This is the official PyTorch implementation of "SAGE: Saliency-Guided Mixup with Optimal Rearrangements" accepted at BMVC'20 (paper, talk, poster). Some parts of the codes are borrowed from Puzzle Mix (link).
We also provide implementations for training with mixup and its variants, in particular the ones that leverage input saliency in the data augmentation process: Mixup, SaliencyMix, Manifold Mixup, CutMix, Puzzle Mix and Co-Mixup.
@inproceedings{ma2022sage,
title= {SAGE: Saliency-Guided Mixup with Optimal Rearrangements},
author = {Ma, Avery and Dvornik, Nikita and Zhang, Ran and Pishdad, Leila and Derpanis, Konstantinos G and Fazly, Afsaneh},
booktitle = {British Machine Vision Conference (BMVC)},
year = {2022}
}
This code has been tested with
python 3.9.7
pytorch 1.10.1
cudnn 7605
torchvision 0.11.2
gco-wrapper (https://github.com/Borda/pyGCO)
matplotlib 2.1.0
numpy 1.21.2
six 1.12.0
Dataset will be downloaded at [data_path] and the results will be saved at [save_path]. If you want to run codes without saving results, please set --log_off True.
- To reproduce SAGE with PreActResNet18 for 300 epochs, run:
python3 main.py --dataset cifar10 --data_dir [data_path] --root_dir [save_path] --labels_per_class 5000 --arch preactresnet18 --learning_rate 0.2 --momentum 0.9 --decay 0.0001 --epochs 300 --schedule 100 200 --gammas 0.1 0.1 --method sage --blur_sigma 1.0 --seed 1 --kernel_size 5 --ngpu 1 --workers 2 --mixup_alpha 1.0 --upper_lambda 0.6 --rand_pos 0.01 --update_ratio 0.7
- To reproduce SAGE with PreActResNet18 for 300 epochs, run:
python3 main.py --dataset cifar100 --data_dir [data_path] --root_dir [save_path] --labels_per_class 500 --arch preactresnet18 --learning_rate 0.2 --momentum 0.9 --decay 0.0001 --epochs 300 --schedule 100 200 --gammas 0.1 0.1 --method sage --blur_sigma 1.0 --seed 1 --kernel_size 5 --ngpu 1 --workers 2 --mixup_alpha 1.0 --upper_lambda 0.5 --rand_pos 0.01 --update_ratio 0.7
- To reproduce SAGE with ResNext29-4-24 for 300 epochs, run:
python3 main.py --dataset cifar100 --data_dir [data_path] --root_dir [save_path] --labels_per_class 500 --arch resnext29_4_24 --learning_rate 0.2 --momentum 0.9 --decay 0.0001 --epochs 300 --schedule 100 200 --gammas 0.1 0.1 --method sage --blur_sigma 1.0 --seed 1 --kernel_size 5 --ngpu 1 --workers 2 --mixup_alpha 1.0 --upper_lambda 0.7 --rand_pos 0.01 --update_ratio 0.5
- To reproduce SAGE with WideResNet16-8 for 300 epochs, run:
python3 main.py --dataset cifar100 --data_dir [data_path] --root_dir [save_path] --labels_per_class 500 --arch wrn16_8 --learning_rate 0.2 --momentum 0.9 --decay 0.0001 --epochs 300 --schedule 100 200 --gammas 0.1 0.1 --method sage --blur_sigma 1.0 --seed 1 --kernel_size 5 --ngpu 1 --workers 2 --mixup_alpha 1.0 --upper_lambda 0.7 --rand_pos 0.01 --update_ratio 0.5
We follow the hyper-parameter configurations described in Co-Mixup to reproduce the baseline results, simply replace [baseline_method] with vanilla, saliencymix, input, manifold, cutmix, puzzle and comix.
Options of arch: preactresnet18, resnext29_4_24 and wrn16_8.
python3 main.py --dataset cifar10 --data_dir [data_path] --root_dir [save_path] --labels_per_class 5000 --arch [arch] --learning_rate 0.2 --momentum 0.9 --decay 0.0001 --epochs 300 --schedule 100 200 --gammas 0.1 0.1 --method [baseline_method]
python3 main.py --dataset cifar100 --data_dir [data_path] --root_dir [save_path] --labels_per_class 500 --arch [arch] --learning_rate 0.2 --momentum 0.9 --decay 0.0001 --epochs 300 --schedule 100 200 --gammas 0.1 0.1 --method [baseline_method]
MIT License