Efficient Test-Time Adaptation with Cache-based Dynamic Adapter

📄 Official Paper:
Efficient Test-Time Adaptation of Vision-Language Models
🧠 Authors: Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, Eric Xing

🎓 This project was developed as part of the “Trends and Applications in Computer Vision” course taught by Prof. M. Mancini and G. Boato.

🔍 Overview

This repository extends the official implementation of TDA (Test-Time Adaptation) by exploring its efficiency, robustness, and flexibility in different real-world scenarios.

📂 Resources:

• 📊 Final Presentation
• 📚 Related Works Report
• 🖥️ Related Works Slides

👥 Project by:
Juan Camacho Mohedano •
Andrea De Carlo •
Samuele Bolotta

For setup and base implementation details, refer to the original README_official.md

🚀 Our Contributions

📌 1. Benchmarking on Diverse Datasets

We evaluated TDA on:

• Class Distribution Shifts (CD)
• Out-of-Distribution (OOD) test cases

Example:
CIFAR-10-C (non-iid stream) results:

📄 Code: tda_cd_benchmark.ipynb

⚙️ 2. Hyperparameter Sensitivity under Budget Constraints

Performance analysis across:

• Adaptation budgets
• Stream orderings

📄 Code: tda_runner_experiments.py

🧩 3. Waiting List Strategy

Added Waiting List to improve robustness:

• Helped on ImageNet
• Less effective on CIFAR10-C

📄 Code: tda_runner_with_waiting.py

📁 Files Added

📌 Notes

• Experiments are reproducible and tested under realistic compute constraints.
• Visualizations and explanations are in the Final Presentation.

📬 Contact

For inquiries, feel free to reach out to any of us on LinkedIn.

⭐ If you found this useful, consider checking out the official paper and citing it!