Tent Example: Image Corruptions
This example compares a baseline without adaptation (base, or source), test-time normalization that updates feature statistics during testing (norm), and our method for entropy minimization during testing (tent).
Each experiment measures accuracy (%) on corrupted data. Clean accuracy is 94.78%.
- dataset: CIFAR-10-C (https://github.com/hendrycks/robustness/), with 15 corruption types and 5 levels
- model: WRN-28-10, the default for RobustBench (https://github.com/RobustBench/robustbench)
Created on March 22|Last edited on April 28
Comment
python cifar10c.py --cfg cfgs/source.yamlpython cifar10c.py --cfg cfgs/norm.yamlpython cifar10c.py --cfg cfgs/tent.yaml
results
Run set
3
Run set
3
Run set
3
Run set
3
Run set
3
Add a comment