GAN for Increasing CNN Accuracy

GAN for Increasing CNN Accuracy

‍

The code in this repository is part of my Master thesis: “Using data generated by a GAN to improve the accuracy of a CNN image classification model in the case of a small amount of initial data".

Checkout the code in my Github: https://github.com/nicolotognoni

‍

ABSTRACT

There are certain real-world situations where it’s either impossible or too expensive to gather a huge quantity of data in order to effectively train deep learning models.

The objective of the thesis would be to circumvent this challenge by using additional data artificially generated by GAN models.

We want to see if the additional data created using the initial data, will improve the performance of the model.

We expect there will be a certain threshold in the number of data initially available to train the GAN models, under which it would be inefficient to use this technique. We want to find out the minimum number of data (images in our case) needed to improve the deep learning model.

INFO

• StyleGAN2_ADA.ipynb - code to train the StyleGAN
• generate_synthetic_data.ipynb - code to generate synthetic images from a pre-trained pkl file, previously created with styleGAN
• CNN.ipynb - Convolutional Neural Network we created to test the accuracy before and after the implementation of synthetic data in the original dataset

‍