Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time
April 30, 2018 19:55
April 30, 2018 20:48
June 21, 2018 17:40
April 30, 2018 20:19
November 22, 2019 15:42
May 1, 2018 01:24


[IJCAI18] SSR-Net: A Compact Soft Stagewise Regression Network for Age Estimation

Code Author: Tsun-Yi Yang

Last update: 2019/09/19 (Renew the morph2 dataset link)

Real-time webcam demo



Paper authors

Tsun-Yi Yang, Yi-Husan Huang, Yen-Yu Lin, Pi-Cheng Hsiu, and Yung-Yu Chuang


This paper presents a novel CNN model called Soft Stagewise Regression Network (SSR-Net) for age estimation from a single image with a compact model size. Inspired by DEX, we address age estimation by performing multi-class classification and then turning classification results into regression by calculating the expected values. SSR-Net takes a coarse-to-fine strategy and performs multi-class classification with multiple stages. Each stage is only responsible for refining the decision of the previous stage. Thus, each stage performs a task with few classes and requires few neurons, greatly reducing the model size. For addressing the quantization issue introduced by grouping ages into classes, SSR-Net assigns a dynamic range to each age class by allowing it to be shifted and scaled according to the input face image. Both the multi-stage strategy and the dynamic range are incorporated into the formulation of soft stagewise regression. A novel network architecture is proposed for carrying out soft stagewise regression. The resultant SSR-Net model is very compact and takes only 0.32 MB. Despite of its compact size, SSR-Net’s performance approaches those of the state-of-the-art methods whose model sizes are more than 1500x larger.


  • Keras
  • Tensorflow
  • GTX-1080Ti
  • Ubuntu


pip install mtcnn
conda install -c conda-forge moviepy
conda install -c cogsci pygame
conda install -c conda-forge requests
conda install -c conda-forge pytables


There are three different section of this project.

  1. Data pre-processing
  2. Training and testing
  3. Video demo section We will go through the details in the following sections.

This repository is for IMDB, WIKI, and Morph2 datasets.

1. Data pre-processing

cd ./data
python --db imdb --output imdb.npz
python --db wiki --output wiki.npz
python --output morph_db_align.npz

2. Training and testing

The experiments are done by randomly choosing 80% of the dataset as training and 20% of the dataset as validation (or testing). The details of the setting in each dataset is in the paper.

For MobileNet and DenseNet:

cd ./training_and_testing

For SSR-Net:

cd ./training_and_testing
  • Note that we provide several different hyper-parameters combination in this code. If you only want a single hyper-parameter set, please alter the command inside "".

Plot the results: For example, after the training of IMDB dataset, you want to plot the curve and the results. Copy "", "", and "" into "./imdb_models". The following command should plot the results of the training process.


3. Video demo section

Pure CPU demo command:

cd ./demo

# Or you can use

KERAS_BACKEND=tensorflow CUDA_VISIBLE_DEVICES='' python TGOP.mp4 '3'
  • Note: You may choose different pre-trained models. However, the morph2 dataset is under a well controlled environment and it is much more smaller than IMDB and WIKI, the pre-trained models from morph2 may perform ly on the in-the-wild images. Therefore, IMDB or WIKI pre-trained models are recommended for in-the-wild images or video demo.

  • We use dlib detection and face alignment in the previous experimental section since the face data is well organized. However, dlib cannot provide satisfactory face detection for in-the-wild video. Therefore we use mtcnn as the detection process in the demo section.

Real-time webcam demo

Considering the face detection process (MTCNN or Dlib) is not fast enough for real-time demo. We show a real-time webcam version by using lbp face detector.

cd ./demo
  • Note that the covered region of face detection is different when you use MTCNN, Dlib, or LBP. You should choose similar size between the inference and the training.
  • Also, the pre-trained models are mainly for the evaluation of the datasets. They are not really for the real-world images. You should always retrain the model by your own dataset. In webcam demo, we found that morph2 pre-trained model actually perform better than wiki pre-trained model. The discussion will be included in our future work.
  • If you are Asian, you might want to use the megaage_asian pre-trained model.
  • The Morph2 pre-trained model is good for webcam but the gender model is overfitted and not practical.

4. Extension

Training the gender model

We can reformulate binary classification problem into regression problem, and SSR-Net can be used to predict the confidence. For example, we provide gender regression and demo in the code for the extension.

Training the gender network:

cd ./training_and_testing

Note that the score can be between [0,1] and the 'V' inside SSR-Net can be changed into 1 for general propose regression.

Third Party Implementation