Awesome

TrivialAugment

This is the official implementation of TrivialAugment (https://arxiv.org/abs/2103.10158), as was used for the paper. TrivialAugment is a super simple, but state-of-the-art performing, augmentation algorithm.

We distribute this implementation with two main use cases in mind. Either you only use our (re-)implementetations of practical augmentation methods or you start off with our full codebase.

Use TrivialAugment and Other Methods in Your Own Codebase

If you are using PyTorch and just need the best working variant of TrivialAugment, we recommend to use the TrivialAugmentWide implementation in torchvision: https://pytorch.org/vision/main/generated/torchvision.transforms.TrivialAugmentWide.html.

If you want more control, we recommend to simply copy over the file aug_lib.py to your codebase. You can now instantiate the augmenters TrivialAugment, RandAugment and UniAugment like this:

augmenter = aug_lib.TrivialAugment()

And simply use them on a PIL images img:

aug_img = augmenter(img)

This format also happens to be compatible with torchvision.transforms. If you do not have Pillow or numpy installed, do so by calling pip install Pillow numpy. Generally, a good position to augment an image with the augmenter is right as you get it out of the dataset, before you apply any custom augmentations.

The default augmentation space is fixed_standard, that is without AutoAugments posterization bug and using the set of augmentations used in Randaugment. This is the search space we used for all our experiments, that do not mention another augmentation space. You can change the augmentation space, though, with aug_lib.set_augmentation_space. This call for example

aug_lib.set_augmentation_space('fixed_custom',2,['cutout'])

will change the augmentation space to only ever apply cutout with a large width or nothing. The 2 here gives indications in how many strength levels the strength ranges of the augmentation space should be divided. If an augmentation space includes sample_pairing, you need to specify a set of images with which to pair before each step: aug_lib.blend_images = [LIST OF PIL IMAGES].

Our recommendation is to use the default fixed_standard search space for very cheap setups, like Wide-Resnet-40-2, and to use wide_standard for all other setups by calling aug_lib.set_augmentation_space('wide_standard',31) before the start of training.

Use Our Full Codebase

Clone this directory and cd into it.

git clone automl/trivialaugment
cd trivialaugment

Install a fitting PyTorch version for your setup with GPU support, as our implementation only support setups with at least one CUDA device and install our requirements, we used python 3.8 and pytorch 1.5.0 for our experiments:

pip install -r requirements.txt
# Install a pytorch version, in many setups this has to be done manually, see pytorch.org

Now you should be ready to go. Start a training like so:

python -m TrivialAugment.train -c confs/wresnet40x2_cifar100_b128_maxlr.1_ta_fixedsesp_nowarmup_200epochs.yaml --dataroot data --tag EXPERIMENT_NAME

For concrete configs of experiments from the paper see the comments in the papers LaTeX code around the number you want to reproduce. For logs and metrics use a tensorboard with the logs directory or use our aggregate_results.py script to view data from the tensorboard logs in the command line.

Confidence Intervals

Since in the current literature we rarely found confidence intervals, we share our implementation in evaluation_tools.py.

Citation

If you use TrivialAugment in scientific publications, we would appreciate a citation of the following paper:

TrivialAugment: Tuning-free Yet State-of-the-Art Data Augmentation
Samuel Müller and Frank Hutter
ICCV 2021 oral

Link to publication.

@InProceedings{Muller_2021_ICCV,
    author    = {M\"uller, Samuel G. and Hutter, Frank},
    title     = {TrivialAugment: Tuning-Free Yet State-of-the-Art Data Augmentation},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {774-782}
}

Shoutouts

This repository uses code from https://github.com/ildoonet/pytorch-randaugment and from https://github.com/tensorflow/models/tree/master/research/autoaugment.