danielkonecny / generating-faces

Generating faces with GANs and analyzing embedding space distribution for different classes as a Bachelor's Thesis at BUT FIT.

Generating Faces with Generative Adversarial Networks

Requirements

• Python 3 64bit, preferably pip, Jupyter Notebook
• Libraries: NumPy, TensorFlow, matplotlib, imageio, PIL, glob, os, json, math, IPython.display, joblib, pandas, SciPy

Programmed models

5 models were implemented.

• mnist32x32: for handwritten digits.
• lfw32x32bw: for LFW database (grayscale and cropped).
• lfw64x48bw: for LFW database (grayscale).
• lfw64x48color: for LFW database.
• ffhq128x128: for FFHQ database.

General information about models

Each model has a similar folder and code structure.

• analysis: saved models used for latent space analysis.
• animations: animations of training progress in gif format.
• datasets: saved datasets ready for training.
• features: manually labeled samples for analysis.
• grids: examples generated during the training.
• images: generated images of different kinds.
• models: saved models of generator.
• weights: saved weights of all models.
• ..._train.ipynb: code for training of a GAN.
• ..._generate.ipynb: code for generating images.
• ..._analyze.ipynb: code for analyzing images.