TorchLean API

Docs HomeInstallationGuideExamplesGraphsTools

NN.Examples.Quickstart.AutogradBasics

Quickstart: Autograd #

Use autograd.grad for scalar tensor functions and autograd.model.grad when model state is also being differentiated. Add value := true when the scalar value is needed with the gradient.

Jacobians, Hessians, JVPs, VJPs, and gradient stopping live in NN.Examples.DeepDives.AutogradTransforms.

Run:

scripts/lake.sh exe torchlean quickstart_autograd
source
def NN.Examples.Quickstart.AutogradBasics.exeName :
String

Subcommand name, kept in one place so the usage text and the error messages agree.

Instances For
    source
    def NN.Examples.Quickstart.AutogradBasics.meanSquare {shape : Spec.Shape} :
    TorchLean.autograd.Function shape []

    Mean squared magnitude of one tensor.

    Instances For
      source
      def NN.Examples.Quickstart.AutogradBasics.model :
      TorchLean.nn.Sequential [2] [1]

      A single affine model, y = W x + b.

      Instances For
        source
        def NN.Examples.Quickstart.AutogradBasics.runDemo :
        IO Unit

        Run the two common differentiation paths.

        Instances For
          source
          def NN.Examples.Quickstart.AutogradBasics.usage :
          String

          Help text; this demo takes no flags of its own.

          Instances For
            source
            def NN.Examples.Quickstart.AutogradBasics.main (args : List String) :
            IO Unit

            Entry point. CLI.dropDashDash lets the runner pass -- through, and requireNoArgs rejects stray flags rather than silently ignoring them.

            Instances For