NPY loader tutorial (NumPy/PyTorch interop)

This tutorial shows how to train from .npy files (NumPy arrays), similar to a common PyTorch workflow where you:

1. prepare X.npy / y.npy in Python (NumPy / PyTorch),
2. then train a model in TorchLean by loading those files.

Generate small deterministic .npy files with python3 NN/Examples/Data/generate_toy_data.py:

• NN/Examples/Data/toy_regression_X.npy (shape 25×2, dtype float32)
• NN/Examples/Data/toy_regression_y.npy (shape 25×1, dtype float32)

Build:

• lake build NN.Examples.Data.Loaders.Npy

The tutorial code is compiled with the rest of TorchLean. For command-line model training, use the torchlean executable examples in NN/Examples/Models.

Optional flags (tutorial-specific):

• --data-dir PATH (default: NN/Examples/Data)
• --x PATH, --y PATH (override the .npy files)
• --seed S (controls shuffling and model initialization)
• --batch N
• --epochs E

Public API used here:

• Data.fromNpy (metadata)
• Data.fromNpySupervised (typed dataset from disk)
• Data.Transforms.Compose
• Data.batchLoader
• train.fitLoaderWith

def NN.Examples.Data.Loaders.Npy.nSamples : ℕ

def NN.Examples.Data.Loaders.Npy.inDim : ℕ

def NN.Examples.Data.Loaders.Npy.outDim : ℕ

def NN.Examples.Data.Loaders.Npy.mkModel {batch : ℕ} : API.nn.M (API.nn.Sequential (Shape.Mat batch inDim) (Shape.Mat batch outDim))

A small 2-layer MLP 2 -> 8 -> 1.

def NN.Examples.Data.Loaders.Npy.loadDataset (xPath yPath : System.FilePath) {α : Type} [API.Semantics.Scalar α] [API.Runtime.Scalar α] : IO (Except String (Runtime.Autograd.Train.Dataset (API.sample.Supervised α (Shape.Vec inDim) (Shape.Vec outDim))))

Load the .npy tensors, print their metadata, then apply a small input transform.

This file is intentionally interop-first: it shows the path from NumPy/PyTorch exports on disk to TorchLean's normal training API.

def NN.Examples.Data.Loaders.Npy.main (args : List String) : IO Unit