เรียก backward() เพื่อรับเกรเดียนต์

Once your graph is built, a single call does all the calculus. Run backward() on your output and gradients flow back to every input.

You call backward on the final number you want to minimize, usually the loss. It must be a single scalar so autograd knows where to start.

After backward runs, each leaf tensor's .grad holds the derivative of the output with respect to that tensor. The math is already done.

For y equal to x squared, calculus says the derivative is 2x. At x equal to 2 that is 4, and PyTorch fills x.grad with exactly 4.0.

backward() travels the graph from output to inputs, applying the chain rule at every node. You wrote only the forward math and got all of this free.

If your output depends on several tracked tensors, one backward call fills the .grad of each of them at once. That is how whole models update.

By default the graph is freed after backward to save memory, so calling it twice errors. You build a fresh graph on the next forward pass.

Need to call backward again on the same graph? Pass retain_graph equals true. You rarely need this, so skip it unless an error tells you to.

If your output is a vector, backward needs a gradient argument saying how to weight each element. Most of the time you sum to a scalar instead.

That single backward call is the famous backward pass of training. Forward computes predictions, backward computes how to fix the weights.

The numbers in .grad tell each weight which way to move. An optimizer reads them and nudges the weights to lower the loss a little.

Call backward() on a scalar loss and autograd applies the chain rule across the graph, filling each tensor's .grad with its derivative. That is the backward pass. 🔁