rheehot/tf-alae

(Unofficial) Tensorflow implementation of Adversarial Latent Autoencoder (ALAE, Pidhorskyi et al., 2020)

tf-alae

(Unofficial) Tensorflow implementation of Adversarial Latent Autoencoder (ALAE, Pidhorskyi et al., 2020)

• Original paper, Adversarial Latent Autoencoder [arXiv:2004.04467]
• Full code is based on original repository [GIT].

Usage

To train the mnist model

# to train mnist-alae
python mnist_mlp.py train

# to train style-alae
python mnist_style.py train

To open tensorboard summary

tensorboard --logdir summary

Currently, lsunbed-StyleALAE is experimental.

# to train lsunbed
python lsunbed_style.py train

To use released checkpoints, download files from release and unzip it.

Following is example of MNIST-MLP.

import json
from mnist_mlp import MnistAlae

with open('settings.json') as f:
    settings = json.load(f)

alae = MnistAlae(settings)
alae.load_weights('./mnist_mlp/mnist_mlp')

Jupyter notebook

• mnist_expr.ipynb: MNIST interpolation with Mlp-ALAE
• mnist_style.ipynb: MNIST interpolation with Style-ALAE

Learning Curve

Mlp-ALAE + MNIST

Style-ALAE + MNIST

• In the original paper, they claim that latent reconstruction has better perceptual quality than image one so that they do not use image reconstruction loss in the original repository.
• But for more easy training, this repository use image reconstruction loss as pretraining at half of the epochs in each resolution level.

Sample

MNIST-MLP 0 ~ 4 polymorph

MNIST-Style 0 ~ 4 polymorph