Skip to content

Latest commit

 

History

History
31 lines (20 loc) · 992 Bytes

README.md

File metadata and controls

31 lines (20 loc) · 992 Bytes

Importance Weighted Autoencoders (IWAE)

Link to paper: https://arxiv.org/abs/1509.00519

AnalyticalIWAE: IWAE calculating loss manually

PytorchIWAE: IWAE using built-in torch functions to evaluate and calculation loss.
Includes example of algorithm very easy to apply to existing VAE (although a bit slower)

ConvIWAE: An example of convolutional IWAE, not integrated with main script, only as example

Importance Weighted Autoencoders - Gaussian encoder and decoder

Pytorch IWAE Loss Curve:

MNIST

Pytorch IWAE 60 epoch results:

MNIST sampled sampels

Training gif

Giffygifgif1

Importance Weighted Autoencoders - Gaussian encoder, Bernoulli decoder

Analytical IWAE Loss Curve:

MNIST sampled sampels

Analytical IWAE 60 epoch results:

MNIST sampled sampels

Training gif

Giffygifgif2