Efficient TTA with Cache-based Dynamic Adapter (TDA)

Efficient Test-Time Adaptation of Vision-Language Models
Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, Eric Xing

Course-project for "Trends and Applications in Computer Vision" of prof. M. Mancini and G. Boato.

Here you can find our final presentation for the results.

Here you can find the report about the related works we studied with our presentation.

by Juan Camacho Mohedano, Andrea De Carlo, Samuele Bolotta

Please refer to the official README of the original project for the configuration of the original code.

Our Contributions

What we did:

• Benchmark on different datasets, both OOD and CD with failure cases on CIFAR-10-C (non-iid data stream)

• We evaluated how the performance changed w.r.t. changing hyperparameters and the orders of data presented considering budget-aware constraints

• We tried to mitigate the issues adding a Waiting List to the model, which improved performance on ImageNet but didn’t help on more challenging dataset like as CIFAR10-C

You can find better details in the final presentation

file added:

• tda_cd_benchmark.ipynb
• tda_runner_experiments.py
• tda_runner_with_waiting.py