Hands-On One-shot Learning with Python

Preprocessing

For training a Siamese network, we need to apply a special kind of preprocessing to the dataset. While preprocessing the dataset, we have to carefully create pairs of data points as follows:

• Pairs of similar images
• Pairs of dissimilar images

The following diagram illustrates an example of a Siamese network objective for Omniglot:

We also need to create labels accordingly for similar data points (y=1), and dissimilar data points (y=0); then, each pair is fed to the Siamese architecture. At the end of the layer, the Siamese network uses a differentiating form of the loss function to learn the differentiating features across layers. ...