Python Deep Learning - Second Edition

Data augmentation

One of the most efficient regularization techniques is data augmentation. If the training data is too small, the network might start to overfit. Data augmentation helps counter this by artificially increasing the size of the training set. Let's use an example. In the MNIST and CIFAR-10 examples, we've trained the network over multiple epochs. The network will "see" every sample of the dataset once per epoch. To prevent this, we can apply random augmentations to the images, before using them for training. The labels will stay the same. Some of the most popular image augmentations are:

• Rotation
• Horizontal and vertical flip
• Zoom in/out
• Crop
• Skew
• Contrast and brightness adjustment

The emboldened augmentations are shown in ...