Having learned to define model architectures in PyTorch by comparing with Keras, we now turn to the procedures for training and evaluating these models. If you are accustomed to TensorFlow Keras, you have likely used model.compile() to specify the optimizer and loss function, and then model.fit() to manage the training iterations. PyTorch adopts a more explicit methodology, where you construct the training loop from fundamental components. This approach provides increased flexibility and a more transparent understanding of the entire training process.

This chapter will guide you through building these custom training and evaluation loops in PyTorch. We will examine how loss functions (from torch.nn or torch.nn.functional) and optimization algorithms (within the torch.optim package) are implemented and utilized, contrasting them with their TensorFlow Keras equivalents. You will learn the standard PyTorch training cycle: executing a forward pass to get predictions, calculating the loss, computing gradients via loss.backward(), and updating model parameters using optimizer.step(). Furthermore, we will address how to structure model evaluation, effectively track performance metrics, and incorporate mechanisms for training control analogous to Keras Callbacks.