CBEGAN

• A simple conditional version of the Boundary Equilibrium Generative Adversarial Networks (CBEGANs)
• M. Mirza et. al, Conditional Generative Adversarial Nets, 2014
• D. Berthelot et. al, BEGAN: Boundary Equilibrium Generative Adversarial Networks, 2017
• We generate male or female face images with latent code (i.e. z) conditioned on gender-attribute vector.

data preperation

• download CelebA dataset: download Align&Croppped images, Attribute Annotations, and Train/Val/Test partitions
• Run parseCelebA_gender_faceCrop.py (take care of specifying celebRawImgRoot and list_attr_celeba.data.txt)
• We found out if we use celeba raw image without face-crop, the algorithm does not generate male or female images well.

Results

• Train with main_CBEGAN.py: You should check meta parameters properly in training.
• CUDA_VISIBLE_DEVICES=x python main_CBEGAN.py --dataroot /path/to/CelebA/gender_facecrop/train --exp /path/to/dir/for/checkpoints --cond_size 2
• Run interpolateCond.py: We generate z with uniform distribution (from -1 to 1) and then set condition vector to [1, -1] or [-1, 1]
• CUDA_VISIBLE_DEVICES=x python interpolateCond.py --exp /path/to/dir/for/saving/result --netG /path/to/your/netG_epoch_xx.pth