NengoTorch

NengoTorch lets users keep normal Nengo model construction and run the backend through PyTorch tensors. Build models with nengo.Network, nengo.Node, nengo.Ensemble, nengo.Connection, and nengo.Probe, then use nengo_torch.Simulator.

Install

When published to PyPI:

python -m pip install nengo-torch

From a source checkout:

python -m pip install .

For editable development:

python -m pip install -e .

For examples that load real datasets:

python -m pip install ".[examples]"

For Linux CUDA environments where Triton kernels should be available:

python -m pip install ".[triton]"

Verify The Install

Run the installed smoke test:

nengotorch-smoke --device auto

Verify the fallback-free fast/core path:

nengotorch-smoke --device auto --mode fast --require-compile-safe

Or through Python:

python -m nengo_torch --device auto --json

On GPU/Triton machines:

nengotorch-smoke --device cuda --require-triton

Minimal Use

import nengo
import torch
import nengo_torch

with nengo.Network(seed=1) as net:
    inp = nengo_torch.InputNode(size_out=1)
    out = nengo.Node(size_in=1)
    nengo.Connection(inp, out, synapse=None)
    probe = nengo.Probe(out)

x = torch.tensor([[[1.0], [2.0], [3.0]]])

with nengo_torch.Simulator(net, minibatch_size=1, progress_bar=False) as sim:
    y = sim.forward({inp: x}, n_steps=x.shape[1])[probe]

print(y.shape)

For Colab-specific setup details, see docs/colab.md. For broader install scenarios, see docs/install.md.

License

MIT. See LICENSE.