traceTorch Documentation

traceTorch is a PyTorch library for stateful recurrent layers, built primarily for spiking neural networks.

It keeps model code close to ordinary PyTorch while handling the parts that usually make recurrent and spiking models awkward: hidden-state initialization, hidden-state resets, truncated history, state persistence, parameter constraints, and explicit feature-last composition.

The short version is:

• Inherit models from tt.Model.

• Use traceTorch layers inside ordinary PyTorch modules.

• Loop over timesteps yourself.

• Call reset_states() when a new sequence starts.

• Call detach_states() when you want online or truncated learning.

Where to go first

If you are new to traceTorch, read the documentation in this order:

1. Installation shows how to install the package and run examples.

2. Quickstart builds the smallest useful model and training loop.

3. Introduction explains the mental model: hidden states, timestep loops, SNN naming, and layer choices.

4. Examples walks through the runnable scripts in the repository.

5. Tutorials covers save/load, compile/decompile, and custom layer creation.

6. Reference links the API docstrings.

What traceTorch provides

tt.snn

32 leaky-integrator-based SNN layers. The family covers continuous readout layers, binary firing layers, ternary firing layers, scaled ternary layers, dual sign-specific traces, synaptic traces, recurrent traces, and combinations of those mechanisms.

tt.rnn

SimpleRNN, LSTM, and GRU layers adapted to traceTorch’s state-management style. They process one timestep at a time and keep their hidden states internal.

tt.Model and tt.Layer

The core abstractions. tt.Layer gives each layer states, constrained parameters, compilation, and per-state shape rules. tt.Model recursively manages every traceTorch layer in a model tree.