Training Pipeline

Training Pipeline

Generated with Grok Build: Grok 4.5 · xAI Imagine (/imagine)

Module: src/training.jl

ModelStep

Abstract type for typed callables:

struct MyStep <: ModelStep end
(::MyStep)(model, spikes::SpikeBatch) = ...

Plain Functions and other callables with methods are also accepted.

train_step!

train_step!(model, spikes::SpikeBatch, loss_fn;
            forward_fn = nothing,
            rule::Symbol = :eprop,
            optimizer = default_optimizer(),
            kwargs...)

train_step!(model, spikes, loss_fn, model_step; rule=:eprop, kwargs...)

Argument	Contract
model	User SNN / trainable object
spikes	SpikeBatch
loss_fn	output -> Number (must be scalar)
forward_fn / positional step	(model, spikes::SpikeBatch) -> output (required)
rule	:eprop or :ottt

Returns: (model, TrainingState).

Behavior today (e986a66)

1. Validate forward_fn is callable (ArgumentError if missing / not callable).
2. Zygote.withgradient(model) around loss_fn(model_step(m, spikes)).
3. Coerce loss to Float32.
4. Branch on rule — currently prints stub messages for :eprop / :ottt (full eligibility update not wired).
5. Pack TrainingState(loss=…, gradients=…).

default_optimizer is a tiny placeholder lambda, not Optimisers.jl.

Unknown rule → error("Unknown training rule: …").

Related

• Learning Rules
• Status and Roadmap

---

Last updated: July 29, 2026 Updated by: Grok Build: Grok 4.5 Package tip reference: e986a66 (main, v0.2.0) Devin DeepWiki: commit e986a66 · SynapticDistill.jl