C3AE: Age_estimation - Implemented in Notebook (Google Colab)

C3AE: Facial Age Estimation Project Implementation in Google Jupyer/Colab notebook. A Very Easy Unofficial Implementation of C3AE(Age estimation) model in Google Colab using keras. I have kept the Code very simple & wrote too many comments, almost on every line.... for better understanding of beginners like me:) & whoever wants to build something on top it....

Results

Project Strcuture

C3AE-keras/
    |->datasets                (datasets should be placed here, e.g wiki_crop)
    |->detector                (contains face landmarks detector model. (used with dlib))
    |->model_saved             (will save weights here (best trained model weights)
    |->net_training            (contain scripts to define network and utils)
    |->preprocessing_scripts   (contain scripts for dataset preprocessing. initially added WIKI, IMDB & morph preprocessing modules. Which will load datasets, detect faces+landmarks, encode images & save as feather.)

Code tested with

0. Google Colab
1. tensorflow : 2.3.0
2. keras : 2.4.3
3. cv2 : 4.1.2
4. dlib : 19.18.0
5. feather : 0.4.1
6. numpy : 1.18.5
7. pathlib

Steps

Open "C3AE_Notebook_Implementation.ipynb" as a notebook, and start executing each cell step by step. as follows,

I have divided each part into sections. As follows.

1. Clone repo : to get the code
2. Download Datasets : This will Download datasets and extract into "C3ae_keras/datasets/". (Initially contains links of wiki and imdb only, because morph is not publically available)
3. Preprocess datasets : Go to "preprocess_WIKI-IMDB.py" script first to change "dataset_name" varible to "wiki" or "imdb" for preprocessing coresponding dataset.
4. Train: 1. import libs 2. load pre processed dataset 3. train(0.8)/test(0.2)_split 4. Make Data Generator(because dataset is too large) 5. build net, train, and display history
5. Inference : Place an image in "/content/test.jpg" and run all cells in this section.

Refrence and Credits

1. Original paper : C3AE: Exploring the Limits of Compact Model for Age Estimation
2. Starter code Taken from: StevenBanama's C3AE implementation (https://github.com/StevenBanama/C3AE). Great Work!