thrindex 0.4.0

The open compiler for spiking neuromorphic compute.
Author spiking neural networks in Python. Compile them to neuromorphic hardware.

Website · Docs · Discussions

What is it?

thrindex is the full-stack toolkit for spiking neural networks (SNNs) — from model authoring to deployment on neuromorphic hardware.

It gives you a PyTorch-native authoring API for building and training SNNs with surrogate gradients, a deterministic Rust simulation engine for fast, reproducible model validation, and a compiler + runtime that targets real neuromorphic chips — all through one unified SDK.

import thrindex as thx
import thrindex.snn as snn

model = snn.Sequential(
    snn.Dense(784, 1000),
    snn.LIF(threshold=1.0, tau_mem=5.0),
    snn.Dense(1000, 10),
    snn.LIF(threshold=1.0, tau_mem=5.0),
)

# train with surrogate gradients — standard PyTorch loop
loss = thx.train.rate_loss(model(encoded_input), labels)
loss.backward()

Install

pip install thrindex

Requires Python 3.11+ and PyTorch 2.0+.

Quickstart

import torch
import thrindex as thx
import thrindex.snn as snn
from thrindex.encoders import rate
from thrindex.train import rate_loss

# Build a two-layer LIF network
model = snn.Sequential(
    snn.Dense(784, 1000),
    snn.LIF(threshold=1.0, tau_mem=5.0, reset="subtract"),
    snn.Dense(1000, 10),
    snn.LIF(threshold=1.0, tau_mem=5.0, reset="subtract"),
)

# Rate-encode your input (explicit generator — no global RNG state)
gen = torch.Generator()
gen.manual_seed(42)
spikes = rate(x_batch, T=25, generator=gen)   # [T, batch, 784]

# Forward + loss
spk_out = model(spikes)                        # [T, batch, 10]
loss = rate_loss(spk_out, labels)

# Standard PyTorch backward
loss.backward()
optimizer.step()

Key features

LIF neuron

The LIF module follows the exact per-timestep update order:

(i)   mem_t    = α · mem_prev + x_t          # leak + integrate
(ii)  spk_t    = H(mem_t − threshold)         # fire  (Heaviside; surrogate in backward)
(iii) mem_out  = mem_t − spk_t · threshold   # reset (subtract mode)

where α = exp(−dt / τ_mem) (exact exponential discretization, dt = 1 ms).

Canonical parameters: threshold, tau_mem, tau_syn, reset, delay.

Encoders

from thrindex.encoders import rate, latency, delta

# Rate: Bernoulli-sample — always pass an explicit generator
gen = torch.Generator(); gen.manual_seed(0)
spikes = rate(x, T=25, generator=gen)          # [T, *x.shape]

# Latency: higher value spikes earlier — deterministic, no RNG
spikes = latency(x, T=25)

# Delta: fires on change above threshold — deterministic, no RNG
spikes = delta(x_sequence, T=25, threshold=0.1)

Stack

python/thrindex/       Python SDK — strictly typed, zero logic
crates/thrindex-py/    PyO3 bridge (maturin)
crates/thrindex-sim/   Behavioral simulator — deterministic, seeded, CPU-parallel  (M2)
crates/thrindex-compiler/  Compiler: capture → validate → quantize              (M3)
crates/thrindex-numerics/  Fixed-point core — zero dependencies, consumed bit-for-bit

Rust engines are MIT/Apache-2.0. The Python SDK wraps them with no logic of its own.

Status

M5 complete — templates, public docs, CI snippet tests, conformance suite, and energy benchmark harness are all shipped. M6 (first silicon bring-up) is next.

Issues and discussions are open.
See SECURITY.md for responsible disclosure.

License

Apache 2.0 — see LICENSE.