CLEAM_code

Real_GAN (GAN)

1)Training attribute classifier

a) Preparing the dataset

Download the dataset from https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html Pepare a pre-processed version with the following source code. In Path ./Real_GAN/CLEAM/attributeClassifier: Run the following with data_dir changed to the location of the images and label_dir as the label directory

python pre-process.py

In Path ./Real_GAN/CLEAM/attributeClassifier: Split data into even train/test/val. Ammend class_idx to the respective SA.

python data_split.py

b) Train Attribute Classifier

In Path ./Real_GAN/CLEAM/attributeClassifier/src_ResNet-18: train attribute classifier, change class_idx to the respective SA

python train_attribute_clf --class_idx 20

In Path ./Real_GAN/CLEAM/attributeClassifier/src_ResNet-18: Validate attribute classifier, change class_idx to the respective SA

python validate_acc.py --class_idx 20

2) Preparing data for real GAN

Download dataset from annonymous link https://drive.google.com/drive/folders/1ENslNLyK6EEG2qj5YLZ3Qu3rFijJWEqB?usp=sharing and copy them into ./Real_GAN/GeneratedData/datasetName/samples a)In Path ./Real_GAN/preprocess: Preprocess dataset to .npz format, edit dataDir for new dataset

python numpyData.py

3) Run CLEAM

a) In Path ./Real_GAN/CLEAM

Run for CLEAM approximation. Please edit attributeDict dictionary with the validated classifier's accuracy

python fairness_classifier_mturk_celebAHQ_StyleGAN_Resnet18.py

Real_GAN (DGM)

1) Download CelebA-HQ dataset

Download dataset and 'CelebAMask-HQ-attribute-anno' from https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

2) Evaluate CLIP's accuracy

a)In path ./Real_GAN/CLEAM/CLIP

Change the labelpath and dataPath in the python script to your paths

python celebAHQ_realsamples_labeller_measure_alpha.py

3) Evluate the DGM generated samples

a)In path ./Real_GAN/CLEAM/CLIP

update the --acc to the measured accuracy in (2) update the --dataPath to where the data is located update the --SA to the respective senstive attribute {Gender,Smiling}

python CLEAM_CLIP.py