This is the official code for the paper "Quantized rewiring: hardware-aware training of sparse deep neural networks" for training sparse deep neural networks while considering hardware limitations.
"Quantized rewiring: hardware-aware training of sparse deep neural networks."
Petschenig, Horst and Robert Legenstein.
Neuromorphic Computing and Engineering 3.2 (2023): 024006.
https://doi.org/10.1088/2634-4386/accd8f
You will have to install PyTorch and PyTorch Lightning to run this code. The dependencies are listed in environment.yml. If you use Conda, you can install the environment via
conda env create -f environment.yml --name quantized_rewiring
to install all required packages and dependencies.
In this task we have tested the applicability of our approach on the well-known sequential MNIST benchmark which is a difficult temporal credit-assignment problem. To start training, run
python train_seq_mnist.py
The logs directory will contain .csv files that track the training progress.