- To evaluate the repaired model, execute
eval_repaired.py
by running:
python3 eval_repaired.py <clean validation data directory> <test data directory> <model directory>
.
E.g., python3 eval_repaired.py data/clean_validation_data.h5 data/sunglasses_poisoned_data.h5 models/sunglasses_bd_net.h5
. to test on the sunglasses poisoned data
python3 eval_repaired.py data/clean_validation_data.h5 data/clean_validation_data.h5 models/sunglasses_bd_net.h5
. to test on the clean validation data
This script would create the repaired model and show the accuracy for the badnet as well as the repaired model The script takes around 5-10 mins per run. (shows progress after every 1000 images processes)