Official repo for Lightweight Facial Attractiveness Prediction Using Dual Label Distribution (Submitted to TCDS).

The code is tested on a CentOS server with a NVIDIA V100 GPU. Run the following commands to install the dependencies.

conda create --name 2D_FAP python=3.8.8
conda activate 2D_FAP
conda install pytorch=1.8.1 torchvision=0.9.1 torchaudio=0.8.1 cudatoolkit=10.1 -c pytorch
pip install pillow thop torchsummary

Extract the archive files of datasets into the /data folder. The download links are provided as follows.

• Google Drive: https://drive.google.com/file/d/1ObjJ8RwYBVI71Eyo_wsM1dWzI0JbhotA/view
• Baidu Netdisk: https://pan.baidu.com/s/1YWalTRX4YwV7XGMcCXJ5Dg?pwd=157g (Pass: 157g)

• Google Drive: https://drive.google.com/file/d/1KB9F2eASZgKrnL3FUaR_59XUKjns_Aim/view
• Baidu Netdisk: https://pan.baidu.com/s/13sGf7HypD_z_ZwWdPIfdug?pwd=dpmj (Pass: dpmj)

Place the pretrained models with .pth format under the /models folder.

• Google Drive: https://drive.google.com/drive/folders/1-wStPO8K3fN9Xijp4gYbG_Sl64K5AFwE
• Baidu Netdisk: https://pan.baidu.com/s/1WPeVsRph3vJPNVIxV8EpjQ?pwd=wgnv (Pass:wgnv)

Notice that the best settings have been configured for both datasets.

# SCUT-FBP5500
python 2D_FAP.py --dataset 5500

# SCUT-FBP
python 2D_FAP.py --dataset 500

loss1, loss2, and loss3 in the code stand for $L_{ad}$, $L_{rd}$, and $L_{score}$ in the paper, respectively.

• --fold: int

  • Fold for cross validation on SCUT-FBP5500. Default 1.
  • Available options: 1, 2, 3, 4, 5

• --alpha: int

  • Weight for $L_{ad}$. Default 1.

• --beta: int

  • Weight for $L_{rd}$. Default 1.

• --gamma: int

  • Weight for $L_{score}$. Default 1.

• --dataset: int

  • Adopted dataset. Default 5500. (SCUT-FBP5500)
  • Available options: 5500(SCUT-FBP5500), 500(SCUT-FBP).

• --aligned: bool

  • Aligned images, only available on SCUT-FBP5500. Default True.

• --MTCNN: bool

  • MTCNN-processed SCUT-FBP5500 images. Default False.

• --sample: str

  • The adopted distribution in the attractiveness distribution. Default 'L'. (Laplace Distribution)
  • Available options: 'L' (Laplace Distribution), 'G' (Gaussian Distribution).

• --loss1: str

  • Different forms of $L_{ad}$. Default 'ED'. (Euclidean Distance)
  • Available options: 'L1' (L1 Distance), 'ED' (Euclidean Distance), 'KL' (Kullback–Leibler divergence).

• --loss1_option: str

  • Take the sum/average over a mini-batch as $L_{ad}$ for the batch. Default 'mean'.
  • Available options: 'sum', 'mean'.

• --loss3: str

  • Different forms of $L_{score}$. Default '3'.

  • Available options

    • '1': $L_{score}= \ln {[\frac12(\exp{(\hat y^{(i)}-y^{(i)})}+\exp{(y^{(i)}-\hat y^{(i)})})]}$

    • '2': $L_{score}= \ln{(1+|\hat y^{(i)}-y^{(i)}|)}$

    • '3': $L_{score}=\exp{(|\hat y^{(i)}-y^{(i)}|)-1}$

    • '4': $L_{score}= \ln{(|\hat y^{(i)}-y^{(i)}|+\sqrt{1+|\hat y^{(i)}-y^{(i)}|^2})}$

• --loss3_option: str

  • Take the sum/average over a mini-batch as $L_{score}$ for the batch. Default 'sum'.
  • Available options: 'sum', 'mean'.

• --losses: str

  • The losses used in training. Default '123'.

• --lr: float

  • Initial learning rate. Default 0.001.

• --network: str

  • Adopted network architecture. Default 'mobilenet'. (MobileNetV2)
  • Available options
    • 'mobilenet': ImageNet-pretrained MobileNetV2.
    • 'mobilenet_m': ImageNet-pretrained MobileNetV2 without the pretrained parameters on the last fc layer.
    • 'mobilenetv3_large': ImageNet-pretrained MobileNetV3_large.
    • 'mobilenetv3_small': ImageNet-pretrained MobileNetV3_small.
    • 'resnet18': ImageNet-pretrained ResNet-18.
    • 'resnet50': ImageNet-pretrained ResNet-50.
    • 'vgg16': ImageNet-pretrained VGG16.
    • 'vgg19': ImageNet-pretrained VGG19.
    • 'attnet_m': Modified AttNet [1] that adapts our task.
    • 'hmt': Modified HMTNet [2] that adapts our task.

• --local_rank: int

  • Local rank for DistributedDataParallel. Default -1.

• --batch: int

  • Batch size. Default 256.

• --count: bool

  • Model parameter counting. Default False.

• --interval: float

  • Interval length, namely $\Delta l$ in the paper. Default 0.1.

• --min_score: int

  • Minimum attractiveness score of the dataset. Default 1.

• --max_score: int

  • Maximum attractiveness score of the dataset. Default 5.

• --device: str

  • GPU Device ID. Default 0.

• SCUT-FBP5500: https://github.com/HCIILAB/SCUT-FBP5500-Database-Release
• SCUT-FBP: http://www.hcii-lab.net/data/SCUT-FBP/EN/introduce.html

[1] Gao, Bin-Bin, et al. "Learning expectation of label distribution for facial age and attractiveness estimation." arXiv preprint arXiv:2007.01771 (2020).

[2] Xu, Lu, Heng Fan, and Jinhai Xiang. "Hierarchical multi-task network for race, gender and facial attractiveness recognition." IEEE International conference on image processing (ICIP). IEEE, 2019.