requires_grad und der Berechnungsgraph

Training needs derivatives, but you never compute them by hand. PyTorch's autograd watches your math and works out every gradient for you. 🤖

A tensor only earns gradients when you ask. Set requires_grad to true and PyTorch starts recording every operation done to it.

Each math step on a tracked tensor is logged as a node. Together these nodes form a computation graph describing how your result was built.

Think of the graph as a recipe: inputs at the top, operations in the middle, your final loss at the bottom. Autograd reads it backward to find gradients.

PyTorch uses a dynamic graph: it is created as your code runs, not ahead of time. Normal Python loops and ifs just work inside it.

Any tensor made from a tracked one is also tracked. Here y remembers it came from x, so its grad_fn points back to that squaring step.

Tensors you create directly are leaf nodes. Tensors produced by operations are interior nodes that link back toward those leaves.

Every computed tensor carries a grad_fn telling autograd which operation made it, like PowBackward or AddBackward. Leaves have no grad_fn.

If a tensor has requires_grad false, autograd ignores it and builds no graph. That saves memory whenever you do not need gradients.

The graph is what makes the backward pass possible. Without this recorded history, PyTorch would have no way to know how the loss depends on each weight.

The beauty is you only code the forward math. Autograd silently assembles the graph so the gradients are ready the moment you ask for them.

Set requires_grad and PyTorch records your math into a computation graph. That recorded history is what autograd later reads backward to find every gradient. ✨