Implementation of Autoencoding beyond pixels using a learned similarity metric in Keras.

EE298 Group 4;

Peralta, Daryl Mendaros, Jonathan Aslan, Cha-dash

• tensorflow >=1.4
• keras >= 2.1.4
• OpenCV >= 3.4.0
• numpy

For training:

python train.py --dataset [path to dataset]

Ex.

python train.py --dataset /home/daryl/datasets/img_align_celeba

For testing:

python test.py --dataset_path [path dataset] --encoder_path [path to encoder] --decoder_path [path to decoder]

Ex.

python test.py --dataset_path '/home/daryl/datasets/img_align_celeba' --encoder_path checkpoints/encoder_chk-vaegan_complete_demo.hdf5 --decoder_path checkpoints/decoder_chk-vaegan_complete_demo.hdf5 --dataset_path /home/daryl/datasets/img_align_celeba

Checkpoints can be found here.

CelebA dataset

• 202599 celebrity images

VAEGAN encoder model:

VAEGAN decoder model:

VAEGAN discriminator model:

VAEGAN encoder model for training (only the encoder is trainable):

VAEGAN decoder model for training (only the decoder is trainable):

VAE generation from noise:

VAE autoencoder input:

VAE autoencoder reconstruction:

GAN generation from noise:

VAEGAN generation from noise:

VAEGAN autoencoder input:

VAEGAN autoencoder reconstruction:

VAEGAN generation from noise:

VAEGAN autoencoder input:

VAEGAN autoencoder reconstruction:

VAEGAN generation from noise:

VAEGAN autoencoder input:

VAEGAN autoencoder reconstruction: