Skip to content

Latest commit

 

History

History
165 lines (144 loc) · 9.68 KB

README.md

File metadata and controls

165 lines (144 loc) · 9.68 KB

a) Fair Generative Model Via Transfer Learning

(Read me is abstracted from Choi etal source code and some modifications have been made for our purpose)

1) Data setup:

(a) Download the CelebA dataset here (http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html) into the data/ directory (if elsewhere, note the path for step b). Of the download links provided, choose Align&Cropped Images and download Img/img_align_celeba/ folder, Anno/list_attr_celeba.txt, and Eval/list_eval_partition.txt to data/. (b) Download the UTKFace dataset here (https://susanqq.github.io/UTKFace/) into the data/ directory (c) Preprocess the CelebA dataset for faster training:

python preprocess_celeba.py --data_dir=/path/to/downloaded/dataset/celeba/ --out_dir=../data --partition=train

You should run this script for --partition=[train, val, test] to cache all the necessary data. The preprocessed files will then be saved in data/. (d) Preprocess the UTKFace dataset for fast training

python3 preprocess_UTKFace.py --data_dir=/path/to/downloaded/dataset/UTKFace/ --out_dir=../data/UTKFace

e)FID: For CelebA, we have provided unbiased FID statistics in the source directory. For Multi CelebA run ./BIGGAN/notebook/multi-attribute data and unbiased FID splits.ipynb . For UTKFace run ./BIGGAN/notebook/UTKFace unbiased FID splits.ipynb f)

2) For training of Resnet Classifier

  • Preprocessing files for the given LA & Training classifier

a) Train attribute classifier with CelebA (Single Attribute)

python train_attribute_clf.py celeba ./results/celeba/attr_clf

b) Train attribute classifier with CelebA (Multi Attribute)

python train_attribute_clf.py celeba ./results/celeba/multi_clf -- multi=True

c) Train attribute classifier with UTKFace

python train_attribute_clf.py UTKFace ./results/UTKFace/attr_clf

3) Generate the various Perc and bias spilits

  • Prepare data-split
  • The density ratio classifier should be trained for the appropriate bias and perc setting, which can be adjusted in the script below:
python get_density_ratios.py [celeba UTKFace] [ResNet18 CNN5 CNN3] --perc=[0.1, 0.25, 0.5, 1.0] --bias=[90_10, multi]

By employing --ncf argument, you can control the complexity of CNN classifiers. Note that the best density ratio classifier will be saved in its corresponding directory under ./data/[celeba UTKFace FairFace], you should transfer this to ./data.

2)To Train the Baseline Model

  • Use ./BIGGAN/src/Base_imp_weight for (Choi etal)
  • Use --reweight 1 for Choi et.al Importance-weights
  • Select the --perc setting e.g. {0.25,0.1,0.05,0.025}
  • Select the --bias settings e.g. {90_10,multi}, if you are testing multi you have to append --multi 1 for the model to reference the correct FD classifier
python train.py --shuffle --batch_size 128 --parallel --num_G_accumulations 1 --num_D_accumulations 1 \
--num_D_steps 4 --G_lr 2e-4 --D_lr 2e-4 --dataset CA64 --data_root ../../data --G_ortho 0.0 --G_attn 0 \
--D_attn 0 --G_init N02 --D_init N02 --ema --use_ema --ema_start 1000 --save_every 1000 --test_every 1000 --num_best_copies 2 --num_save_copies 1 \
--loss_type hinge --seed 777 --num_epochs 200 --start_eval 40 --reweight 1 --alpha 1.0 --bias 90_10 --perc 1.0 --name_suffix celeba_90_10_perc1.0_Baseline

3) FairTL and FairTL++

To Train the Pre-trained Model for fairTL++ and fairTL

  • Use ./BIGGAN/src/FairGAN++ for our proposed work
  • Use --reweight 0 for regular traing on the reference + bias data
  • Select the perc setting e.g. {0.25,0.1,0.05,0.025}
  • Select the --bias settings e.g. {90_10,multi}, if you are testing multi you have to append --multi 1 for the model to reference the correct FD classifier
python train.py --shuffle --batch_size 128 --parallel --num_G_accumulations 1 --num_D_accumulations 1 \
--num_D_steps 2 --G_lr 5e-4 --D_lr 2e-4 --dataset CA64 --data_root ../../data --G_ortho 0.0 --G_attn 0 \
--D_attn 0 --G_init N02 --D_init N02 --ema --use_ema --ema_start 1000 --save_every 1000 --test_every 1000 --num_best_copies 2 --num_save_copies 1 \
--loss_type hinge --seed 777 --num_epochs 200 --start_eval 40 --reweight 0 --alpha 1.0 --bias 90_10 --perc 1.0 --name_suffix celeba_90_10_perc1.0_pretrained

FairTL

  • Use ./BIGGAN/src/FairGAN++ for our proposed work
  • To perform fairTL we select the new "./weights/###_copy0" pre-trained weights and copy them into a new file of your own naming choice e.g., celeba_90_10_perc1.0_pretrained_Linear_Prob
  • Include --dummy 2 to train on the uniform D_ref dataset only
python train.py --shuffle --batch_size 8 --parallel --num_G_accumulations 1 --num_D_accumulations 1 \
--num_D_steps 2 --G_lr 5e-4 --D_lr 2e-4 --dataset CA64 --data_root ../../data --G_ortho 0.0 --G_attn 0 \
--D_attn 0 --G_init N02 --D_init N02 --ema --use_ema --ema_start 1000 --save_every 1000 --test_every 1000 --num_best_copies 2 --num_save_copies 1 \
--loss_type hinge --seed 777 --num_epochs 300 --start_eval 40 --reweight 0 --alpha 1.0 --bias 60_40 --perc 1.0 --name_suffix celeba_90_10_perc1.0_pretrained\
--resumePaused --load_weight copy0 --dummy 2

(FairTL++) Perform linear probing

  • Use ./BIGGAN/src/FairGAN++ for our proposed work
  • To perform linear probing we select the new "./weights/###_copy0" pre-trained weights and copy them into a new file of your own naming choice e.g., celeba_90_10_perc1.0_pretrained_Linear_Prob
  • Include --dummy 2 to train on the uniform D_ref dataset only
python train_LP.py --shuffle --batch_size 8 --parallel --num_G_accumulations 1 --num_D_accumulations 1 \
--num_D_steps 2 --G_lr 5e-4 --D_lr 2e-4 --dataset CA64 --data_root ../../data --G_ortho 0.0 --G_attn 0 \
--D_attn 0 --G_init N02 --D_init N02 --ema --use_ema --ema_start 1000 --save_every 1000 --test_every 1000 --num_best_copies 2 --num_save_copies 1 \
--loss_type hinge --seed 777 --num_epochs 200 --start_eval 40 --reweight 0 --alpha 1.0 --bias 90_10 --perc 1.0 --name_suffix celeba_90_10_perc1.0_pretrained\
--resumePaused --load_weights=copy0 --dummy 2

(FairTL++) Perform Fine Tuning

  • Use ./BIGGAN/src/FairGAN++ for our proposed work
  • Similarly select the new "./weights/###_copy0" weights and copy them into a new file e.g., celeba_90_10_perc1.0_pretrained_Linear_Prob_Fine_tuning
  • Include --dummy 2 to train on the uniform D_ref dataset only
python train_FT.py --shuffle --batch_size 8 --parallel --num_G_accumulations 1 --num_D_accumulations 1 \
--num_D_steps 2 --G_lr 5e-4 --D_lr 2e-4 --dataset CA64 --data_root ../../data --G_ortho 0.0 --G_attn 0 \
--D_attn 0 --G_init N02 --D_init N02 --ema --use_ema --ema_start 1000 --save_every 1000 --test_every 1000 --num_best_copies 2 --num_save_copies 1 \
--loss_type hinge --seed 777 --num_epochs 200 --start_eval 40 --reweight 0 --alpha 1.0 --bias 90_10 --perc 1.0 --name_suffix celeba_90_10_perc1.0_pretrained\
--resumePaused --load_weights=copy0 --dummy 2

b) Fair Generative Model Via Transfer Learning From StyleGANv2

1) Preprocessing Raw data into the respective SA splits

python sort_data.py

2) Preprocessing the data into the zip file StyleGAN2 code requires

  • zip the file for training
  • Select the source data i.e., ./StyleGANv2/stylegan2-ada-pytorch-main/Prepare_data/training/Gender/GenderTrainSamples_0.025/
  • Select the output dataset name (this can be anything you want)
python dataset_tool.py --source=*source of Dref* --dest=~/datasets/*dataset_name*.zip

2) Train attribute classifier

  • Use the directiory ./StyleGANv2/stylegan2-ada-pytorch-main//metrics/FD
  • Then copy the data from 3)Gender_FFHQ_Clf_Training_data.pt and Gender_FFHQ_Clf_Training_labels.pt into ./StyleGANv2/stylegan2-ada-pytorch-main//metrics/FD/data/ and post process the data to train/Test/split with:
python Prepare_training_data.py

then train the classifier with (Fill sensitive attribute(SA) name):

python train_attribute_clf.py -attribute *SA*

3)Resume sensitive adaptation with styleGANv2

  • Training the pre-trained styleGANv2 on the new Dref data with linear probing
  • Fill training datafile with output file from 2)
  • Fill network pickle file
python train_2D_LP.py --outdir=~/training-runs --data=~/mydataset/*training datafile* --gpus=1 --resume=https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan2/versions/1/files/stylegan2-ffhq-256x256.pkl --aug=noaug --kimg 200

followed by fine-tuining:

  • network pickle file is the output of the LP file
python train_2D.py --outdir=~/training-runs --data=~/mydataset/*training datafile* --gpus=1 --resume=*network pickle file* --aug=noaug --kimg 800

4)Measuring Fairness and FID

  • Select FID ref file from 2)FID reference files e.g., .`/training/Gender/GenderFIDSamples/
  • Select saved network from the training file
  • Select SA to evaluate
python cal_FD --attribute *SA* --network *saved network*

python calc_metrics.py --metrics=fid50k_full --data=~/datasets/*FID ref file* --mirror=1 --network *saved network*