Introduction

The introduction is the conceptual part of the documentation. Read it before the examples if traceTorch feels unusual.

traceTorch is intentionally small and opinionated: layers own their hidden states, models manage those states recursively, and sequences are processed one timestep at a time. Once that model clicks, the rest of the library feels like ordinary PyTorch.

Recommended order:

1. The Ethos of traceTorch explains why traceTorch exists and what problems it chooses to solve.

2. Stateful Models explains hidden states, timestep loops, reset/detach, and state persistence.

3. The SNNs of traceTorch explains SNN dynamics, surrogate outputs, quantization, and the 32-layer naming scheme.

4. Layer Map gives a compact map of the SNN, RNN, SSM, core, functional, and plotting modules.

Contents:

• The Ethos of traceTorch
  • Hidden states stay hidden
  • State management stays explicit
  • One timestep at a time
  • Composition over special cases
  • Opinionated defaults
  • What traceTorch is not
• Stateful Models
  • Why inherit from tt.Model?
  • Timestep loops
  • When to reset
  • Lazy state shapes
  • Saving states
• The SNNs of traceTorch
  • The membrane trace
  • Spikes, probabilities, and quantization
  • Surrogate gradients
  • Synaptic traces
  • Recurrent traces
  • Choosing a layer
  • The layer families
  • Common parameters
  • Working on non-last dimensions
  • Practical choices
• Layer Map
  • SNN layers
  • RNN layers
  • SSM layers
  • Core layers
  • Functional helpers
  • Plotting helpers