Currently deep learning requires large volumes of training data to fit accurate models. In practice, however, there is often insufficient training data available and augmentation is used to expand the dataset. Historically, only simple forms of augmentation, such as cropping and horizontal flips, were used. More complex augmentation methods have recently been developed, but it is still unclear which techniques are most effective, and at what stage of the learning process they should be introduced. This paper investigates data augmentation strategies for image classification, including the effectiveness of different forms of augmentation, dependency on the number of training examples, and when augmentation should be introduced during training. The most accurate results in all experiments are achieved using random erasing due to its ability to simulate occlusion. As expected, reducing the number of training examples significantly increases the importance of augmentation, but surprisingly the improvements in generalization from augmentation do not appear to be only as a result of augmentation preventing overfitting. Results also indicate a learning curriculum that injects augmentation after the initial learning phase has passed is more effective than the standard practice of using augmentation throughout, and that injection too late also reduces accuracy. We find that careful augmentation can improve accuracy by +2.83% to 95.85% using a ResNet model on CIFAR-10 with more dramatic improvements seen when there are fewer training examples. Source code is available at https://git.io/fjPPy