Introduction

Deep learning code eventually needs a framework. The reason is practical: real models have many parameters, many intermediate tensors, repeated training steps, device placement, saved state, and gradients that are too tedious to write by hand.

PyTorch enters here, after the main ideas are already visible. You have seen tensors, batches, forward passes, losses, gradients, optimizers, initialization, regularization, evaluation, embeddings, convolution, residual paths, and numerical stability. PyTorch gives short names to those moves.

That order matters. If a framework appears too early, it can make learning feel fast while hiding the actual computation. If it appears too late, the reader never sees how real experiments are written. This chapter is the bridge.

The goal is not to memorize a framework API. The goal is to read a small PyTorch training script and say what each line is compressing.

In this chapter:

• tensors are arrays with shape, dtype, and device information
• requires_grad marks values that need derivatives
• modules package a forward rule with trainable parameters
• losses turn predictions into scalar training signals
• optimizers update parameters from gradients
• data loaders feed mini-batches to the training loop
• saved state records the pieces needed to continue or inspect an experiment

The framework is useful because it removes mechanical repetition. It should not remove understanding.