PRIME: A Few Primitives Can Boost Robustness to Common Corruptions

This is the official repository of PRIME, the data agumentation method introduced in the ECCV 2022 paper "PRIME: A Few Primitives Can Boost Robustness to Common Corruptions". PRIME is a generic, plug-n-play data augmentation scheme that consists of simple families of max-entropy image transformations for conferring robustness against common corruptions. PRIME leads to significant improvements in corruption robustness on multiple benchmarks.

Pre-trained models

We provide different models trained with PRIME on CIFAR-10/100 and ImageNet datasets. You can download them from here.

Setup

This code has been tested with Python 3.8.5 and PyTorch 1.9.1. To install required dependencies run:

$ pip install -r requirements.txt

For corruption robustness evaluation, download and extract the CIFAR-10-C, CIFAR-100-C and ImageNet-C datasets from here.

Usage

We provide a script train.py for PRIME training on CIFAR-10/100, ImageNet-100 and ImageNet. For example, to train a ResNet-50 network on ImageNet with PRIME, run:

$ python -u train.py --config=config/imagenet_cfg.py \
    --config.save_dir=<save_dir> \
    --config.data_dir=<data_dir> \
    --config.cc_dir=<common_corr_dir> \
    --config.use_prime=True

Detailed configuration options for all the datasets can be found in config.

Results

Results on ImageNet/ImageNet-100 with a ResNet-50/ResNet-18 (†: without JSD loss)

Dataset	Method	Clean ↑	CC Acc ↑	mCE ↓
ImageNet	Standard	76.1	38.1	76.1
ImageNet	AugMix	77.5	48.3	65.3
ImageNet	DeepAugment	76.7	52.6	60.4
ImageNet	PRIME†	77.0	55.0	57.5
ImageNet	PRIME	75.3	56.4	55.5
ImageNet-100	Standard	88.0	49.7	100.0
ImageNet-100	AugMix	88.7	60.7	79.1
ImageNet-100	DeepAugment	86.3	67.7	68.1
ImageNet-100	PRIME	85.9	71.6	61.0

Results on CIFAR-10/100 with a ResNet-18

Dataset	Method	Clean ↑	CC Acc ↑	mCE ↓
CIFAR-10	Standard	95.0	74.0	24.0
CIFAR-10	AugMix	95.2	88.6	11.4
CIFAR-10	PRIME	93.1	89.0	11.0
CIFAR-100	Standard	76.7	51.9	48.1
CIFAR-100	AugMix	78.2	64.9	35.1
CIFAR-100	PRIME	77.6	68.3	31.7

Citing this work

@inproceedings{PRIME2022,
    title = {PRIME: A Few Primitives Can Boost Robustness to Common Corruptions}, 
    author = {Apostolos Modas and Rahul Rade and Guillermo {Ortiz-Jim\'enez} and Seyed-Mohsen {Moosavi-Dezfooli} and Pascal Frossard},
    year = {2022},
    booktitle = {European Conference on Computer Vision (ECCV)}
}