This is the implementation of "Wasserstein Autoencoder" with MMD corresponding to 2-dimensional shape.
Original paper: I. Tolstikhin, O. Bousquet, S. Gelly, and B. Schölkopf. Wasserstein auto-encoders. In International Conference on Learning Representations, 2018. link

Please build the source file according to the procedure.

$ mkdir build
$ cd build
$ cmake ..
$ make -j4
$ cd ..

• Large-scale CelebFaces Attributes (CelebA) Dataset
  This is a large-scale face attributes dataset with more than 200K celebrity images, each with 40 attribute annotations.
  Link: official

Please create a link for the dataset.
The following hierarchical relationships are recommended.

datasets
|--Dataset1
|    |--train
|    |    |--image1.png
|    |    |--image2.bmp
|    |    |--image3.jpg
|    |
|    |--valid
|    |--test
|
|--Dataset2
|--Dataset3

You should substitute the path of training data for "<training_path>", test data for "<test_path>", respectively.
The following is an example for "celebA".

$ cd datasets
$ mkdir celebA
$ cd celebA
$ ln -s <training_path> ./train
$ ln -s <test_path> ./test
$ cd ../..

Please set the shell for executable file.

$ vi scripts/train.sh

The following is an example of the training phase.
If you want to view specific examples of command line arguments, please view "src/main.cpp" or add "--help" to the argument.

#!/bin/bash

DATA='celebA'

./WAE2d_MMD \
    --train true \
    --epochs 300 \
    --dataset ${DATA} \
    --size 256 \
    --batch_size 16 \
    --gpu_id 0 \
    --nc 3

Please execute the following to start the program.

$ sh scripts/train.sh

Please set the shell for executable file.

$ vi scripts/test.sh

The following is an example of the test phase.
If you want to view specific examples of command line arguments, please view "src/main.cpp" or add "--help" to the argument.

#!/bin/bash

DATA='celebA'

./WAE2d_MMD \
    --test true \
    --dataset ${DATA} \
    --test_in_dir "test" \
    --test_out_dir "test" \
    --size 256 \
    --gpu_id 0 \
    --nc 3

If you want to test the reconstruction error of the image, the above settings will work.
If you want to test the denoising of the image, set "test_in_dir" to "directory of noisy images" and "test_out_dir" to "directory of output ground truth".
However, the two file names must correspond.

Please execute the following to start the program.

$ sh scripts/test.sh

Please set the shell for executable file.

$ vi scripts/synth.sh

The following is an example of the synthesis phase.
If you want to view specific examples of command line arguments, please view "src/main.cpp" or add "--help" to the argument.

#!/bin/bash

DATA='celebA'

./WAE2d_MMD \
    --synth true \
    --dataset ${DATA} \
    --size 256 \
    --gpu_id 0 \
    --nc 3

If you want to generate image, the above settings will work.

Please execute the following to start the program.

$ sh scripts/synth.sh

Please set the shell for executable file.

$ vi scripts/sample.sh

The following is an example of the sampling phase.
If you want to view specific examples of command line arguments, please view "src/main.cpp" or add "--help" to the argument.

#!/bin/bash

DATA='celebA'

./WAE2d_MMD \
    --sample true \
    --dataset ${DATA} \
    --sample_total 100 \
    --size 256 \
    --gpu_id 0 \
    --nc 3

If you want to generate image, the above settings will work.

Please execute the following to start the program.

$ sh scripts/sample.sh

This code is inspired by WAE-pytorch.