Reduce the ML-cost with quantization and pruning

I wanted to try and test what this publication talks about : https://arxiv.org/abs/2307.02973

So I made a little study of the technology presented in the publication in this project.

This project demonstrates the impact of model design choices on both energy consumption and economic cost. It analyzes the weight importance within a neural network, estimates the total FLOPs required for inference, and explores how quantization and pruning affect both model performance and resource efficiency.

Project Structure

• core/: Contains the main functionality for analyzing weight usage.

  • weightusageanalyzer.py: Functions to compute weight importance for TensorFlow/Keras and PyTorch models that come from my repo : https://github.com/AngelLagr/weight-usage-analyser

• notebooks/: Jupyter notebook for demonstration purposes.

  • demo.ipynb: A notebook that loads the keras wine dataset, trains a simple model, and does a study of its cost with and without pruning and quantization.

• models/: Defines the neural network architecture.

  • simple_model.py: A simple neural network model with one hidden layer.

• data/: Functions for loading and preprocessing datasets.

  • load_wine_dataset.py: Functions to load and preprocess the keras wine dataset.

• quantization/: Contains functions for model quantization.

  • quantize.py: Functions to analyze the impact of quantization on model performance and resource usage.

• requirements.txt: Lists the dependencies required for the project.

Setup Instructions

1. Clone the repository:

   git clone https://github.com/AngelLagr/reduce-ml-cost-with-quantization-pruning.git
   

2. Install the required packages (Python version == 3.10.13):

   pip install -r requirements.txt
   

Usage

1. Open the Jupyter notebook:

   jupyter notebook notebooks/demo.ipynb
   

2. Follow the instructions in the notebook to load the Breast Cancer dataset, train the model, and visualize the weight importance.

Components Overview

• Weight Usage Analyzer: The core functionality for analyzing the importance of weights in neural networks and help for pruning it.
• Simple Model: A basic neural network architecture to demonstrate the weight usage analysis.
• Data Loading: Functions to handle the Breast Cancer dataset, including normalization and splitting.
• Quantization: Tools to reduce model size and analyze the impact on performance and energy efficiency.

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for more details.