MetaFBP: Learning to Learn High-Order Predictor for Personalized Facial Beauty Prediction

• 🔔This is the official (Pytorch) implementation for the paper "MetaFBP: Learning to Learn High-Order Predictor for Personalized Facial Beauty Prediction", ACM MM 2023.
• 🛖This repository is built on the dragen1860/MAML-Pytorch. You can also view the project for details: https://github.com/dragen1860/MAML-Pytorch

🛠️Setup

Runtime

The main python libraries we use:

• Python 3.8
• torch 1.8.1
• numpy 1.19.2

Datasets

Please create a directory named datasets in current directory, then download these following datasets and unzip into datasets:

Download link: GoogleDrive or Quark

You can also change the root directory of datasets by modifying the default value of the argument --data-root in train_fea.py[L34], train.py[L40], and test.py[L59]:

# train_fea.py, Line 34parser.add_argument('--data-root', type=str, default='./datasets')
# train.py, Line 40parser.add_argument('--data-root', type=str, default='./datasets')
# test.py, Line 59parser.add_argument('--data-root', type=str, default='./datasets')

🎢Run

After finishing above steps, your directory structure of code may like this:

MetaFBP/
|–– data/
|–– dataset/
|–– FBP5500/
|–– FBPSCUT/
|–– US10K/
|–– model/
|–– util/
README.md
test.py
test_fea.py
train.py
train.sh
train_fea.py
train_fea.sh

1. First of all, please train the universal feature extractor for each dataset:

   bash train_fea.sh PFBP-SCUT5500
   bash train_fea.sh PFBP-SCUT500
   bash train_fea.sh PFBP-US10K

   Usage of train_fea.sh:

   bash train_fea.sh {arg1=dataset}
   

   • dataset specifies which dataset to train on, available ones are: PFBP-SCUT5500,PFBP-SCUT500, PFBP-US10K

2. Once the universal feature extractor is ready, you can run the experiments of PFBP task. For example, the following cmd runs the experiment of MetaFBP-R on PFBP-SCUT5500 benchmark with 5-way K-shot regression:

   bash train.sh MetaFBP-R PFBP-SCUT5500

   Usage of train.sh:

   bash train.sh {arg1=model} {arg2=dataset}
   

   • model specifies which model to use, available ones are: Base-MAML,MetaFBP-R, MetaFBP-T
   • dataset specifies which dataset to train and test on, available ones are: PFBP-SCUT5500,PFBP-SCUT500, PFBP-US10K

📌Citation

If you would like to cite our work, the following bibtex code may be helpful:

@inproceedings{lin2023metafbp,
title={MetaFBP: Learning to Learn High-Order Predictor for Personalized Facial Beauty Prediction},
author={Lin, Luojun and Shen, Zhifeng and Yin, Jia-Li and Liu, Qipeng and Yu, Yuanlong and Chen, Weijie},
booktitle={Proceedings of the 31st ACM International Conference on Multimedia},
year={2023},
}

🔗Acknowledgements

• Our code is built on dragen1860/MAML-Pytorch - https://github.com/dragen1860/MAML-Pytorch

⚖️License

This source code is released under the MIT license. View it here