DCGAN

This is my Tensorflow implementation of Deep Convolutional Generative Adversarial Networks in Tensorflow proposed in the paper Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks. The main contribution comes from the tricks which stablize the training of Generative Adversarial Networks. The proposed architecture is as followed.

This project is an implementation of famous DCGAN algorithm. I have created a package named dc gan which contains the actual implementation of the algorithm.

Prerequisites

• Python 3.3+
• Tensorflow 2.3.1
• NumPy

Project Structure

Data

'data' folder contains the data for training the generator and discriminator.If you want to train the generator and discriminator on your dataset there is only one step to do this just replace the images in data/train folder with your images in data

weights

weights folder contains the saved weights for the generator and discriminator

Model

model class contains the code for DCGAN Model

functions :

summary(): This function prints the the layer wise information about generator and discriminator

train(): This function train the generator and discriminator 
    parameters : 
        dataset : a tensor of shape (None,None, 64,64,3)  which means dataset should be supplied in forms of batches
        epochs  : number of epochs to train the generator and discriminator

save: to save the weights of the generator and discriminator after training the model
    parameters: 
        path : base path to save the folder of weights of the generator and discriminator
        name : name for weights folder 

load_weights: to load the weights of the saved generator and discriminator
        parameters :
            path : path to load the weights of the generator and discriminator , this path should target the folder containing weights

generate_images(): This function generate multiple images and save them to directory
    parameters : 
        count : the number of image to be generated 
        save_dir : the directory to save the generated images

generate_image() :To generate a image 
    return :
        a 3D images tensor

Results During Training

MNIST

• Generated samples (200th epochs)

• First 50 epochs

CFAR

• Generated samples (500th epochs)

• First 40 epochs

Training tricks

• To avoid the fast convergence of the discriminator network
  • The generator network is updated more frequently.
  • Higher learning rate is applied to the training of the generator.
• One-sided label smoothing is applied to the positive labels.
• Gradient clipping trick is applied to stablize training

Related works

• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks by Radford

• Fashion MNIST

• Training Tips i borrowed from shaohua0116 training