nndst

Basic Usage

Prerequisites

The code has the following dependencies:

• python 3.7
• pytorch
• torchvision
• Pillow (PIL)
• scipy
• networkx

to generate networkx directed network graph for both structured (EB) and unstructured (FreeTickets) do:

bash run.sh

You will be ask to answer the following prompts:

• "Cuda: " -> determines which gpu to run on
• "Dataset: " -> select dataset(s) to use [cifar10, cifar1000, imagenet]
• "Models: " -> select models to use (consult supported models in common_models/models.py)
• "batchsize: " -> select batch-size to train
• "Epochs: " -> select total epochs
• "seed: " -> set seed for random generator (important as it will also be use to genreate shared random init weights values for both structured and unstructured)
• "sparsity: " -> set sparsity level for Dynamic sparsity training (0.1-1.0)
• "ouput graph dir: " -> output dir to save networkx's gml graphs

Sample prompts

• "Cuda: " 6
• "Dataset: " cifar10, cifar100
• "Models: " resnet18 resnet34 resnet50 vgg16
• "batchsize: " 512
• "Epochs: " 100
• "seed: " 69
• "sparsity: " 0.3
• "ouput graph dir: " graphs/