CUSP

Official implementation for paper Custom Structure Preservation in Face Aging (Accepted on ECCV '22)

Preparation

DEX model

The procedure to obtain the weights needed for the DEX age classifier are described in HRFAE Github's README.md.

Docker image

For improved efficiency it is recommended to use a Docker image.

docker build -t i2idocker .

Data

FFHQ Restricted Range

FFHQ-RR dataset is built following its corresponding instructions available on HRFAE: High Resolution Face Age Editing.

FFHQ Lifespan

FFHQ-LS dataset is downloaded from FFHQ Aging Dataset Github.

Models

Available pretrined models

Two models have been released depending on the dataset used:

• Trained on FFHQ-RR dataset: Google Drive.
• Trained on FFHQ-LS dataset: Google Drive.

Train a new model

FFHQ-RR

DOCKER_WORKDIR=/workdir
USER_DIR=/<USER_PATH>/
GPU_ID=0
RELATIVE_DATASET_PATH=../datasets/ffhq_augm
docker run --gpus all --rm --user "$(id -u):$(id -g)" --workdir $DOCKER_WORKDIR -v $USER_DIR:$DOCKER_WORKDIR -e HOME=$DOCKER_WORKDIR -w $DOCKER_WORKDIR/cusp-pytorch i2idocker:latest python train.py --outdir ./training-runs_ffhqaug --gpus $GPU_ID --cfg 224 --dataset ffhq_aug --age_np $RELATIVE_DATASET_PATH/ffhq_aug_tr.npy --age_np_test $RELATIVE_DATASET_PATH/ffhq_aug_ts.npy --data $RELATIVE_DATASET_PATH/images --classifier_path ./dex_imdb_wiki.caffemodel.pt --downsamples 5 --bias False --class_w 0.1 --skip_layers 2 --skip_kind linear --age_margin 0 --rgb_attention False --rgb_reg none --soft_margin True --act_reg l2 --learn_mask none --mixing_prob 0.0 --disc_class True --cmap_kind number --fake_rec false --skip_grad_blur gb --blur_skip true --blur_msk random --cycle_w 10 --bottleneck_class false --style_enc true --class_kind max

The dataset is described through two numpy files for train and test, respectively ffhq_aug_tr.npy and ffhq_aug_ts.npy. It is shaped (num_images, 2), the first column is the relative filename path from --data and the second column is DEX's predicted age.

array([['30032.png', '60'],
       ['14936.png', '42'],
       ['47014.png', '66'],
       ...,
       ['37665.png', '59'],
       ['31158.png', '30'],
       ['04330.png', '41']], dtype='<U9')

FFHQ-LS

DOCKER_WORKDIR=/workdir
USER_DIR=/<USER_PATH>/
GPU_ID=0
RELATIVE_DATASET_CSV_PATH=../datasets/FFHQaging.csv
RELATIVE_DATASET_IMAGES_PATH=../datasets/FFHQaging
RESNET_FFHQ_LS_PATH=./trained_models/resnet_ffhq_ls.pt
docker run --gpus all --rm --user "$(id -u):$(id -g)" --workdir $DOCKER_WORKDIR -v $USER_DIR:$DOCKER_WORKDIR -e HOME=$DOCKER_WORKDIR -w $DOCKER_WORKDIR/cusp-pytorch i2idocker:latest python train.py --outdir ./training-runs-ffhqlat --gpus 3 --cfg 256 --dataset ffhq_lat --csv $RELATIVE_DATASET_CSV_PATH --data $RELATIVE_DATASET_IMAGES_PATH --classifier_path $RESNET_FFHQ_LS_PATH --downsamples 5 --bias False --class_w 0.1 --skip_layers 2 --skip_kind linear --act_reg l2 --skip_grad_blur gb --learn_mask none --mixing_prob 0.0 --disc_class true --fake_rec false --cycle_w 10 --blur_skip true --blur_msk random --cmap_kind number

The dataset is described with a single csv file, FFHQaging.csv, similar to the following, where the column is_train describes if the image is used for training or testing.

fname	agebin_0-2	agebin_3-6	agebin_7-9	agebin_10-14	agebin_15-19	agebin_20-29	agebin_30-39	agebin_40-49	agebin_50-69	agebin_70-120	is_train
00000.png	1	0	0	0	0	0	0	0	0	0	TRUE
00001.png	0	0	0	0	0	0	1	0	0	0	TRUE
...	...	...	...	...	...	...	...	...	...	...	...
69999.png	0	0	0	0	0	0	0	0	1	0	FALSE

(Google Colab) Load a pretrained model

You can run the sample notebook in Google Colab.

Citation

If used in your research, please cite the following paper:

@inproceedings{gomez2022custom,
  title={Custom structure preservation in face aging},
  author={Gomez-Trenado, Guillermo and Lathuili{\`e}re, St{\'e}phane and Mesejo, Pablo and Cord{\'o}n, {\'O}scar},
  booktitle={European Conference on Computer Vision},
  pages={565--580},
  year={2022},
  organization={Springer}
}

License

The base code was forked from StyleGAN2-ADA-pytorch