Skills Reference¶

btorch ships with built-in skills that encode canonical usage patterns for neuromorphic modeling. If you use AI agent, you can invoke these skills explicitly (e.g., "use the btorch-snn-modelling skill") to get vetted, context-aware assistance.

This page summarizes what each skill covers and links to the relevant source files and examples.

btorch-snn-modelling¶

When to invoke it: Whenever you are building or training spiking neural networks with btorch.

What it covers:

• Stateful modules — MemoryModule, init_net_state, reset_net, checkpointing
• The dt environment — environ.context(dt=...) usage
• Training loops — plain PyTorch and Lightning integration
• Checkpointing — saving/loading memories_rv with state_dict()
• Truncated BPTT — detach_net for long sequences
• Common pitfalls — forgetting dt, wrong state names, missing resets

Key references:

• Skill source: skills/btorch-snn-modelling/SKILL.md
• Full training loop: skills/btorch-snn-modelling/references/training_example.md
• Plain PyTorch example: examples/fmnist.py
• Lightning example: examples/fmnist_lightning.py
• Tests: tests/models/test_mem_load_save.py

omegaconf-config¶

When to invoke it: When you need structured configuration, CLI overrides, or launcher-to-worker option forwarding.

What it covers:

• Dataclass-first config — defaults live in Python, not YAML
• Composition — nested dataclasses for common + task-specific settings
• Variant selection — dataclass unions with _type_
• Option forwarding — to_dotlist for spawning worker processes
• Diff utilities — diff_conf for comparing configs

Key references:

• Skill source: skills/omegaconf-config/SKILL.md
• Utilities: btorch/utils/conf.py
• Tests: tests/utils/test_conf.py
• Guide: Configuration Guide

How to Reference Skills¶

When prompting an agent, be explicit:

"Use the btorch-snn-modelling skill to help me write a training loop with truncated BPTT."

"Use the omegaconf-config skill to set up a batched parameter sweep with launcher-to-worker forwarding."