taufikiqbalr / realtime-deep-prediction-tensorflow Goto Github PK

This repository is inspired from the following github pages:

• FaceNet
• UTKFace
• Race/Gender recognition
• Age/Gender Estimate

More information: https://upaspro.com/tesorflow-facenet/

Install dlib with instruction from dlib.

conda create --name <myenv> python=3.5
conda activate <myenv>
pip install opencv-python
pip install -r requirements.txt

• Race/Gender
• ARG

Source

The labels of each face image is embedded in the file name, formated like [age]_[gender]_[race]_[date&time].jpg

• [age] is an integer from 0 to 116, indicating the age
• [gender] is either 0 (male) or 1 (female)
• [race] is an integer from 0 to 4, denoting White, Black, Asian, Indian, and Others (like Hispanic, Latino, Middle Eastern).
• [date&time] is in the format of yyyymmddHHMMSSFFF, showing the date and time an image was collected to UTKFace

Handle error files from UTKFace:

mv 39_1_20170116174525125.jpg.chip.jpg 39_0_1_20170116174525125.jpg.chip
mv 61_1_20170109150557335.jpg.chip.jpg 61_1_3_20170109150557335.jpg.chip
mv 61_1_20170109142408075.jpg.chip.jpg 61_1_1_20170109142408075.jpg.chip

Put UTKFace folder in the same level of hierarchy of either two version of CMPT726 (namely ARG or RaceGender)

gender_labels
0: male, 1: female
{'0': 12391, '1': 11317}

make_features.py

Create a folder inside CMPT726. Functionality:

• get images from directory: UTKFace/*.jpg
• split data in to training, validation and test set
• data augmentation to balance the representation each race in the total population
• save to `.tfrecords files

multitask_model.py

Functionality:

• read_and_decode: read, decode data from training .tfrecord; and preprocess the data ()
• build_model: load network graph and add layers
• a few options of how to add layers, add_layer as the most basic one, add_layer_v2 added additional FC before auxiliary branches
• compute losses

trainer.py

python trainer.py --project_dir /data/gender_race_face/ --model_name model1_gender_0001 --learning_rate 0.0001 --num_epoch 2

Before running:

• Create 'models' inside 'CMPT726' directory
• download the pretrained VGGFace2 weights
• create 'detectors' inside 'models' directory
• create 'logs' inside 'CMPT726' folder
• create 'model_1' and 'model_2' inside the 'logs' directory

Functionality

• set training parameters and log directory
• train steps
• log to tensorboard

evaluator.py

python evaluator.py --project_dir /data/gender_race_face/ --model_name model1_gender_0001 

Functionality:

• evaluate trained model
• log to tensorboard

run_evaluator.sh

Functionality:

• run specified evaluator 50 times in a loop

model_analyst.py

python model_analyst.py --experiments_log experiemnts_combined4.txt --model_name model2_combined --project_dir /data/gender_race_face/

Dependencies: numpy
Functionality:

• if --experiments_log is not null, stats for multiple runs of the model will be computed
• if --model_name is not null, predictions will be genereated and saved in pickle file
• a hidden function additional_analysis will take a pickle file and compute confusion matrix and draw montages for detailed error analysis
• for additional_analysis, additional dependences: imutils, open-cv are required

model_comparator.py

python model_comparator.py ----gender_logs experiemnts_gender.txt --race_logs experiemnts_gender.txt --multitask_logs experiemnts_combined4.txt

Dependencies: scipy
Functionality:

• if --gender_logs and multitask_logs are not null, t test is performed to compare gender models
• if --race_logs and multitask_logs are not null, t test is performed to compare gender models

predictor.py

Dependencies: dlib Also, might have to use an older version of python.

• preprocess: chop and align the face
• predict use saved model

Training parameters after 50 epocks: 0.63, 0.38, and 0.15 cross-entropy loss for age, race and gender evaluation, respectively.
Validation accuracy with bining: 73, 84, 93% for age, race and gender prediction, respectively.