/generating-faces

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

Primary LanguageJupyter NotebookMIT LicenseMIT

Generating Faces with Generative Adversarial Networks

Requirements

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.