Probabilistic Pruning of Neural Networks

Overview

This repository contains the resources for implementing probabilistic pruning on Deep Neural Networks (DNNs). The main notebooks, resnet_optimized.ipynb and EfficientNet_optimized.ipynb, along with the supporting script pruning_funcs.py, are detailed below.

Primary Resources

resnet_optimized.ipynb and EfficientNet_optimized.ipynb:

These notebooks are dedicated to training and testing the ResNet and EfficientNet models. They include comprehensive instructions for model definition, training procedures, and the application of various pruning techniques to assess sparsity and accuracy. These notebooks serve as the primary sources for experimental analysis and contain outputs that are integral to our reports. Due to the probabilistic nature of the experiments, results may vary slightly in each run, reflecting the stochastic nature of the process.

State Files

resnet_state, efficientnet_state:

These binary files store the trained model weights for the ResNet and EfficientNet models, enabling reproducibility and further experimentation without the need to retrain from scratch.

Utility Script

pruning_funcs.py:

This script encompasses all the core logic functions necessary for pruning models documented in the notebooks above.

Legacy Files

deprecated_working_trained.ipynb, deprecated_state:

These files contain earlier iterations of our work and are maintained for historical reference.

Dependencies

To run the notebooks and scripts provided in this repository, the following packages are required:

• torch
• torchvision
• numpy
• matplotlib
• timm

This setup guide aims to facilitate a structured approach to exploring probabilistic pruning within DNNs, ensuring clarity and consistency in replication and experimentation.