A Reproducible Benchmark of Visual Attention Mechanisms in CNNs

This project provides a professional, end-to-end Python pipeline to investigate the impact of modern visual attention mechanisms (Squeeze-and-Excitation, CBAM) against a standard ResNet-style baseline. The benchmark is run across four diverse datasets to test for generalization.

The entire system is built with MLOps best practices in mind, featuring modular code, command-line configuration, and comprehensive experiment tracking with MLflow.

Key Findings: The Impact of Attention

The experiments reveal that attention mechanisms provide a significant, task-dependent performance boost, especially on complex and fine-grained datasets.

Dataset	Baseline (None)	SE (Squeeze-and-Excitation)	CBAM (Convolutional Block)
CIFAR10	86.97%	87.01% (+0.04%)	85.89% (-1.08%)
CIFAR100	57.26%	59.28% (+2.02%)	58.07% (+0.81%)
Oxford102Flowers	28.13%	37.53% (+9.40%)	38.74% (+10.61%)
EuroSAT	97.41%	97.61% (+0.20%)	97.37% (-0.04%)

Tech Stack & Skills Demonstrated

• Frameworks: Python, PyTorch, Torchvision
• MLOps: MLflow (for experiment tracking), argparse (for configuration management)
• Engineering Practices: Modular & Reusable Code Structure, GPU Memory Management, Object-Oriented Programming (OOP), Reproducible Research Methodology.
• Tools: Git, Bash, Jupyter (for initial exploration)

How to Reproduce the Results

This project is structured for easy reproduction on any machine with a CUDA-enabled GPU.

1. Clone the repository:

   git clone https://github.com/chillguycode/visual_attention_benchmarking.git
   cd visual_attention_benchmarking

2. Install dependencies:

   pip install -r requirements.txt

3. Run a training experiment: The run_experiment.py script is the main entry point. All parameters are configurable via the command line.

   # Example: Run the SE model on the CIFAR100 dataset for 30 epochs
   python run_experiment.py --dataset CIFAR100 --attention_type se --epochs 30

   # Example: Run the baseline model on EuroSAT
   python run_experiment.py --dataset EuroSAT --attention_type none

4. View and Compare All Experiment Results: All parameters and metrics are automatically logged with MLflow. To launch the tracking dashboard and compare your runs, execute:

   mlflow ui

   Then navigate to http://127.0.0.1:5000 in your browser.