Day 6 - Using the Keras Functional API

• The functional API can be used to build nonsequential networks (e.g. wide & deep neural networks) which can make it possible for neural networks to learn both deep patterns (using the deep path) and simple rules (through thee short path).

• A model created using the functional API has to be compiled, trained, evaluated and used to make predictions the same way as a model created using the sequential API.

• A model can have multiple inputs. A different subset of features can be passed through different inputs.

• When calling the fit, evaluate and predict methods, a list of matrices has to be passed - one per input.

• A model can also have multiple outputs when:
  • The task demands it - predicting the position (regression - x and y coordinates) and class of an object (classification).
  • The are multiple independent tasks based on the same data. Multiple outputs is better than training a single neural network per task because a single neural network can learn features in the data that are useful across tasks.
  • It is used as a regularization technique - auxiliary outputs can be added to ensure lower layers are learning something useful on their own without relying on the rest of the network.

• When using multiple outputs, each output needs its own loss function.

• An instance of a Keras Model has to be created specifying which inputs and outputs to use.

• Keras will compute all losses and add them up to get the final loss used for training.

• It is possible to set different weights for each loss function (e.g. to give the final output loss more importance than the auxiliary output loss) using the loss_weights parametere.

• When training the model, labels need to be provided for each of the outputs.

• When evaluating the model, the individual losses and the total loss is returned.

• The predict method returns a prediction for each of the outputs.