Model training and evaluation

Although the network architecture is not that of an image classifier and the labels are not scalars, semantic segmentation can be seen as a traditional classification problem and therefore the training and evaluation processes can be the same.

For this reason, instead of writing a custom training loop, we can use the compile and fit Keras models to build the training loop and execute it respectively.

Data preparation

To use the Keras fit model, the tf.data.Dataset object should generate tuples in the (feature, label) format, where feature is the input image and label is the image label.

Therefore, it is worth defining some functions that can be applied to the elements produced by tf.data.Dataset...