We have introduce and analysis our project more detail in ML_project.ppt
We choose the first 40000 Align&Cropped Face Images in CelebA for training with their gender attribute
ICGAN introduce an encoder to determine specific representation of generated images, which inverse the mapping of cGAN, given an input image x, to obtain its representation as a latent variable z and a conditional vector y, and we can modify z and y to re-generate the original image with complex variations.
- train cgan:
- D_loss = -(log(D(x,y)) + log(1-D(G(z,y'),y')))
- G_loss = -log(D(G(z,y),y))
- train enz:
- z_loss = l2_loss(z-enz(G(z,y')))
- train eny:
- y_loss = l2_loss(y-eny(x))
image source:ICGAN Paper