Learning a complex affine map from a real loss

Fit z ↦ (2-i)*z + (1/2+3i/4) at four points on the unit circle. The model, parameters, predictions, and checkpoint remain complex; only the squared-residual objective is real. autograd.complex.grad differentiates both coordinates and nn.sgdStep updates them together. This small example uses the coordinate-forward reference algorithm, not a complex reverse engine.

def NN.Examples.Models.Operators.ComplexRegression.model : TorchLean.nn.Sequential [1] [1]

An ordinary linear builder, evaluated with complex scalar arithmetic.

def NN.Examples.Models.Operators.ComplexRegression.objective {α : Type} [TorchLean.Storage α] [Context α] [Atan2 α] (state : TorchLean.nn.State (TorchLean.Complex α) (Runtime.Autograd.Model.Layers.Seq.stateShapes model)) : IO α

Half the mean squared complex residual, returned in the real component type.

def NN.Examples.Models.Operators.ComplexRegression.run (steps : ℕ := 50) (checkpoint : Option System.FilePath := none) : IO Unit

Train both complex components, optionally persisting the exact final state.

def NN.Examples.Models.Operators.ComplexRegression.main (args : List String) : IO Unit

Run with an optional step count and checkpoint path.