Face Generation with Deep Convolutional Generative Adversarial Network

This repo implements Deep Convolutional Generative Adversarial Network (DCGAN) (Paper Link) to generate 32x32 RGB images of faces. The DCGAN architecture is implemented in PyTorch. The network can be modified to generate faces of higher resolution by training it with more higher resolution images.

Dataset

For training CelebFaces Attributes Dataset (CelebA) has been used. The processed dataset contains 32x32 images to reduce compute time.

• Original Dataset Link

• Processed Dataset Link

Setup

Run following commands

 pip install -r requirements.txt  

Files

All code for training and network architecture can be found in DCGAN_for_Image_Generation.ipynb. For more detailed explanation of core concept Original Training Notebook/dlnd_face_generation.ipynb can be studied.

Screenshots

Sample 1	Sample 2

Compute Resource

With batch size = 64 and model defined in training notebook, It takes 1 min (Approx) to run a single epoch on following hardware resources available on Google Colab

    Intel(R) Xeon(R) CPU @ 2.20GHz [Core(s) per socket:  1 | Thread(s) per core:  2 ]
    Tesla T4 [CUDA Version: 10.1]