This repository contains code and documentation for image generation using Generative Adversarial Networks (GANs). GANs are a class of machine learning models designed to generate realistic-looking data, particularly images. The model consists of a generator and a discriminator, both trained simultaneously through adversarial training.
Generative Adversarial Networks were introduced by Ian Goodfellow and his colleagues in 2014. GANs consist of two neural networks, the generator and the discriminator, engaged in a two-player minimax game.
-
Generator: The generator creates new data instances that resemble a given dataset.
-
Discriminator: The discriminator evaluates the authenticity of a given image, determining whether it is from the real dataset or generated by the generator.
The generator and discriminator are trained simultaneously, with the generator trying to create increasingly realistic data, and the discriminator trying to become more adept at distinguishing real data from generated data. This adversarial process results in the generator creating data that is indistinguishable from real data.
The GAN architecture consists of two main components:
-
Generator: The generator takes random noise as input and transforms it into data that resembles the training dataset. It typically consists of a series of layers (often transposed convolutional layers) that progressively upsample the input noise.
-
Discriminator: The discriminator evaluates the authenticity of an input image and outputs a probability score. It typically consists of convolutional layers that downsample the input image.
During training, the generator and discriminator are trained iteratively. The generator produces fake images, and the discriminator evaluates them along with real images from the dataset. The generator's objective is to create images that are so realistic that the discriminator cannot distinguish them from real images. The discriminator's objective is to become better at distinguishing between real and fake images.
The adversarial process continues until the generator generates images that are indistinguishable from real images according to the discriminator.
Once trained, the generator can take random noise as input and generate realistic images.
This repository contains code for generating dog images using Generative Adversarial Networks (GANs). The GAN architecture consists of a generator and a discriminator trained simultaneously to create realistic dog images.
DOGSGAN.ipynb: Jupyter Notebook containing the code for training a GAN to generate dog images.
The Discriminator class is responsible for evaluating the authenticity of input images. It consists of a series of convolutional layers with leaky ReLU activation and batch normalization.
The Generator class generates new images. It consists of transposed convolutional layers with batch normalization and ReLU activation, producing realistic dog images.
The dog images dataset is loaded using the ImageDataGenerator from Keras. Image augmentation techniques such as shear, zoom, and horizontal flip are applied to increase dataset diversity.
The GAN is trained over multiple epochs. The training loop alternates between updating the weights of the discriminator and generator. Real and fake images are used to calculate adversarial losses.
Generated images are saved in the /content/Results/ directory during training. You can visualize the generated images using the saved PNG files.
Tensorboard logs are not implemented in the notebook. You can consider adding Tensorboard visualization for monitoring training progress.
Feel free to customize the code and experiment with different hyperparameters, architectures, and datasets. You can also train the GAN on your own dataset by replacing the training data.
This project was inspired by the groundbreaking work on GANs by Ian Goodfellow and his colleagues.
This project is licensed under the MIT License - see the LICENSE file for details.