Hyperflux Experiments

Overview

This script allows you to run predefined experiments for training and pruning various neural network models (LeNet300, ResNet50, VGG19) on datasets such as MNIST, CIFAR-10, CIFAR-100, and ImageNet. The focus is on exploring different model sparsities. You also have codes to train networks from scratch for Cifar10/100.

Prerequisites

• Python 3.x
• Required dependencies: Ensure you have all necessary packages installed. You might have a requirements.txt file for this.
• Pytorch dependencies are not included in requirements.txt, make sure to install them from the official website

Usage + examples

The script is controlled via command-line arguments to select and run specific experiments.

Basic Command

python entry.py --experiment <experiment_name> [--sparsity <percentage>]

Running bottleneck experiment

python entry.py --experiment lenet300_bottleneck

Running convergencec MNIST experiment

python entry.py --experiment lenet300_convergence

Run resnet50 cifar100 at 98% sparsity

python entry.py --experiment resnet50_cifar100 --sparsity 98

Run Vgg19 cifar10 at 95% sparsity

python entry.py --experiment vgg19_cifar10 --sparsity 95

For the imagenet experiments, we created predefined files to run for each sparsity, as they are more sensible to regrowth and pruning. The commands are the following

For 96.5% sparsity

python run_experiments.py --experiment resnet50_imagenet_96.5

For 95% sparsity

python run_experiments.py --experiment resnet50_imagenet_95

For 90% sparsity

python run_experiments.py --experiment resnet50_imagenet_90