Model Optimization Workshop

This workshop provides a hands-on experience for optimizing machine learning models to improve performance and reduce costs. You'll learn various optimization techniques including quantization, pruning, knowledge distillation, and fine-tuning.

Dependency Management

This workshop uses a minimal dependency approach where each notebook manages its own dependencies:

• Each notebook installs only the packages it needs
• Version requirements are specified directly in each notebook
• Common settings are stored in workshop_config.json

Workshop Structure

1. Introduction and Setup - Environment setup and model download
2. Quantization - Applying quantization techniques
3. Pruning - Implementing pruning techniques
4. Knowledge Distillation - Creating smaller student models
5. Fine-tuning - Adapting models to specific tasks
6. Cost Analysis - Analyzing cost implications and ROI
7. Resource Cleanup - Cleaning up AWS resources

Prerequisites

• AWS account with access to SageMaker
• Basic understanding of machine learning and deep learning concepts
• Familiarity with Python and PyTorch

Setup Instructions

Automated Setup with CloudFormation

The easiest way to set up the workshop environment is using the provided CloudFormation template:

1. Navigate to the cloudformation directory
2. Follow the instructions in the README.md file to deploy the CloudFormation stack
3. After the stack is deployed, follow the "Getting Started" instructions to clone the repository and set up the workshop environment

Manual SageMaker Setup

If you prefer to set up the environment manually:

1. Launch a SageMaker notebook instance with the following specifications:

   • Instance type: ml.t3.xlarge (minimum) or ml.g4dn.xlarge (recommended for GPU acceleration)
   • Use the standard ML AMI provided by SageMaker
   • Attach an IAM role with the necessary permissions (AmazonSageMakerFullAccess and S3 access)

2. Clone this repository to your SageMaker notebook instance:

   git clone https://github.com/SaRedfiche/model-optimization-workshop.git
   cd model-optimization-workshop

3. Install the required packages:

   pip install -r requirements.txt

4. Open the first notebook 01_introduction_and_setup.ipynb and follow the instructions.

Local Setup (Alternative)

If you prefer to run the workshop locally:

1. Clone this repository:

   git clone https://github.com/SaRedfiche/model-optimization-workshop.git
   cd model-optimization-workshop

2. Create and activate a virtual environment:

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install the required packages:

   pip install -r requirements.txt

4. Configure AWS credentials:

   aws configure

5. Open the first notebook 01_introduction_and_setup.ipynb and follow the instructions.

Recommended Instance Types

• Notebook Instance:

  • ml.t3.medium: Good for initial notebooks and smaller models (4 vCPUs, 16 GB memory)
  • ml.t3.large: Better for more demanding notebooks (2 vCPU, 8 GiB memory)
  • ml.g4dn.xlarge: For GPU-accelerated tasks (4 vCPU, 16 GiB memory, 1 GPU)

• Optimization Instance (for distributed processing):

  • ml.c5.xlarge: Good for CPU-based quantization (4 vCPU, 8 GiB memory)
  • ml.g4dn.xlarge: For GPU-accelerated tasks (4 vCPU, 16 GiB memory, 1 GPU)

Documentation

• cloudformation/README.md - Instructions for deploying with CloudFormation
• model_optimization_techniques.md - Overview of model optimization techniques
• quantization_deep_dive.md - Detailed explanation of quantization methods
• pruning_techniques.md - Detailed explanation of pruning techniques
• cost_estimation_for_model_optimization.md - Guide to estimating cost savings

Helper Scripts

• update_notebooks.py - Script to update all notebooks to use standardized imports
• check_imports.py - Script to check for import conflicts across notebooks

License

This project is licensed under the MIT License - see the LICENSE file for details.