How does Spyx compare to PyTorch-based SNN libraries?

Spyx uses JAX and Haiku for higher performance via just-in-time compilation, but it has a steeper learning curve if you're not familiar with these tools. It excels in GPU speed and custom neuron flexibility, whereas PyTorch libraries might offer broader ecosystem integration.

How to install Spyx with GPU support on Linux?

First, install JAX with GPU support following the official instructions, then use 'uv add spyx' for the library. Ensure CUDA and cuDNN are configured correctly to leverage the performance benefits mentioned in the README.

Is Spyx good for neuroevolution of spiking neural networks?

Yes, Spyx explicitly supports neuroevolution algorithms for optimizing SNN architectures and parameters, as detailed in the features and research papers, making it a strong choice for evolutionary approaches in neuromorphic computing.

What GPU memory do I need for training with Spyx?

Spyx maintains the entire dataset in GPU vRAM, so for small networks, 6GB may suffice, but for large SNNs or neuroevolution, higher memory cards are recommended. Check the hardware requirements section for specifics.

Can I use Spyx on Google Colab for free?

Not easily, as it doesn't work with Colab's older TPU versions, and GPU instances might have memory constraints. You might need to adjust datasets or use paid tiers for optimal performance with Spyx's vRAM demands.

How do I define a custom neuron model in Spyx?

The library is designed for extensibility; you can extend base classes to integrate custom dynamics, as outlined in the documentation. Examples in the research notebooks provide guidance on plugging in user-defined models.

Does Spyx support converting artificial neural networks to SNNs yet?

ANN2SNN conversion is a planned feature, so it's not currently available. Users interested in this capability may need to wait for future updates or implement their own methods using Spyx's flexible framework.

Open-Awesome

Spyx

MITJupyter Notebookv1.0.0

A compact spiking neural network library built on JAX and Haiku, offering high-performance training via surrogate gradient descent and neuroevolution.

Visit Website GitHub

137 stars14 forks0 contributors

What is Spyx?

Spyx is a spiking neural network library built on JAX and Haiku, designed for high-performance training of SNNs. It supports surrogate gradient descent and neuroevolution methods, enabling efficient optimization of spiking models for neuromorphic computing and neuroscience research.

Target Audience

Researchers and developers working in neuromorphic computing, computational neuroscience, or machine learning who need efficient tools for training and experimenting with spiking neural networks.

Value Proposition

Spyx offers a unique combination of PyTorch-like flexibility and custom CUDA kernel-level performance, making it ideal for prototyping custom neuron models while achieving state-of-the-art training speeds on GPU hardware.

Overview

Spyx: Spiking Neural Networks in JAX

Use Cases

Best For

Training spiking neural networks with surrogate gradient methods
Neuroevolution experiments for optimizing SNN architectures
Research requiring custom neuron model definitions
High-performance SNN training on GPU hardware
Prototyping advanced spiking models like SpikingRWKV
Benchmarking SNN performance against other frameworks

Not Ideal For

Projects running on Google Colab's TPUs due to JAX version incompatibility
Teams with limited GPU memory, as Spyx requires storing entire datasets in vRAM for optimal performance
Researchers needing immediate, out-of-the-box ANN2SNN conversion or advanced blocks like SpikingRWKV
Developers unfamiliar with JAX or Haiku, given the framework's reliance on these tools for setup and usage

Pros & Cons

Pros

Extreme GPU Performance

Leverages JAX's just-in-time compilation and GPU acceleration to achieve high training speeds, with benchmarks showing efficient use similar to custom CUDA kernels.

Flexible Neuron Models

Designed for easy definition and integration of user-defined dynamics, allowing researchers to prototype custom spiking behaviors without deep framework changes, as highlighted in the README.

Dual Training Methods

Supports both surrogate gradient descent and neuroevolution for SNN optimization, providing multiple pathways as evidenced by the included research and master's thesis references.

Modern Development Workflow

Uses uv for package management and includes automated release scripts, facilitating smooth development and deployment with tools like MkDocs for documentation.

Cons

High Hardware Demands

Requires significant GPU vRAM to store entire datasets during training, which can be prohibitive for systems with limited memory or shared resources, as noted in the hardware requirements.

Incomplete Feature Set

Key capabilities like ANN2SNN conversion and SpikingRWKV are listed as planned for the future, meaning current users may face gaps or need to implement these themselves.

Complex Initial Setup

Installation involves separate setup for GPU-accelerated JAX, adding extra steps compared to more integrated libraries, and compatibility issues arise with older TPU versions like on Google Colab.

Frequently Asked Questions

Related Projects

Google Research

Probabilistic reasoning and statistical analysis in TensorFlow

Stars4,421

Forks1,125

Last commit16 days ago

BRAX

Massively parallel rigidbody physics simulation on accelerator hardware.

Stars3,208

Forks345

Last commit15 days ago

Mctx

Monte Carlo tree search in JAX

Stars2,645

Forks208