Autograd Train

NN.Runtime.Autograd.Train is the curated umbrella for TorchLean's dynamic-tape training helpers.

This layer is intentionally about training-loop infrastructure, not model definitions:

• Core gives tagged errors plus typed value/gradient extraction from shape-erased tape data.
• Dataset gives pure, deterministic datasets/loaders with seeded shuffling.
• Trainer and Logging give small report/logging abstractions for example loops and tests.
• Eval averages reports over samples or batches while checking metric names.
• TapeM contains ergonomic tape-building helpers for params, constants, and mean losses.
• TensorLoader and IoLoader convert small in-memory/CSV/NPY datasets into typed tensors.
• Optim connects parameter tables, schedulers, and canonical optimizer equations.

The public model/training API in NN.API.* builds on these pieces. This umbrella exists so tests, examples, and downstream users can import one stable training helper surface instead of memorizing the internal file split.