Fixed-Sample Training Helpers (API)

Many runnable examples in NN/Examples/Models/* follow the same pattern:

1. build a model with nn.withModel,
2. wrap it as a ScalarModuleDef (model + supervised loss),
3. load or synthesize one supervised sample (x, y),
4. run steps optimizer updates on that fixed sample, and
5. either print loss0 -> loss1 or write a TrainLog curve.

This module keeps that loop in one place so examples stay short and consistent.

Scope:

• it trains against one fixed sample supplied by the caller;
• it is model-agnostic: callers supply the loss wrapper and optimizer constructor;
• it is backend-agnostic: callers can use it on CPU or CUDA via TorchLean.Options.

For dataset-backed training, use the TorchLean.Trainer facade exported by NN or the shared model-zoo loader helpers.

structure NN.API.Models.TrainFixed.LossPair (α : Type) : Type

Before/after scalar losses for a fixed-sample training run.

• loss0 : α
• loss1 : α

def NN.API.Models.TrainFixed.instReprLossPair.repr {α✝ : Type} [Repr α✝] : LossPair α✝ → ℕ → Std.Format

@[implicit_reducible]

instance NN.API.Models.TrainFixed.instReprLossPair {α✝ : Type} [Repr α✝] : Repr (LossPair α✝)

def NN.API.Models.TrainFixed.steps {α : Type} [Semantics.Scalar α] [DecidableEq Shape] [ToString α] [Runtime.Scalar α] [Runtime.Autograd.Torch.Internal.CudaBridge.TensorConv α] {σ τ : Shape} (mkModel : nn.M (nn.Sequential σ τ)) (mkModuleDef : (model : nn.Sequential σ τ) → Runtime.Autograd.TorchLean.ScalarModuleDef (TorchLean.NN.Seq.paramShapes model) [σ, τ]) (mkOptim : (Float → α) → (paramShapes : List Shape) → Runtime.Autograd.TorchLean.Optimizer α paramShapes) (cast : Float → α) (opts : Runtime.Autograd.Torch.Options) (sample : SupervisedSample α σ τ) (steps : ℕ) (cudaMemWatch : ℕ := 0) : IO (LossPair α)

One fixed-sample run for an arbitrary scalar backend.

def NN.API.Models.TrainFixed.curveFloat {σ τ : Shape} (mkModel : nn.M (nn.Sequential σ τ)) (mkModuleDef : (model : nn.Sequential σ τ) → Runtime.Autograd.TorchLean.ScalarModuleDef (TorchLean.NN.Seq.paramShapes model) [σ, τ]) (mkOptim : (paramShapes : List Shape) → Runtime.Autograd.TorchLean.Optimizer Float paramShapes) (opts : Runtime.Autograd.Torch.Options) (sample : SupervisedSample Float σ τ) (steps : ℕ) (cudaMemWatch : ℕ := 0) : IO Runtime.Training.Curve

Fixed-sample run specialized to Float, returning a full per-step curve.