Quickstart: Training a Small MLP

Regression with a two-layer MLP on a synthetic grid:

1. define the model with nn.Sequential!,
2. build an in-memory dataset with Data.fromTensors,
3. create a Trainer with an objective and optimizer,
4. call trainer.train, then predict with the result.

Run:

• scripts/lake.sh exe torchlean quickstart_mlp
• scripts/lake.sh exe torchlean quickstart_mlp --steps 20 --seed 3

Flags: --steps N, --seed S, --arithmetic, --execution, --device, and --show-backend.

def NN.Examples.Quickstart.SimpleMlpTrain.exeName : String

Command name used in diagnostics and by the top-level example runner.

def NN.Examples.Quickstart.SimpleMlpTrain.model : TorchLean.nn.Builder (TorchLean.nn.Sequential [2] [1])

A two-layer ReLU MLP 2 -> 8 -> 1.

def NN.Examples.Quickstart.SimpleMlpTrain.target (x : TorchLean.Tensor Float [2]) : TorchLean.Tensor Float [1]

Piecewise-linear regression target 0.8 relu(x₁ + x₂) - 0.4 relu(x₂ - x₁) + 0.2.

def NN.Examples.Quickstart.SimpleMlpTrain.data : TorchLean.Trainer.Dataset [2] [1]

Twenty-five grid points in [-1, 1]² with their targets.

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

Command-line help for the simple MLP quickstart.

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

Entry point. Parses the training flags, then runs the loop; --steps defaults to 200, which is enough for the printed loss to visibly fall without the demo taking long.