Defactify 4 AAAI 2025 Workshop

Train model

When training model with prompt --noise_sigma x, means add noise with sigma from 0 to x randomly.

python classifier.py --features_selected rgb
python classifier.py --features_selected rgb --noise_sigma 0.2

Generate validation result

python classifier.py --features_selected rgb --evaluate

Use Grad-cam

python classifier.py --features_selected rgb --grad_cam 

Inference

python classifier.py --features_selected rgb --inference
python classifier.py --features_selected rgb --inference --labels_file ../data/test/captions.xlsx

feature_extraction with Robustness test

python feature_extraction.py --data_path ../data/val --results_path ../data/val_compression --error --frequency --flat_structure --compression_quality 80
python feature_extraction.py --data_path ../data/val --results_path ../data/val_crop --error --frequency --flat_structure --crop_factor 0.8
python feature_extraction.py --data_path ../data/val --results_path ../data/val_blur --error --frequency --flat_structure --blur_sigma 2
python feature_extraction.py --data_path ../data/val --results_path ../data/val_noise --error --frequency --flat_structure --noise_sigma 0.1