This repository is used for finding and predicting from an image one or multiple human's gender and age(confidence scores provided for both age and gender). 94% acc for gender and MAE of 4.2 for age can be achieved after just 32 epochs of training.

• python3, pytorch
• pip3 install --upgrade opencv-python, dlib, imutils, skimage
• Download FaceAligner and save shape_predictor_68_face_landmarks.dat in models/
• Download FaceDetector and save mmod_human_face_detector.dat in models/

1. train the model using python train.py, weight will be stored in models/
2. or download pretrained weight from url[pending]
3. put your test image in pics/val/
4. run python evaluate.py

call eval_live() function in evaluate.py

1. Using cleaned IMDB-WIKI dataset[1] for training (IMDB-WIKI dataset contains 50%+ mislabeled images[2]).
2. Using FG-NET dataset[3] for testing.
3. Train a model based on ResNet-18,
   • the output is 2 neuron represents probs of male&female plus 100 neurons represents probs of being age 0-99.
   • auto detect if use GPU or even multiple GPUs for training.
   • auto reduce learning rate when we have no loss reduce on val dataset for >N epochs.
   • auto freeze CNN laters and train only last FCN layers when first epoch.
   • auto load and save weights, log training loss and metadatas after each epoch.
   • more detains can be found on src file train.py and configuration file config.ini

1. detect and align faces using dlib.
2. predict age, gender and confidence scores(probability of each gender and variance of age).

[1] Rothe R, Timofte R, Van Gool L. Deep Expectation of Real and Apparent Age from a Single Image Without Facial Landmarks [J/OL]. International Journal of Computer Vision. 126 (2). 2018, Apr: 144–157.

[2] Antipov G, Baccouche M, Berrani S-A et al. Effective training of convolutional neural networks for face-based gender and age prediction [J/OL]. Pattern Recognition. 72. 2017, December: 15–26.

[3] Panis G, Lanitis A, Tsapatsoulis N et al. Overview of research on facial ageing using the FG-NET ageing database [J]. IET Biometrics. 5 (2). 2016: 37–46.

Feel free to mail [email protected] for any pretrained weight or quetsions and bug report about this repo :)