Hands-On Machine Learning for Algorithmic Trading

How to prepare the data using image augmentation

CIFAR10 can also be downloaded from Keras, and we similarly rescale the pixel values and one-hot encode the ten class labels.

We first train a two-layer feedforward network on 50,000 training samples for training 20 epochs to achieve a test accuracy of 44.22%. We also experiment with a three-layer convolutional net, with 500 K parameters for 67.07% test accuracy (see notebook).

A common trick to enhance performance is to artificially increase the size of the training set by creating synthetic data. This involves randomly shifting or horizontally flipping the image, or introducing noise into the image.

Keras includes an ImageDataGenerator for this purpose, which we can configure and fit to the ...