/Age-Gender-Ethnicity-Predictor

Predicts age,gender,ethnicty of image of person uploaded

Primary LanguageJupyter NotebookMIT LicenseMIT

Age-Gender-Ethnicity-Predictor

Predicts age,gender,ethnicty of image of person uploaded

Go to the app to test :)

may take few minutes to load

zoom-out to 75% on PC for better experience.

Details of Files present in Repository

  • images

    • contains images needed for frontend

  • static

    • css
      • contains CSS files
    • js
      • contains Java script files

  • templates

    • contains HTML files

  • uploads

    • image uploaded during prediction are temporarily stored in this folder. Make sure this folder is not empty during deployment, as git will not track empty folders.

  • app.py : python file

  • mobilenet_model.zip : pretrained model (lucky that the compressed model is less than 25MB 😄)

  • Procfile : needed for Heroku to know what commands it need to run.

  • requirements.txt : to specify the libraries needed to install to Heroku.

  • runtime.txt : to specify specific version of python.

Details of Project

Age Gender Ethnicity Predictor is web-app deployed on Heroku.

  • Model Training

    Go to my training notebook on kaggle
    • Library used : TensorFlow
    • Pretrained Model : MobileNetV2 (Due to 500MB limit on heroku I used smaller model. A big model with better training will give better results)
    • Technique used : MultiLabel classification.
    • Dataset : UTK Face Dataset
    • My model will output a 19D array as output when given an image as input (resized to (200,200,3))
    • ID 0-11 gives information about Age
      • MultiClass classification.
      • if argmax(predicted[:12] = 3) that means the age is between 3*10+1 and 3*10+10 ie., 31-40.
    • ID 12-13 gives information about Gender
      • Binary Classification
      • 0- Male 1- Female
    • ID 14-18 gives information about Ethnicity
      • MultiClass Classification
      • 0- White, 1- Black, 2- Asian, 3-Indian, 4-Others

Note : All the labels used are defined in UTK Face Dataset

Things I couldn't do because of memory constraint 😢

  • Using MTCNN to detect whether the face is present in uploaded image or not.
  • Using better pretrained models for training on UTK face dataset.