Automatic Differentiation

The common transforms are named directly:

• autograd.grad differentiates scalar tensor functions and accepts (value := true);
• autograd.model.grad differentiates model losses and accepts (value := true);
• Trainer is the ordinary training API and adds optimizer updates, batching, devices, and logs.

Namespace completion after autograd. exposes the function transforms. Completion after autograd.model. exposes the model transforms and loss namespace. Named options such as value appear in each operation's signature instead of creating additional AndValue declarations.

Start with NN.Examples.Quickstart.AutogradBasics. Full Jacobians, JVPs, Hessians, and Hessian-vector products are kept in NN.Examples.DeepDives.AutogradTransforms.