Masking Augmentation for Supervised Learning

Byeongho Heo, Taekyung Kim, Sangdoo Yun, Dongyoon Han

NAVER AI LAB

Official PyTorch implementation of MaskSub "Masking Augmentation for Supervised Learning" | arxiv.

Abstract

Pre-training using random masking has emerged as a novel trend in training techniques. However, supervised learning faces a challenge in adopting masking augmentations, primarily due to unstable training. In this paper, we propose a novel way to involve masking augmentations dubbed Masked Sub-model (MaskSub). MaskSub consists of the main-model and sub-model; while the former enjoys conventional training recipes, the latter leverages the benefit of strong masking augmentations in training. MaskSub addresses the challenge by mitigating adverse effects through a relaxed loss function similar to a self-distillation loss. Our analysis shows that MaskSub improves performance, with the training loss converging even faster than regular training, which suggests our method facilitates training. We further validate MaskSub across diverse training recipes and models, including DeiT-III, MAE fine-tuning, CLIP fine-tuning, ResNet, and Swin Transformer. Our results show that MaskSub consistently provides significant performance gains across all the cases. MaskSub provides a practical and effective solution for introducing additional regularization under various training recipes.

Updates

• Feb 28, 2024: Arxiv paper update
• Jun 21, 2023: Codes for deit, mae, swin, and resnet are released
• Jun 21, 2023: Arxiv paper is released

Getting Started

You can find MaskSub training command at each folder.

• deit/ : DeiT-III training "DeiT III: Revenge of the ViT" original repo
• mae/ : MAE finetuning "Masked Autoencoders Are Scalable Vision Learners" original repo
• swin/ : Swin Transformer training "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows" original repo
• resnet_rsb/ : ResNet training with RSB recipe "ResNet strikes back: An improved training procedure in timm" original repo

Method preview

Pseudo-code for MaskSub

It shows basic mechanism of MaskSub with simple code.

# For drop probability p
for (x, y) in data_loader:
    o1, o2 = model(x, drop_prob=0), model(x, drop_prob=p)
    loss = CrossEntropy(o1, y)
    loss += CrossEntropy(o2, softmax(o1.detach()))
    (loss/2).backward()
    optimizer.step()

Practical code for MaskSub 50%

In practice, we use gradient accumulation technique to prevent GPU memory issues. Also, we use loss_scaler for mixed precision.

for (x, y) in data_loader:
    optimizer.zero_grad()
            
    # Main model
    outputs = model(x), 
    loss = criterion(outputs, y)
    loss_scaler(loss/2, optimizer, retain_graph=False, update_grad=False)
    
    # Sub-model with masking
    outputs_sub = model(x, augsub='masking', augsub_ratio=0.5)
    loss = criterion(outputs_sub, F.softmax(outputs.detach()))
    loss_scaler(loss/2, optimizer, retain_graph=False, update_grad=True)

Performances

DeiT-III

Architecture	# params	FLOPs	400 epochs	+ MaskSub	800 epochs	+ MaskSub
ViT-S/16	22.0 M	4.6 G	80.4	81.1 (+0.7)	81.4	81.7 (+0.3)
ViT-B/16	86.6 M	17.5 G	83.5	84.1 (+0.6)	83.8	84.2 (+0.4)
ViT-L/16	304.4 M	61.6 G	84.5	85.2 (+0.7)	84.9	85.3 (+0.4)
ViT-H/14	632.1 M	167.4 G	85.1	85.7 (+0.6)	85.2	85.7 (+0.5)

MAE finetuning

Architecture	Finetuning Epochs	Baseline	+ MaskSub
ViT-B/16	100	83.6	83.9 (+0.3)
ViT-L/16	50	85.9	86.1 (+0.2)
ViT-H/14	50	86.9	87.2 (+0.3)

Swin Transformer

Architecture	# Params	FLOPs	Baseline	+ MaskSub
Swin-T	28.3 M	4.5 G	81.3	81.4 (+0.1)
Swin-S	49.6 M	8.7 G	83.0	83.4 (+0.4)
Swin-B	87.9 M	15.4 G	83.5	83.9 (+0.4)

ResNet

Architecture	# Params	FLOPs	Baseline	+ MaskSub
ResNet50	25.6 M	4.1 G	79.7	80.0 (+0.3)
ResNet101	44.5 M	7.9 G	81.4	82.1 (+0.7)
ResNet152	60.2 M	11.6 G	81.8	82.8 (+1.0)

License

Licensed under CC BY-NC 4.0

AugSub
Copyright (c) 2023-present NAVER Cloud Corp.
CC BY-NC-4.0 (https://creativecommons.org/licenses/by-nc/4.0/)

How to cite

@article{heo2023masksub,
    title={Masking Augmentation for Supervised Learning},
    author={Heo, Byeongho and Kim, Taekyung and Yun, Sangdoo and Han, Dongyoon},
    year={2023},
    journal={arXiv preprint arXiv:2306.11339},
}