Burstprop: A Python repository from jordan-g

Training on MNIST, CIFAR-10 and ImageNet Using Burstprop

Code from "Payeur, A., Guerguiev, J., Zenke, F., Richards, B., & Naud, R. (2020). Burst-dependent synaptic plasticity can coordinate learning in hierarchical circuits. bioRxiv.

System Requirements

This software requires a working Python 3.x installation as well as the following libraries:

torch
torchvision
numpy
tqdm
tensorboardx

It has been tested on CentOS Linux 7 and Python 3.6.8.

Note: Training convolutional networks requires a GPU accessible to PyTorch.

Installation

Python 3.x can be installed by downloading the appropriate installer: https://www.python.org

All of the required libraries can be installed using pip in the command line:

pip install torch torchvision numpy tqdm tensorboardx

When running the scripts to train on MNIST or CIFAR-10 for the first time, these datasets will be downloaded automatically by the torchvision library. For training on ImageNet, the dataset must be downloaded by the user (http://www.image-net.org/challenges/LSVRC/2012/), and the path to the dataset must be supplied as a command line argument. Note that downloading ImageNet can take many hours, due to the large size of the dataset.

Instructions

There are three main scripts to run:

train_mnist.py for training networks on MNIST
train_cifar10.py for training networks on CIFAR-10
train_imagenet.py for training networks on ImageNet

When running a script, a path to a directory where results will be stored must be provided:

python3 train_mnist.py path/to/results/directory

When training on ImageNet, the path to the dataset must be provided as a second argument:

python3 train_imagenet.py path/to/results/directory path/to/imagenet/dataset

Each script has a collection of arguments that can be provided to set training hyperparameters, which can be listed by using the --help argument:

python3 train_cifar10.py --help

Demo

The Jupyter notebook mnist_demo.ipynb contains a simple demo for training on MNIST using learned feedback weights and observing weight alignment. To run the notebook, Jupyter must be installed (https://jupyter.org/install). The notebook should take around 10 minutes to run on a laptop.

License