Quickstart: Autograd Basics

Tour of the public autograd APIs beyond ordinary training:

• autograd.model.* for model-level VJP, Jacobian, and loss gradients.
• autograd.model.OutputLoss.* for reusable scalar losses on model outputs.
• autograd.fn1.* for Jacobian / Hessian APIs on single-input tensor functions.
• nn.functional.detach for stop-gradient behavior.

Run: lake exe torchlean quickstart_autograd

@[reducible, inline]

abbrev NN.Examples.Quickstart.AutogradBasics.XShape : Spec.Shape

@[reducible, inline]

abbrev NN.Examples.Quickstart.AutogradBasics.YShape : Spec.Shape

@[reducible, inline]

abbrev NN.Examples.Quickstart.AutogradBasics.WShape : Spec.Shape

@[reducible, inline]

abbrev NN.Examples.Quickstart.AutogradBasics.BShape : Spec.Shape

The tour stays on the public Float autodiff surface. It avoids:

• runtime tape/session code,
• hand-written parameter-shape bookkeeping,
• runtime scalar dispatch.

Instead, it uses TorchLean.autograd.* directly on a tiny fixed payload.

def NN.Examples.Quickstart.AutogradBasics.model : TorchLean.nn.Sequential XShape YShape

def NN.Examples.Quickstart.AutogradBasics.mseLoss : API.autograd.model.OutputLoss YShape YShape

def NN.Examples.Quickstart.AutogradBasics.detachedMSELoss : API.autograd.model.OutputLoss YShape YShape

def NN.Examples.Quickstart.AutogradBasics.squareFn : API.autograd.fn1.Fn XShape XShape

def NN.Examples.Quickstart.AutogradBasics.sumsqFn : API.autograd.fn1.Fn XShape TorchLean.Shape.scalar

structure NN.Examples.Quickstart.AutogradBasics.Internal.DemoPayload (α : Type) : Type

Deterministic model/sample payload for the autograd walkthrough.

The example stays small: one Linear layer, one input vector, one target vector, and one fixed parameter-direction for JVP/HVP queries.

• W : TorchLean.Tensor.T α WShape
• b : TorchLean.Tensor.T α BShape
• x : TorchLean.Tensor.T α XShape
• y : TorchLean.Tensor.T α YShape
• vW : TorchLean.Tensor.T α WShape
• vb : TorchLean.Tensor.T α BShape

def NN.Examples.Quickstart.AutogradBasics.Internal.demoPayloadF : DemoPayload Float

Fixed Float tensors used by the walkthrough.

def NN.Examples.Quickstart.AutogradBasics.Internal.modelParams {α : Type} (payload : DemoPayload α) : API.autograd.model.Params model α

Parameter pack for the single Linear layer in model.

def NN.Examples.Quickstart.AutogradBasics.Internal.paramDirection {α : Type} (payload : DemoPayload α) : API.autograd.model.Params model α

Direction vector in parameter space used for JVP/HVP examples.

def NN.Examples.Quickstart.AutogradBasics.Internal.unpackLinearParams {α : Type} (params : API.autograd.model.Params model α) : TorchLean.Tensor.T α WShape × TorchLean.Tensor.T α BShape

Unpack this tutorial's single Linear-layer parameter pack.

def NN.Examples.Quickstart.AutogradBasics.Internal.runDemo : IO Unit

Run the Float autograd walkthrough.

def NN.Examples.Quickstart.AutogradBasics.usage : String

Command-line help for the Float autograd quickstart.

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

CLI entrypoint for the Float autograd quickstart.