Skip to content

Plugin for Sinabs, implementing the EXODUS algorithm for training SNNs efficiently with BPTT

License

Notifications You must be signed in to change notification settings

synsense/sinabs-exodus

Repository files navigation

DOI

sinabs-exodus

Sinabs-exodus is a plugin to the sinabs spiking neural network library. It can provide massive speedups in training and inference on GPU.

The tool is based on EXODUS, a formulation of backpropagation-through-time with surrogate gradients, that allows for efficient parallelization. EXODUS stands for EXact calculation Of Derivatives as Update to SLAYER. It builds upon the SLAYER1 algorithm, but uses mathematically accurate gradients and tends to be more robust to surrogate gradient scaling, making training less prone to suffer from exploding or vanishing gradients.

Some of the code in this library is loosely based upon slayerPytorch, the python implementation of SLAYER.

If you use any of this code please cite the following publication:

@article{bauer2022exodus,
  title={EXODUS: Stable and Efficient Training of Spiking Neural Networks},
  author={Bauer, Felix Christian and Lenz, Gregor and Haghighatshoar, Saeid and Sheik, Sadique},
  journal={arXiv preprint arXiv:2205.10242},
  year={2022}
}

Additionally, you also may cite the current version of the code directly by clicking at 'Cite this repository'.

Getting started

Prerequisites

EXODUS uses CUDA for efficient computation, so you will need a CUDA-capable GPU, and a working installation of CUDA.

If you have CUDA installed, you can use the command

$ nvcc -V

to see the installed version. The last line should say something like Build cuda_xx.x....., where x.xx is the version. Note that

$ nvidia-smi

does not show you the installed CUDA version, but only the newest version your Nvidia driver is compatible with.

You should also make sure that you have a PyTorch installation that is compatible with your CUDA version. To verify this, open a python console and run

import torch
print(torch.__version__)

The part after the + in the output is the CUDA version that PyTorch has been installed for and should match that of your system.

Installation

Installation from PyPI

The easiest way to install sinabs-exodus is via pip, from the Python Package Index (PyPI):

$pip install sinabs-exodus

Installation from source

You can also clone this repository and install from there, for instance if you want to use a specific branch. After cloning, the package can simply be installed via pip. This is a namespace package meaning that once installed this will be sharing its namespace with sinabs package.

$ pip install . 

Do not install in editable (-e) mode.

Usage

If you have used sinabs before, using EXODUS is straightforward, as the APIs are the same. You just need to import the spiking or leaky layer classes that you want to speed up from sinabs.exodus.layers instead of sinabs.layers.

Supported classes are:

  • IAF
  • LIF
  • ExpLeak

For example, instead of

from sinabs.layers import IAF

iaf = IAF()

do

from sinabs.exodus.layers import IAF

iaf = IAF()

Conversion to and from Sinabs classes

EXODUS provides convenience functions for converting EXODUS objects to their counterparts in Sinabs and vice versa in the sinabs.exodus.conversion module. In the following example, a new object exodus_model is created that is the same as sinabs_model, but with all sinabs-based layers being replaced with EXODUS equivalents, where possible. The original sinabs_model can be any torch.nn.Module object. Currently, classes that can be converted to and from EXODUS are: IAF, IAFSqueeze, LIF, LIFSqueeze, ExpLeak, and ExpLeakSqueeze.

from torch.nn import Sequential, Conv2d, AvgPool2d
from sinabs.layers import IAF
from sinabs.exodus import conversion

# This could be any torch module
sinabs_model = Sequential(Conv2d(3, 4, 1), AvgPool2d(2), IAF())

# Convert sinabs layers to exodus layers
exodus_model = conversion.sinabs_to_exodus(sinabs_model)

Converting from EXODUS to Sinabs:

new_sinabs_model = conversion.exodus_to_sinabs(exodus_model)

Frequent Issues

CUDA is not installed or version does not match that of torch

If during installation you get an error, such as

RuntimeError:
The detected CUDA version (...) mismatches the version that was used to compile
PyTorch (...). Please make sure to use the same CUDA versions.

or

OSError: CUDA_HOME environment variable is not set. Please set it to your CUDA install root.

CUDA is either not installed properly on your system or the version does not match that of torch (see above). If you do have the correct version installed and the error still comes up, try to make sure that the environment variables such as PATH and LD_LIBRARY_PATH contain references to the correct directories. Please refer to NVIDIA's installation instructions for more details on how to do this for your system.

The same holds if, while using EXODUS, you get an error like:

 undefined symbol: _ZN2at4_ops5zeros4callEN3c108ArrayRefIlEENS2

or similar.

License

Sinabs-exodus is published under AGPL v3.0. See the LICENSE file for details.

Footnotes

Footnotes

  1. Sumit Bam Shrestha and Garrick Orchard. "SLAYER: Spike Layer Error Reassignment in Time." In Advances in Neural Information Processing Systems, pp. 1417-1426. 2018.