A deep convolutional GAN was trained from scratch on the Large scale CelebFaces dataset[1] consisting of 200k images of faces.
A random sample vector as well as a dropout was used in both the Discriminator and the Generator to induce noise.
Various values of dropouts were tried.
Download the images into data/celeba/
folder in the same directory.
./train.py --d1 [argument] --d2 [argument]
d1 and d2
are dropout values that are applied on the odd and even layers of the network respectively.They can be independently chosen.
./sample_generator.py --d1 [argument] --d2 [argument]
Model state dictionaries for models with different dropouts values are provided for trained models in models
folder
Some generated face images are stored in samples
folder with the d1
and d2
used.
Decrease single mode generator outputs by adding more random noise to the generator network
[1] http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
[2] https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html