Generation of new, realistic looking faces by DC-GAN trained on celebrity faces
The project is about generating new, realistic looking images. I have taken CelebA dataset (you can download here) which contains images of celebrities. The images in CelebA images has been cropped to remove parts of the image that don't include a face, then resized down to 64x64x3 NumPy images.
Generative Adversarial Networks contains two networks.
- Generator: Generates fake images
- Discriminator: Detects if an image is real or fake
These two models fight against each other and both of them develop at the same time. The generator tries to produce images in such a way that discriminator should think those images are real. On the other hand, discriminator tries to correctly identify the real images and the fake images produced by generator. Both the models are built using Convolutional Neural Networks(Transpose convolutions in generator), BatchNorm layers and fully connected layers. From reading the Original DCGAN paper, they say:
All weights were initialized from a zero-centered Normal distribution with standard deviation 0.02.
Some of the hyper parameters chosen in the project are referred from the Original DCGAN paper like learning rate, beta1, beta2 for Adam optimizer and values for weight initialization and finally the model is trained on the GPU. The input images and model produced images are shown below. For detailed description, refer the notebooks.
- Python: language
- NumPy: library for numerical calculations
- Matplotlib: library for data visualisation
- Pytorch: a deep learning framework by Facebook AI Research Team for building neural networks
- torchvision: package consists of popular datasets, model architectures, and common image transformations for computer vision
To use this project, clone the repo
git clone https://github.com/Surya-Prakash-Reddy/Face-Generation.git
After cloning, you can use the dlnd_face_generation.ipynb notebook to modify the notebook or generate realistic looking faces. If you want to learn how to generate images and replicate the notebook, you can use dlnd_face_generation.html file.