GAN-based Semi-supervised learning for Image Classification

GAN-based Semi-supervised learning for Image Classification in 3 datasets: MNIST, DOODLE (subset of Quick, Draw! dataset with 10 classes) and EMNIST-letters with 3 algorithms: KNN, CNN and GAN-SSL

Features

The project will allow you to draw digits on the canvas predicting in real-time image

There will be 3 datasets to choose from:

MNIST: Handwritten digits
DOODLE: subset of Quick, Draw with 10 classes namely: helicopter, car, book, windmill, cat, umbrella, octopus, bird, hat, birthday, cake
EMNIST-letters: a-z alphabet both upper and lower letter

And for type of training:

First 3 is when using 50, 100, 500 labelled data respectively for MNIST and DOODLE. 100, 200, 500 labelled data for EMNIST
Last one is when using all data to train

Requirement

To run this project, you must have Python version >= 3.9 installed.

How to use:

To use this project using git, please do the following.

Clone the repository's main branch into your desired directory using your git command prompt and access the directory by this command.

git clone -b master https://github.com/trinhminh11/GAN-based-Semi-Supervised-Learning

for lighter used (exlude KNN) used:

git clone -b MT https://github.com/trinhminh11/GAN-based-Semi-Supervised-Learning

for even lighter used (only app, without training and KNN) used:

git clone -b demo https://github.com/trinhminh11/GAN-based-Semi-Supervised-Learning

access to the directory:

cd GAN-based-Semi-Supervised-Learning

Install required libraries via requirements.txt

pip install -r requirements.txt

Run the project with

python main.py

python3 main.py

if you want to train, run training_{Algorithm}.ipynb, changes USED_DATA and NUM_LABELLED and DEVICE in config.py

Data Preporcessing

The above datasets' original dimensions are 28 x 28 x 1 RGB images. However, we will resize them to 32 x 32 x 1 images and adjust the pixel values from 0 to 255 to a range of -1 to 1.

K-nearest neighbors

We will implement with euclidean distance with K = 3.

Convolutional Neural network

The Architechture of CNN are represent as follow:

GAN-based Semi-Supervised Learning

We use the same Architechture of CNN for Discriminator and using trained Generator and trained Discriminator with below Loss function mentiond in Improved Techniques for Training GANs (https://arxiv.org/abs/1606.03498):

with

Note

Because the data store in KNN are too large to upload to github, if you want to evaluate KNN when using all labelled data, you should run kNN.ipynb first for each USED_DATA set in config.py to get pickle file for KNN Algorithm