Training loop

The PyTorch training loop repeats, for each batch: resetting the gradients, a forward pass to predict, computing the loss, backpropagation with backward(), and the optimizer update. It alternates with an evaluation phase in eval() mode without gradient computation.

for x, y in loader:
    optimiseur.zero_grad()
    perte = critere(modele(x), y)
    perte.backward()
    optimiseur.step()

The whole glossary