Here is an overview of the files needed to create ImageNet training data augmented with DeepAugment. Alternatively, you can download our EDSR and CAE images directly.

• EDSR_distort_imagenet.py: Creates a distorted version of ImageNet using EDSR (https://arxiv.org/abs/1707.02921)
• CAE_distort_imagenet.py: Creates a distorted version of ImageNet using a CAE (https://arxiv.org/abs/1703.00395)

The above scripts can be run in parallel to speed up dataset creation (multiple workers processing different classes). For example, to split dataset creation across 5 processes, run the following in parallel:

    CUDA_VISIBLE_DEVICES=0 python3 EDSR_distort_imagenet.py --total-workers=5 --worker-number=0
    CUDA_VISIBLE_DEVICES=1 python3 EDSR_distort_imagenet.py --total-workers=5 --worker-number=1
    CUDA_VISIBLE_DEVICES=2 python3 EDSR_distort_imagenet.py --total-workers=5 --worker-number=2
    CUDA_VISIBLE_DEVICES=3 python3 EDSR_distort_imagenet.py --total-workers=5 --worker-number=3
    CUDA_VISIBLE_DEVICES=4 python3 EDSR_distort_imagenet.py --total-workers=5 --worker-number=4

You will need to change the save path and original ImageNet train set path.

In addition to EDSR and CAE, the DeepAugment approach works with randomly sampled architectures. We call an example Noise2Net, which can generate augmentations in memory and in parallel. See the code above.

ResNet-50 + DeepAugment (Noise2Net)

ResNeXt-101 + AugMix + DeepAugment (Noise2Net, EDSR, CAE)