Minimal, runnable introductions to the spectrax core API. Every file
is self-contained (python -m examples.01_basics.<name>) and stays
well under 100 LOC.
| File | Concepts |
|---|---|
01_module_and_forward.py |
Defining a spx.Module subclass, PyTorch-style __init__ + __call__, and running a forward pass. |
02_training_loop.py |
Hand-rolled SGD training loop with spx.value_and_grad, spx.jit(mutable="parameters"), and spx.update. |
03_export_bind.py |
Round-tripping a module via spx.export -> (GraphDef, State) and spx.bind back to a live module. |
04_state_pop_update.py |
State-pytree surgery: spx.tree_state, spx.clone, spx.update, and spx.pop on Parameter/Buffer. |
05_optimizer.py |
Training with spectrax.contrib.Optimizer wrapping optax.adam, inside a jitted step. |
06_multi_optimizer_lora.py |
Separate optimizer policies for base "parameters" and LoRA "lora" collections with MultiOptimizer. |
All examples run on CPU with small sizes (d=32, hidden=64, bs=8)
and finish in a few seconds.