wiretorch

wiretorch builds neural network models from connectomes in PyTorch. You give it a wiring diagram and the cell type of each neuron, and it builds a network with exactly that connectivity. The networks run on GPUs with custom kernels and are differentiable, so they can be trained with any loss: matching recorded spikes, rates or calcium signals, performing a task, or anything else you can write in PyTorch.

It supports a range of single-neuron models, including conductance- and current-based spiking neurons and rate neurons. Weights can be fixed, scaled per cell-type pair, factorised or learnt one by one, and physiological parameters such as time constants and thresholds can be learnt too.

Installation

Install PyTorch first, choosing the CUDA version your driver supports (nvidia-smi shows it), then wiretorch:

pip install torch --index-url https://download.pytorch.org/whl/cu129
pip install wiretorch            # or "wiretorch[wandb]" to log to Weights & Biases

Without the first step pip installs PyPI's default PyTorch, a CUDA 13 build that needs a recent driver. The installation guide explains how to choose, and which GPUs are supported.

import wiretorch as wt

Documentation

wiretorch.readthedocs.io has a quickstart, tutorials (a ring attractor, a spiking network on a synthetic connectome, fitting connection strengths and physiology), and the API reference. Setting up a development environment and running the tests is described under Development.

wiretorch used to be called connectome_snns. That name still imports, with a deprecation warning.