Module 03 Summary

The idea this module was built around

Backpropagation is the chain rule applied backwards through a graph you already built. Work it once by hand and it stops being mysterious.

What you can now do

• Explain why the loss must be a single scalar with a grad_fn
• Verify a gradient by hand and confirm it with autograd
• Order the three steps correctly: zero, backward, step
• Checkpoint on validation and reload a run

The trap this module removed

Gradients accumulate. Two backward passes without clearing give the sum of both. And since PyTorch 2.0, zero_grad() releases the tensors entirely — .grad becomes None, so code reaching for .grad.abs() immediately after raises AttributeError.

Figures worth remembering

The worked chain rule: a.grad 150.0, b.grad 90.0. Training on a known rule recovers [2.0, -3.0] and bias 1.0.

Before you move on

Run your loop on data with a known answer. If it does not recover the rule, the loop is wrong, however nicely the loss curve falls.

Mark complete