Hyperparameters for augmented training using CIFAR-10 dataset

[AdrianSegura220]

Dear Claudio,

First, I would like to say that I appreciate you sharing this repository as an implementation of your paper experiments. Moreover, I cloned the repository, and have a question regarding the hyperparameters for the augmented network trained on the CIFAR-10 dataset:

• In your paper, on page 7, you mention that data augmentation is used to make the original network more robust, using random hyperparameters values for brightness, contrast, hue, and saturation, while keeping the model that lays the best testing performance. I wanted to know whether you could share which were the exact values used, as I would like to have a set of similar results, however doing a hyperparameter search might be a bit complicated in my situation.

Thank you for your time and help, I hope you can help me out. This would also be great for the community following this paper! :)

[AdrianSegura220]

Hey @everyone, if you ever have to find hyperparameters that closely resemble the obtained results for the agumented variant of the Resnet on the CIFAR-10 dataset (results present in fig. 2), you can try using the following ones: 'aug_saturation': 0.6333019759452599 'aug_hue': 0.4542140784634742 'aug_brightness': 0.8260613787076381 'aug_contrast': 0.14959625288049752