This project is an exploration of Generative Models (GM) and its capabilities, focusing on the generation of bicycle images using Wasserstein Generative Adversarial Networks (WGAN-GP) in conjunction with estimators and generators.

WGAN-GP (Wasserstein GAN with Gradient Penalty) is a specific type of generative adversarial network (GAN) utilized for generating realistic data, particularly images.

• Architecture:

  • WGAN-GP: WGAN-GP represents a variant of the GAN framework that places significant emphasis on enhancing training stability and the quality of generated outputs. It leverages the concept of Wasserstein divergence, a metric measuring the dissimilarity between two probability distributions. WGAN-GP introduces a gradient penalty mechanism to control the gradient norms of the discriminator, thereby fostering more stable training and smoother gradients.

• Loss Function:

  • WGAN-GP: In contrast to traditional GANs, WGAN-GP employs the Wasserstein divergence loss, also referred to as the Earth-Mover (EM) loss. This loss measures the dissimilarity between probability distributions and is considered advantageous for improving the quality of generated data. In addition to this loss, WGAN-GP incorporates a gradient penalty component to enhance training stability further.

• Training Stability:

  • WGAN-GP: The primary objective of WGAN-GP is to enhance training stability. By incorporating a gradient penalty, it mitigates issues that can affect some conventional GANs, such as mode collapse, where the generator tends to produce similar-looking samples.

In summary, WGAN-GP is a specialized GAN variant tailored to enhance the training stability and output quality, particularly when generating images.

• Develop and train a powerful WGAN-GP model using the expansive QuickDraw dataset.
• Cultivate a deep understanding of the cutting-edge WGAN-GP architecture and Generative AI techniques.

The Quick Draw dataset is a treasure trove of approximately 50 million drawings, contributed by real artists. For our endeavor, we have curated a subset consisting of 117,555 meticulously crafted bicycle drawings.

Access the QuickDraw Dataset:

• Dataset Repository: https://github.com/googlecreativelab/quickdraw-dataset
• Numpy Bitmap Files: https://console.cloud.google.com/storage/quickdraw_dataset/full/numpy_bitmap
• Bicycle Dataset: https://storage.googleapis.com/quickdraw_dataset/full/numpy_bitmap/bicycle.npy

• Setup: Imports and Parameters.
• Data Exploration: Discovering bicycle drawings in the Dataset.
• Model Architecture: Designing a WGAN-GP (Wasserstein Generative Adversarial Network with Gradient Penalty).
• Model Building: Creating the GAN Model.
• Model Training: Feed data to the model and watch as it learns to generate bicycles.
• Artistic Analysis: Delving into the generated Bicycles.

| View on Kaggle |