Fake or Real Face Image- Autoencoder and CNN classifier
We have two datasets one with 10K rgb unlabeled images with size 256x256 which are used to train an Autoencoder model. The goal is to save its weights and then use the Encoder part to create a classifier and fine tuning upon the second dataset which has 10K rgb images labeled as 'Real' and 10K images labeled as 'Fake'. The second dataset's images are also 256x256 pixels. The Autoencoder was develoveped using Google Colab and jupiter notebook in order to fast prototype and evaluate a few architectures, till I chose one that suits better. Then I used the saved model weights and by keeping the Encoder part of the model I stacked on top of a classifier architecture [where Encoder parameters were set to non-trainable]. This is a basic concept of Fine-Tuning a NN model using two types of datasets (Unlabeled and Labeled).
Both models are based on Convolutional Neural Networks a.k.a CNNs, which are implemented using Keras API of Tensorflow.
| Name | Description |
|---|---|
| datasets | Directory with the 2 types of image datasets |
| saved_models | Includes the models' architecture, weights, etc. |
| classifier_summary | Images showing classifier architecture |
| medver_classifier.py | The classifier train process --run locally |
| medver_autoencoder.ipynb | Research & results of different architectures run on Google Colab, here is the Autoencoder notebook |
| medver_autoencoder.py | Autoencoder --raw python |
| train_classifier_log.csv | Training results of first classifier architecture ** |
| train_classifier_log2.csv | Training results of second classifier architecture ** |