

Fig 2. Comparison of linear probing accuracy on ImageNet-1K.

Pretraining Settings

Environment

For reproducing, please install PyTorch and download the ImageNet dataset. This codebase has been developed with python version 3.8, PyTorch version 1.7.1, CUDA 11.0 and torchvision 0.8.2. For the full environment, please refer to our Dockerfile file.

ViT pretraining 🍺

To pretraining each model, please find the exact hyper-parameter settings at the args column of Table 1. For training log and linear probing log, please refer to the log and eval logs column of Table 1.

ViT-Small pretraining:

To run ViT-small for 100 epochs, we use two nodes of total 8 A100 GPUs (total 512 minibatch size) by using following command:

python -m torch.distributed.launch --nproc_per_node=8 main.py --data_path DATASET_ROOT --output_dir OUTPUT_ROOT --arch vit_small 
--group_teacher_temp 0.04 --group_warmup_teacher_temp_epochs 0 --weight_decay_end 0.2 --norm_last_layer false --epochs 100

To run ViT-small for 300 epochs, we use two nodes of total 16 A100 GPUs (total 1024 minibatch size) by using following command:

python -m torch.distributed.launch --nproc_per_node=16 main.py --data_path DATASET_ROOT --output_dir OUTPUT_ROOT --arch vit_small 
--group_teacher_temp 0.07 --group_warmup_teacher_temp_epochs 30 --weight_decay_end 0.1 --norm_last_layer false --epochs 300

To run ViT-small for 800 epochs, we use two nodes of total 16 A100 GPUs (total 1024 minibatch size) by using following command:

python -m torch.distributed.launch --nproc_per_node=16 main.py --data_path DATASET_ROOT --output_dir OUTPUT_ROOT --arch vit_small 
--group_teacher_temp 0.07 --group_warmup_teacher_temp_epochs 30 --weight_decay_end 0.1 --norm_last_layer false --epochs 800

ViT-Base pretraining:

To run ViT-base for 400 epochs, we use two nodes of total 24 A100 GPUs (total 1024 minibatch size) by using following command:

python -m torch.distributed.launch --nproc_per_node=24 main.py --data_path DATASET_ROOT --output_dir OUTPUT_ROOT --arch vit_base 
--group_teacher_temp 0.07 --group_warmup_teacher_temp_epochs 50 --min_lr 2e-06 --weight_decay_end 0.1 --freeze_last_layer 3 --norm_last_layer 
false --epochs 400

ViT-Large pretraining:

To run ViT-large for 250 epochs, we use two nodes of total 40 A100 GPUs (total 640 minibatch size) by using following command:

python -m torch.distributed.launch --nproc_per_node=40 main.py --data_path DATASET_ROOT --output_dir OUTPUT_ROOT --arch vit_large 
--lr 0.0015 --min_lr 1.5e-4 --group_teacher_temp 0.07 --group_warmup_teacher_temp_epochs 50 --weight_decay 0.025 
--weight_decay_end 0.08 --norm_last_layer true --drop_path_rate 0.3 --freeze_last_layer 3 --epochs 250

Evaluation

We are cleaning up the evalutation code and will release them when they are ready.

Self-attention visualization

Here we provide the self-attention map of the [CLS] token on the heads of the last layer

Fig 3. Self-attention from a ViT-Base/16 trained with Mugs.

T-SNE visualization

Here we provide the T-SNE visualization of the learned feature by ViT-B/16. We show the fish classes in ImageNet-1K, i.e., the first six classes, including tench, goldfish, white shark, tiger shark, hammerhead, electric ray. See more examples in Appendix.

Fig 4. T-SNE visualization of the learned feature by ViT-B/16.

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.

Citation

If you find this repository useful, please consider giving a star ⭐ and citation 🍺:

@inproceedings{mugs2022SSL,
  title={Mugs: A Multi-Granular Self-Supervised Learning Framework},
  author={Pan Zhou and Yichen Zhou and Chenyang Si and Weihao Yu and Teck Khim Ng and Shuicheng Yan},
  booktitle={arXiv preprint arXiv:2203.14415},
  year={2022}
}