NN.Runtime.Autograd.Train: Training Utilities Overview

This directory contains lightweight, mostly-pure utilities for writing training loops on top of TorchLean's runtime autograd tape.

The goal is familiarity (a recognizable workflow) while keeping the code:

• deterministic (pure shuffles; explicit seeds),
• small (easy to audit),
• and compatible with both eager and compiled backends (via the shared tape/TorchLean.Session).

The modules here are narrow in scope: they provide the reusable training infrastructure, while architecture-specific math and constructors stay in NN.GraphSpec.Models.*, NN.API.Models.*, or user code.

Module Map

• NN.Runtime.Autograd.Train.Core: tagged errors, typed extraction of values/grads, and one-step SGD glue.
• NN.Runtime.Autograd.Train.Dataset: deterministic Dataset / DataLoader analogues with seeded shuffle.
• NN.Runtime.Autograd.Train.Eval: evaluation helpers that aggregate step reports.
• NN.Runtime.Autograd.Train.Logging: a small pluggable logger interface.
• NN.Runtime.Autograd.Train.Optim: optimizer and scheduler integration for training loops.
• NN.Runtime.Autograd.Train.TapeM: tape-building helpers for params, constants, and batch mean losses.
• NN.Runtime.Autograd.Train.TensorLoader: in-memory tensor/dataset constructors with shape checks.
• NN.Runtime.Autograd.Train.IoLoader: CSV/NPY loaders for reproducible examples and regression tests.
• NN.Runtime.Autograd.Train.Trainer: a structured trainer loop with reports and optional logging.

The training utilities are intentionally not a second model API. Layers and model constructors live in the architecture/API modules; this directory only supplies the loop machinery that feeds them.

References

• PyTorch data loading (torch.utils.data): https://pytorch.org/docs/stable/data.html
• PyTorch optimizers (torch.optim): https://pytorch.org/docs/stable/optim.html
• A canonical training/eval loop tutorial: https://pytorch.org/tutorials/beginner/basics/optimization_tutorial.html