Ilya Tolstikhin, Olivier Bousquet, Sylvain Gelly and Bernhard Scholkopf
ICLR 2018
There are three qualitative results as proposed in the paper: image reconstrution, interpolation and random-sampling. Image reconstruction and interpolation are applied to test images.
Image reconstruction performs the reconstruction of an image using an encoder and a decoder.
Interpolation steps of two images (x1, x2) on latent space Z are generated.
A latent code z is sampled from a fixed prior distribution on a latent space Z. Then, z is mapped to the image x on input space X.
FID is calculated using 1K samples.
Fréchet Inception Distance (FID) = 99.75676458019791
This folder provides a re-implementation of this paper in PyTorch, developed as part of the course METU CENG 796 - Deep Generative Models. The re-implementation is provided by:
- Ali Abbasi, [email protected]
- Samet Cetin, [email protected]
Please see the jupyter notebook file main.ipynb for a summary of paper, the implementation notes and our experimental results.
Execute following command to install requirements:
$ pip install -r requirements.txt
Execute following command to download pretrained encoder and decoder weights into checkpoint/wae-mmd/ directory:
$ bash download_data.sh