Deep Feature Consistent VAE With PaddlePaddle
November 15, 2021 · View on GitHub
1.Download CelebA and Pretrained VGG19
Download CelebA @AI Studio CelebA and put data at ./dataset
Download pretrained VGG19_bn @PaddlePaddle (password: 25il) and put data at ./paddle_dfcvae/architectures/pretrained
then
sh preparedata.sh
2.Train paddle dfcvae
cd ./paddle_dfcvae
python trainer.py
chose vgg layer feature
vim extract_features @ ./paddle_dfcvae/architectures/dfcvae.py
3.Generated imgs
python generatedImg.py
4.Reconstruct
vae123

vae345

vae123 extract the low features,clearer than vae345
5.Generated

6.Logs , Models and More Imgs
BaiduYun (password: 9vs9)
19967 Reconstruct pics(2x10)
100 Generated pics(10x1)
7.Measure
IS:
| mean | std | |
|---|---|---|
| IS | 1.1636 | 0.0221 |
| cifar10 | 1.1698 | 0.0355 |