/keras_BEGAN

Implementation BEGAN([Boundary Equilibrium Generative Adversarial Networks](https://arxiv.org/pdf/1703.10717.pdf)) by Keras.

Primary LanguagePython

About

Implementation BEGAN(Boundary Equilibrium Generative Adversarial Networks) by Keras.

Version

Developed by these software versions.

  • Mac OS Sierra: 10.12.4
  • Python: 3.5.3
  • Keras: 2.0.3
  • Theano: 0.9.0
  • Pillow: 4.1.0

How to Use

Setup

pip install -r requirements.txt

Create Dataset

Prepare Image

You can use any square images. For example,

images in http://vis-www.cs.umass.edu/lfw/

[new] All images aligned with deep funneling 
(111MB, md5sum 68331da3eb755a505a502b5aacb3c201)

Convert Images to 64x64 pixels

Install imagemagick

For convert command, install imagemagick.

brew install imagemagick

Convert Images

  • ORIGINAL_IMAGE_DIR: dir of original JPG images
  • TARGET_DIR: dir of after converted images
ORIGINAL_IMAGE_DIR=PATH/TO/ORIGINAL/IMAGE_DIR
CONVERTED_DIR=PATH/TO/CONVERTED/IMAGE_DIR

mkdir -p "$CONVERTED_DIR"
for f in $(find "$ORIGINAL_IMAGE_DIR" -name '*.jpg')
do
  echo "$f"
  convert "$f" -resize 64x64 ${CONVERTED_DIR}/$(basename $f)
done

Create Dataset

PYTHONPATH=src python src/began/create_dataset.py "$CONVERTED_DIR"

Training BEGAN

PYTHONPATH=src python src/began/training.py

Images are generated in each epoch into generated/epXXX/ directory.

Training History

FYI: epoch time in training

About 680 sec/epoch

  • Linux

  • Dataset: All images aligned with deep funneling (13194 samples)

  • Intel(R) Core(TM) i7-7700K CPU @ 4.20GHz

  • GeForce GTX 1080

  • Environment Variables

    KERAS_BACKEND=theano
THEANO_FLAGS=device=gpu,floatX=float32,lib.cnmem=1.0

Generate Image

PYTHONPATH=src python src/began/generate_image.py

Generated images are outputted in generated/main/ directory.

Generated Image Examples

Epoch 1
Epoch 25
Epoch 50
Epoch 75
Epoch 100
Epoch 125
Epoch 150
Epoch 175
Epoch 200
Epoch 215

more filters or layers?

Memo

Trouble: Running with Theano 0.9.0 CPU mode on Linux

Theano 0.9.0 CPU mode on Linux seems to have memory leak problem. See: keras-team/keras#5935