Empirical verification of the suggested hyperparameters for data augmentation using the fast.ai library

Data augmentation consists in adding slightly modified copies of the existing data to the training set, which increases the total amount of data and generally results in better results obtained by machine learning algorithms. The fast.ai library has some predefined values for data augmentation hyperparameters for visual data. It is claimed that these predefined parameters are to be the best for most data types, however, no empirical support for this statement has been provided. The aim of this research is to determine whether the suggested hyperparameter values for data augmentation in the fast.ai library are indeed optimal for the highest accuracy for image classification tasks. In order to answer this question, a detailed research was conducted, consisting of a series of experiments for subsequent data augmentation tools (rotation, magnification, contrast change, etc.). Three variables were modified for each tool: 1. maximal and minimal value of transformation (depending on the transformation type), 2. probability of the transformation, 3. padding behaviour. The results of the presented research lead to the conclusion that the suggested values of data augmentation implemented in the fast.ai library provides the good parameters of the model aimed at differentiating male and female faces, however in case of that classification slightly different parameters could be taken into consideration. The results are published in open-source repository (Open Science Framework, DOI:10.17605/OSF.IO/38UJG).