non-variational-autoencoders

Authors: Grigorii Sotnikov, Vladimir Gogoryan, Dmitry Smorchkov, Ivan Vovk

Unsupervised Domain Translation

References

https://arxiv.org/pdf/2001.05017.pdf https://github.com/oripress/ContentDisentanglement

Reproducibility

Link to download aligned images in CelebA https://drive.google.com/open?id=0B7EVK8r0v71pZjFTYXZWM3FlRnM

Link to download list_attr_celeba.txt https://drive.google.com/drive/folders/0B7EVK8r0v71pOC0wOVZlQnFfaGs

Run a command: python preprocess.py --root ./img_align_celeba --attributes ./list_attr_celeba.txt --dest ./glasses_train

After this command in directory /glasses_train will be located 4 files: trainA.txt, trainB.txt, testA.txt, testA.txt

Then by starting from Disentanglement.ipynb a mode can be trained from the config.yml

The other possible option is to use pretrained models which are located at https://drive.google.com/open?id=1sA6BXedG23_ZTuES_udeQlYXrssFq2BU and proceed to visualization.ipynb where beautiful visualizations can be plotted via written functions

Files description

1) disentanglement.ipynb - train UDT model
2) modules.py - parts of the model: E1, E2, Decoder, Disc
3) utils.py - necessary functions for reproducible research
4) preprocess.py - given folder img_align_celeba produces separate dataset partitioning
5) visualization.ipynb - a notebook where you can plot beautiful interpolations between samples in different settings. Source of visualization for the project.

Adversarially Constrained Autoencoder Interpolation

References

https://arxiv.org/pdf/1807.07543.pdf

Files description

1) evaluation.py - fit a FC on top of trained latent representations
2) modules.py - blocks of models
3) train.py - fit ACAI and Baseline models with necessary losses
4) utils.py - load datasets
5) visualize.py - plot interpolated samples during training
6) visualization.ipynb - obtain interpolations for report and score models via FC layer
7) acai_train.ipynb - train baseline AE and ACAI from configs

On the Transformation of Latent Space in Autoencoders

References

https://arxiv.org/pdf/1901.08479.pdf

The main goal of the proposed framework is to transform vectors in the latent space. Reducing distances between vectors in without changing their topological properties should lead to more relevant interpolation between output vectors.

Files description

1) check_2d_minmax.ipynb - get pictures of latent representation for LTAE-M
2) check_2d_standard.ipynb - get pictures of latent representation for LTAE-S
3) check_ae.py - functions for performance measurenment of learned LTAE 
4) compute_hausdorff.py - calculate Hausdorff distance of the model
5) eval_ltae_features.ipynb - check trained LTAE
6) model.py - implemented achitectures
7) tests.ipynb - basic experiments 
8) train.py - train a model 
9) utils.py - useful funcs for reconstruction, interpolatin etc.

rorcde / non-variational-autoencoders Goto Github PK