Autograd : différentiation automatique pour la rétropropagation

Automatic differentiation (autograd) is the engine that computes gradients in PyTorch without requiring the programmer to derive them by hand. Unlike numerical differentiation (finite differences) or symbolic differentiation (algebra), autograd works by recording operations on tensors at runtime and replaying them in reverse. This makes it possible to train models of any architecture efficiently and correctly.

Gradient tracking is disabled by default. Setting requires_grad=True tells PyTorch to record all operations on that tensor so gradients can be computed later. Leaf tensors (parameters) require gradients; intermediate tensors inherit gradient-tracking from their parents automatically. Model parameters created by nn.Module have requires_grad=True set automatically.

As you perform operations on tensors with requires_grad=True, PyTorch builds a dynamic computation graph (a directed acyclic graph of operations). Each node stores the operation and references to its inputs. When .backward() is called, PyTorch traverses this graph in reverse (from output to inputs) using the chain rule to compute partial derivatives for every leaf tensor. The graph is re-built fresh on each forward pass, enabling dynamic architectures.

Calling .backward() on a scalar tensor triggers backpropagation through the entire computation graph, depositing gradients into the .grad attribute of each leaf tensor. If the output is not scalar, you must provide a gradient tensor of the same shape (the upstream gradient). Gradients accumulate by default — always call optimizer.zero_grad() (or tensor.grad.zero_()) before the next backward pass to avoid incorrect updates.

The chain rule states that the derivative of a composition of functions is the product of derivatives. Autograd applies this mechanically through every operation in the graph. Understanding how gradients flow helps debug issues like vanishing gradients (products near zero) and exploding gradients (products greater than 1 repeated many times). Activation function choice (ReLU vs sigmoid) directly affects gradient magnitudes.

During inference (prediction on new data), you do not need gradients — computing them wastes time and memory. Wrapping inference code with torch.no_grad() disables the computation graph entirely, speeding up forward passes and reducing memory usage by up to 50%. This is also used during evaluation loops to get validation metrics without influencing training.

PyTorch accumulates (adds) gradients into .grad each time .backward() is called. This is intentional for some advanced techniques, but during normal training it means you must zero gradients before each backward pass. The standard idiom uses the optimizer's zero_grad() method, which clears gradients for all parameters the optimizer manages. Forgetting this step leads to gradients growing unboundedly, corrupting updates.

Inspecting gradient values is crucial for diagnosing training problems. You can access them via tensor.grad after calling backward. Plotting the gradient norm across layers reveals whether gradients vanish (near zero) or explode (very large) in deep networks. A common debugging hook is to print gradient statistics after every N batches to detect instability early in training.

Sometimes you want to use a computed tensor as a constant input to another computation — without letting gradients flow through it. .detach() returns a new tensor that shares data but is removed from the computation graph. This is used in target networks (reinforcement learning), stopping gradient flow in specific branches of a network, and converting tensors to NumPy arrays for plotting.

PyTorch supports higher-order differentiation. Passing create_graph=True to .backward() keeps the computation graph for the gradient computation itself, allowing you to differentiate through gradients. This is used in meta-learning (MAML), gradient penalty regularisation (Wasserstein GAN), and any technique that optimises with respect to gradients.

Autograd is the foundation of every neural network training loop. The standard four-step pattern is: (1) zero gradients with optimizer.zero_grad(), (2) forward pass to compute predictions and loss, (3) backward pass with loss.backward() to compute gradients, and (4) optimizer step with optimizer.step() to update parameters. Understanding that autograd is doing the heavy calculus work lets you focus on model architecture and hyperparameters.

In this lesson you learned: autograd builds a dynamic computation graph recording every operation on tensors with requires_grad=True, .backward() computes gradients via the chain rule and deposits them in .grad, and gradients accumulate so you must zero them before each training step. Next up we explore building a feedforward network with nn.Module.