sayedmohamedscu / vqgan

Vector-Quantized Generative Adversarial Networks

VQGAN

Vector-Quantized Generative Adversarial Networks:

This repository contains a simple implementaion the code for VQGAN without using Clip :

a deep neural network architecture that can generate images from it used the basic VQVAE followed by a discriminator .

The model consists of an encoder, a decoder, a vector quantization module and a discriminator .

Code

The code for VQGan is divided into two parts:

• train file

• Encoder_decoder_disc_VQ file

train file

This folder contains the code for training VQGan on your own dataset. You need to add your dataset to the data folder called Dataset.

To train the model, run the following command:

python train.py

Encoder_decoder_disc_VQ file

This file contains the code for the VQGan model. It has four parts:

• Encoder
• Decoder
• VQ
• Discriminator

Encoder

The encoder takes an input image and encodes it into a low-dimensional latent vector.

Decoder

The decoder takes the latent vector generated by the encoder and decodes it into an image.

VQ

The VQ module is responsible for learning the discrete latent space and providing the quantized latent code for the generator.

Discriminator

The discriminator is a neural network that takes an image as input and predicts whether it is real or fake. It is used to train the VQGAN model through adversarial learning.

Requirements

The code requires the following packages to be installed:

• torch
• torchvision
• matplotlib
• numpy
• torchsummary
• opencv-python

Dataset

MAinly was 102 Category Flower Dataset

Results