Generative Adversial Network (WGAN-GP)

This project is an impletation of a WGAN-GP on the datatset CelebA with PyTorch.

What is GANs ?

Generator

The role of the Generator is to create images that appear realistic enough to fool the Critic.

• Objective: Produce images that mimic the original dataset.
• Functioning: It takes a random noise vector and transforms it into an image.
• Expected Outcome: That the Critic cannot distinguish the generated images from those actually from the dataset.

Critic (or Discriminator)

The Critic's task is to differentiate generated images from real images from the dataset.

• Objective: Correctly identify whether an image is generated by the Generator or is from the original dataset.
• Functioning: It evaluates images one by one and delivers its verdict.
• Expected Outcome: Maximize its ability to make the correct distinction.

How to train a GAN ?

Technical specifications

Property	Value
Device	MAC M2
Training Time	15 hours
Epochs	15
Training Data	CelebA
Framework	PyTorch
Learning rate	0.0002
Z_DIM	128
BATCH_SIZE	128
Images size	(64,64)
LAMBDA	10
N_CRITIC	4
Optimizer	Adam

How to improve this model ?

This model is not perfect and to improve it we need to consolidate its architecture and increase the number of epochs.