Gan and DcGan on Mnist and Cifar10

**This repository has been prepared for GAN and DCGAN implementation on CIFAR10 and MNIST datasets with using Pytorch and TensorFlow-Keras **

Requirements

To install requirements:

pip install -r requirements.txt

Folder structure:

GAN-DCGAN-Imp-on-Mnist-Cifar10
├── CIFAR10
|   |__DCGAN
|   |__GAN
├── MNIST
    |__DCGAN
    |__GAN

MNIST

The MNIST dataset is a widely used benchmark in the field of machine learning and computer vision. It consists of a collection of handwritten digits from 0 to 9, each represented as a 28x28 grayscale image. MNIST has been extensively utilized for developing and evaluating algorithms in image classification, digit recognition, and deep learning models.

MNIST

See also: PyTorch-GAN

CIFAR10

The CIFAR-10 dataset is a popular dataset used for image classification tasks in machine learning and computer vision. It comprises 60,000 color images in 10 different classes, with each rgb image having a resolution of 32x32 pixels. The classes include common objects such as airplanes, automobiles, birds, cats, dogs, and more. CIFAR10

IMPLEMENTATIONS

• DCGAN To train DCGAN models and save the model for both mnist and cifar10 datasets , You need to run the Train-Code , Train-Code The evaluations are inside the notebook for CIFAR10 dataset implementations.

  Evaluation and FID calculation for MNIST; Eval-Code

• GAN To train GAN models and save the model for both mnist and cifar10 datasets , You need to run the Train-Code , Train-Code

  The evaluations are inside the notebook for CIFAR10 dataset implementations.

  Evaluation and FID calculation for MNIST; Eval-Code

Models

Models are saved for further use.

  CIFAR-10
  ├── DCGAN
  |   |__model
  |   
  ├── GAN
  |   |__model
  |
  MNIST
  ├──DCGAN
  |    |__model
  |        |__gen
  |        |__gan
  |        |__disc
  |    
  ├──GAN
  |    |__model
  |        |__gen
  |        |__gan
  |        |__disc

Installation

$ git clone https://github.com/Alperitoo/GAN-DCGAN-Imp-on-Mnist-Cifar10

Results

Our model achieves the following performance on MNIST:

Model name	FID
GAN	123
-------------------	----------------
DCGAN	55.6
-------------------	----------------

Our model achieves the following performance on CIFAR-10

Model name	FID	Accuracy
GAN	159	0.67
-------------------	----------------	---------------
DCGAN	126	0.985
-------------------	----------------	---------------

References

• https://github.com/eriklindernoren/Keras-GAN
• https://nealjean.com/ml/frechet-inception-distance/
• https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html
• https://github.com/soumith/ganhacks
• https://github.com/mseitzer/pytorch-fid
• DCGAN original paper
• GAN original paper