SnapMix: Semantically Proportional Mixing for Augmenting Fine-grained Data (AAAI 2021)

PyTorch implementation of SnapMix | paper

Method Overview

SnapMix

Cite

@inproceedings{huang2021snapmix,
    title={SnapMix: Semantically Proportional Mixing for Augmenting Fine-grained Data},
    author={Shaoli Huang, Xinchao Wang, and Dacheng Tao},
    year={2021},
    booktitle={AAAI Conference on Artificial Intelligence},
}

Setup

Install Package Dependencies

torch
torchvision 
PyYAML
easydict
tqdm
scikit-learn
efficientnet_pytorch
pandas
opencv

Datasets

create a soft link to the dataset directory

CUB dataset

ln -s /your-path-to/CUB-dataset data/cub

Car dataset

ln -s /your-path-to/Car-dataset data/car

Aircraft dataset

ln -s /your-path-to/Aircraft-dataset data/aircraft

Training

Training with Imagenet pre-trained weights

1. Baseline and Baseline+

To train a model on CUB dataset using the Resnet-50 backbone,

python main.py # baseline

python main.py --midlevel # baseline+

To train model on other datasets using other network backbones, you can specify the following arguments:

--netname: name of network architectures (support 4 network families: ResNet,DenseNet,InceptionV3,EfficientNet)

--dataset: dataset name

For example,

python main.py --netname resnet18 --dataset cub # using the Resnet-18 backbone on CUB dataset

python main.py --netname efficientnet-b0 --dataset cub # using the EfficientNet-b0 backbone on CUB dataset

python main.py --netname inceptoinV3 --dataset aircraft # using the inceptionV3 backbone on Aircraft dataset

2. Training with mixing augmentation

Applying SnapMix in training ( we used the hyperparameter values (prob=1., beta=5) for SnapMix in most of the experiments.):

python main.py --mixmethod snapmix --beta 5 --netname resnet50 --dataset cub # baseline

python main.py --mixmethod snapmix --beta 5 --netname resnet50 --dataset cub --midlevel # baseline+

Applying other augmentation methods (currently support cutmix,cutout,and mixup) in training:

python main.py --mixmethod cutmix --beta 3 --netname resnet50 --dataset cub # training with CutMix

python main.py --mixmethod mixup --prob 0.5 --netname resnet50 --dataset cub # training with MixUp

3. Results

ResNet architecture.

Backbone	Method	CUB	Car	Aircraft
Resnet-18	Baseline	82.35%	91.15%	87.80%
Resnet-18	Baseline + SnapMix	84.29%	93.12%	90.17%
Resnet-34	Baseline	84.98%	92.02%	89.92%
Resnet-34	Baseline + SnapMix	87.06%	93.95%	92.36%
Resnet-50	Baseline	85.49%	93.04%	91.07%
Resnet-50	Baseline + SnapMix	87.75%	94.30%	92.08%
Resnet-101	Baseline	85.62%	93.09%	91.59%
Resnet-101	Baseline + SnapMix	88.45%	94.44%	93.74%
Resnet-50	Baseline+	87.13%	93.80%	91.68%
Resnet-50	Baseline+ + SnapMix	88.70%	95.00%	93.24%
Resnet-101	Baseline+	87.81%	93.94%	91.85%
Resnet-101	Baseline+ + SnapMix	89.32%	94.84%	94.05%

InceptionV3 architecture.

Backbone	Method	CUB
InceptionV3	Baseline	82.22%
InceptionV3	Baseline + SnapMix	85.54%

DenseNet architecture.

Backbone	Method	CUB
DenseNet121	Baseline	84.23%
DenseNet121	Baseline + SnapMix	87.42%

Training from scratch

To train a model without using ImageNet pretrained weights:

python main.py --mixmethod snapmix --prob 0.5 --netname resnet18 --dataset cub --pretrained 0 # resnet-18 backbone

python main.py --mixmethod snapmix --prob 0.5 --netname resnet50 --dataset cub --pretrained 0 # resnet-50 backbone

2. Results

Backbone	Method	CUB
Resnet-18	Baseline	64.98%
Resnet-18	Baseline + SnapMix	70.31%
Resnet-50	Baseline	66.92%
Resnet-50	Baseline + SnapMix	72.17%

Shaoli-Huang / SnapMix

readme