/adversarial-autoencoder

Chainer implementation of adversarial autoencoder

Primary LanguagePythonMIT LicenseMIT

Adversarial Autoencoder

This is the Chainer implementation of Adversarial Autoencoder [arXiv:1511.05644]

この記事で実装したコードです。

Requirements

  • Chainer 1.6+

Running

Create "images" directory in the root or use "image_dir" option to specify the directory that contains training images.

Options:

  • --image_dir
    • Specify the directory that contains the training images.
  • --load_epoch
    • Specify the model you want to load.

Class label

Add the label index (must start 0) to the image filename.

format: [0-9]+_.+\.(bmp|png|jpg)

e.g. MNIST

example

Training

e.g. swiss roll distribution

cd swiss_roll

python train.py

Visualizing:

e.g. swiss roll distribution

cd swiss_roll

python visualize.py --load_epoch 10

MNIST Unsupervised Learning

Uniform (-2.0 ~ 2.0)

1,000 train data

Uniform

9,000 test data

Uniform

MNIST Supervised Learning

10 2D-Gaussian Distribution

1,000 train data

10 2D-Gaussian

9,000 test data

10 2D-Gaussian

Swiss Roll Distribution

1,000 train data

Swiss Roll

9,000 test data

Swiss Roll