Apparent Age Gender Classification

Course

Machine Learning @ NCTU EE (grad) 2016 Sep. - 2017 June

About the Project (from TA's document)

In this project, you are asked to classify the input photos into 4 age groups and 2 genders. You are free to use any methods, tools, and language to finish this work.
If you want to use any work that is not built by you, be sure to specify it in your report. Also, just repeating other people’s work won’t help you get high score in this subject, be sure to include your own idea in this work.

A. Input:
The input will be an RGB photo or Gray level photo with different size. The photos above are from the class child, young, adult, and elder, respectively.
You are only allowed to use the photos in this dataset. (Data augmentation is permitted) We may randomly pick some team’ s work to retrain it to check if the result is far differed from your submission.

B. Target
For each photos, please classify it into one of the following classes. You can predict age and gender jointly or separately, but the final result need to be one of the classes above.
Class is defined as follows:
0: Male Child, 4: Female Child
1: Male Young, 5: Female Young
2: Male Adult, 6: Female Adult
3: Male Elder, 7: Female Elder

Contributor

Daniel You, Vivian Chung

Machine Learning Library Involved

Tensorflow, Dlib, scikit-learn

Reference

Data Conversion
http://blog.csdn.net/u012759136/article/details/52232266

Tensorflow
https://www.tensorflow.org/tutorials/layers

MNIST using TensorFlow
http://terrence.logdown.com/posts/1240896-play-tensorflow-mnist-handwriting-recognition

Face Landmarks with dlib, Opencv, and Python
http://www.pyimagesearch.com/2017/04/03/facial-landmarks-dlib-opencv-python/

Dlib Intro. [Chinese]
https://chtseng.wordpress.com/2016/12/23/dlib-%E5%A5%BD%E7%94%A8%E7%9A%84%E7%9A%84machine-learning%E5%B7%A5%E5%85%B7-%E4%B8%80/

CNN example reference
http://arbu00.blogspot.tw/2017/03/2-tensorflowconvolutional-neural.html

LBP by sk-image
http://scikit-image.org/docs/dev/auto_examples/features_detection/plot_local_binary_pattern.html

About the codes

ML_Final_Project.py: Top module for the project.
data_conversion.py : Convert images to binary data for faster access for TensorFlow.
Modules.py : Modules used throughout the project.
pending_code.py : Codes that might be added to the project.

Roadmap

(Done) Facial Recognition
(Done) Data Augmentation
(Done) Convert to TensorFlow data type
(Done) Implementing CNN with TensorFlow
(Almost Done) Make the model usable
(In Progress) Further tuning the model (Refine the model)
(In Progress) Choice of optimizers and adaptive learning rate if valid
(Coming Soon) Other Image Processing (LBP -> HOG -> ...)
(----) Re-train the model and fine tuning
(----) Demo code preparation
(----) Report

Current Progress

Working on...

Trying to converge the model faster.
Other optimizers?
Structure of the model

Issues

#1 Load testing data as batch for validation.
#2 Unusable model (Cause UNKNOWN)
#3 Out-of-Memory... perhaps not well optimized :(

TO-DOs

Parameterized the model
Reduce training time and computational resources
Better face recognition
Find effective data augmentation methods
[DONE] Increase the amount of data (size of dataset = 39xxx)
Other image processing techniques
Demo code preparation
Adaptive learning rate
Choose proper optimizer
https://www.tensorflow.org/api_guides/python/train#Optimizers

Small scripts to make our life much easier

1. routine

Simple git commit and push at once, and add files if given.

How-to:

Method 1. Run the file as bash

sh routine

Method 2. Run the script as executable

(required step, run the command just once when moving the file across machines)

chomod 777 routine

(execute step)

./routine

Usage:

Just update files that already exist.
Commit message: "routinely commit"

./routine

Add one or multiple new files to the project.
Commit message: "File(s): [list_of_files] added to the project"

./routine [list_of_files]

example:

./routine file1 file2

2. visualize

Open Tensorboard (path adapted to our library path, change it if needed) without TensorBoard install by pip

Usage

./visualize

If TensorBoard is installed

tensorboard --logdir=/path/to/log-directory

e.g. if the log stores in log/

tensorboard --logdir=log/

Installation Reference

Dlib
http://www.pyimagesearch.com/2017/03/27/how-to-install-dlib/
OpenCV
http://www.pyimagesearch.com/2016/12/19/install-opencv-3-on-macos-with-homebrew-the-easy-way/
TensorFlow (build from source)
https://www.tensorflow.org/install/install_sources

Reference for coding

Convert label to one-hot
https://agray3.github.io/2016/11/29/Demystifying-Data-Input-to-TensorFlow-for-Deep-Learning.html

label = tf.stack(tf.one_hot(label-1, nClass))

Fetch mini-batch from queue
https://stackoverflow.com/questions/41978221/tensorflow-next-batch-of-data-from-tf-train-shuffle-batch

init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    threads = tf.train.start_queue_runners(sess=sess)
    minibatch = sess.run([data, label])

Release GPU memory after computation
tensorflow/tensorflow#1578

config = tf.ConfigProto()
config.gpu_options.allow_growth=True
with tf.Session() as sess:
sess = tf.Session(config=config)

Open TensorBoard on local machine from remote server
https://stackoverflow.com/questions/37987839/how-can-i-run-tensorboard-on-a-remote-server

ssh -L local_machine_port:127.0.0.1:6006 username@remote_server

Where 127.0.0.1:6006 represents the port 6006 on remote_server, this command forwards all contents on the remote_server:6006 to your local machine.

e.g. if 16006 is the port that we want to use as local_machine_port, then

ssh -L 16006:127.0.0.1:6006 username@remote_server

To open Tensorboard, we can launch the TensorBoard on the remote server and simply accesss the port 16006 on our local machine.

127.0.0.1:16006

danielyou0230 / apparent-age-gender-classification