Skip to content

Keras implementation of a CNN network for age and gender estimation

License

Notifications You must be signed in to change notification settings

nanfengpo/age-gender-estimation

 
 

Repository files navigation

Age and Gender Estimation

This is a Keras implementation of a CNN for estimating age and gender from a face image [1, 2]. In training, the IMDB-WIKI dataset is used.

Dependencies

  • Python3.5+
  • Keras2.0+
  • scipy, numpy, Pandas, tqdm, tables, h5py
  • dlib (for demo)
  • OpenCV3

Tested on:

  • Ubuntu 16.04, Python 3.5.2, Keras 2.0.3, Tensorflow(-gpu) 1.0.1, Theano 0.9.0, CUDA 8.0, cuDNN 5.0
    • CPU: i7-7700 3.60GHz, GPU: GeForce GTX1080
  • macOS Sierra, Python 3.6.0, Keras 2.0.2, Tensorflow 1.0.0, Theano 0.9.0

Usage

Use pretrained model

Download pretrained model weights for TensorFlow backend:

mkdir -p pretrained_models
wget -P pretrained_models https://www.dropbox.com/s/rf8hgoev8uqjv3z/weights.18-4.06.hdf5

Run demo script (requires web cam)

python3 demo.py

Model weights for Theano backend is also available from here.

Train a model using the IMDB-WIKI dataset

Download the dataset

The dataset is downloaded and extracted to the data directory.

./download.sh

Create training data

Filter out noise data and serialize images and labels for training into .mat file. Please check check_dataset.ipynb for the details of the dataset.

python3 create_db.py --output data/imdb_db.mat --db imdb --img_size 64
usage: create_db.py [-h] --output OUTPUT [--db DB] [--img_size IMG_SIZE] [--min_score MIN_SCORE]

This script cleans-up noisy labels and creates database for training.

optional arguments:
  -h, --help                 show this help message and exit
  --output OUTPUT, -o OUTPUT path to output database mat file (default: None)
  --db DB                    dataset; wiki or imdb (default: wiki)
  --img_size IMG_SIZE        output image size (default: 32)
  --min_score MIN_SCORE      minimum face_score (default: 1.0)

Train network

Train the network using the training data created above.

python3 train.py --input data/imdb_db.mat

Trained weight files are stored as checkpoints/weights.*.hdf5 for each epoch if the validation loss becomes minimum over previous epochs.

usage: train.py [-h] --input INPUT [--batch_size BATCH_SIZE]
                [--nb_epochs NB_EPOCHS] [--depth DEPTH] [--width WIDTH]
                [--validation_split VALIDATION_SPLIT] [--aug]

This script trains the CNN model for age and gender estimation.

optional arguments:
  -h, --help            show this help message and exit
  --input INPUT, -i INPUT
                        path to input database mat file (default: None)
  --batch_size BATCH_SIZE
                        batch size (default: 32)
  --nb_epochs NB_EPOCHS
                        number of epochs (default: 30)
  --depth DEPTH         depth of network (should be 10, 16, 22, 28, ...)
                        (default: 16)
  --width WIDTH         width of network (default: 8)
  --validation_split VALIDATION_SPLIT
                        validation split ratio (default: 0.1)
  --aug                 use data augmentation if set true (default: False)

Train network with recent data augmentation methods

Recent data augmentation methods, mixup [3] and Random Erasing [4], can be used with standard data augmentation by --aug option in training:

python3 train.py --input data/imdb_db.mat --aug

Please refer to this repository for implementation details.

I confirmed that data augmentation enables us to avoid overfitting and improves validation loss.

Use the trained network

python3 demo.py
usage: demo.py [-h] [--weight_file WEIGHT_FILE] [--depth DEPTH] [--width WIDTH]

This script detects faces from web cam input, and estimates age and gender for
the detected faces.

optional arguments:
  -h, --help                show this help message and exit
  --weight_file WEIGHT_FILE path to weight file (e.g. weights.18-4.06.hdf5) (default: None)
  --depth DEPTH             depth of network (default: 16)
  --width WIDTH             width of network (default: 8)

Please use the best model among checkpoints/weights.*.hdf5 for WEIGHT_FILE if you use your own trained models.

Plot training curves from history file

python3 plot_history.py --input models/history_16_8.h5 
Results without data augmentation

Results with data augmentation

The best val_loss was improved from 3.969 to 3.731:

  • Without data augmentation: 3.969
  • With standard data augmentation: 3.799
  • With mixup and random erasing: 3.731

We can see that, with data augmentation, overfitting did not occur even at very small learning rates (epoch > 15).

Network architecture

In the original paper [1, 2], the pretrained VGG network is adopted. Here the Wide Residual Network (WideResNet) is trained from scratch. I modified the @asmith26's implementation of the WideResNet; two classification layers (for age and gender estimation) are added on the top of the WideResNet.

Note that while age and gender are independently estimated by different two CNNs in [1, 2], in my implementation, they are simultaneously estimated using a single CNN.

Results

Trained on imdb, tested on wiki.

References

[1] R. Rothe, R. Timofte, and L. V. Gool, "DEX: Deep EXpectation of apparent age from a single image," ICCV, 2015.

[2] R. Rothe, R. Timofte, and L. V. Gool, "Deep expectation of real and apparent age from a single image without facial landmarks," IJCV, 2016.

[3] H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz, "mixup: Beyond Empirical Risk Minimization," in arXiv:1710.09412, 2017.

[4] Z. Zhong, L. Zheng, G. Kang, S. Li, and Y. Yang, "Random Erasing Data Augmentation," in arXiv:1708.04896, 2017.

About

Keras implementation of a CNN network for age and gender estimation

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 94.7%
  • Python 5.2%
  • Shell 0.1%