Awesome

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

<table> <tr> <th>Epoch 1</th> <th><img src="example/v1/ep001/gen_000.jpg"></th> <th><img src="example/v1/ep001/gen_001.jpg"></th> <th><img src="example/v1/ep001/gen_002.jpg"></th> <th><img src="example/v1/ep001/gen_003.jpg"></th> <th><img src="example/v1/ep001/gen_004.jpg"></th> </tr> <tr> <th>Epoch 25</th> <th><img src="example/v1/ep025/gen_000.jpg"></th> <th><img src="example/v1/ep025/gen_001.jpg"></th> <th><img src="example/v1/ep025/gen_002.jpg"></th> <th><img src="example/v1/ep025/gen_003.jpg"></th> <th><img src="example/v1/ep025/gen_004.jpg"></th> </tr> <tr> <th>Epoch 50</th> <th><img src="example/v1/ep050/gen_000.jpg"></th> <th><img src="example/v1/ep050/gen_001.jpg"></th> <th><img src="example/v1/ep050/gen_002.jpg"></th> <th><img src="example/v1/ep050/gen_003.jpg"></th> <th><img src="example/v1/ep050/gen_004.jpg"></th> </tr> <tr> <th>Epoch 75</th> <th><img src="example/v1/ep075/gen_000.jpg"></th> <th><img src="example/v1/ep075/gen_001.jpg"></th> <th><img src="example/v1/ep075/gen_002.jpg"></th> <th><img src="example/v1/ep075/gen_003.jpg"></th> <th><img src="example/v1/ep075/gen_004.jpg"></th> </tr> <tr> <th>Epoch 100</th> <th><img src="example/v1/ep100/gen_000.jpg"></th> <th><img src="example/v1/ep100/gen_001.jpg"></th> <th><img src="example/v1/ep100/gen_002.jpg"></th> <th><img src="example/v1/ep100/gen_003.jpg"></th> <th><img src="example/v1/ep100/gen_004.jpg"></th> </tr> <tr> <th>Epoch 125</th> <th><img src="example/v1/ep125/gen_000.jpg"></th> <th><img src="example/v1/ep125/gen_001.jpg"></th> <th><img src="example/v1/ep125/gen_002.jpg"></th> <th><img src="example/v1/ep125/gen_003.jpg"></th> <th><img src="example/v1/ep125/gen_004.jpg"></th> </tr> <tr> <th>Epoch 150</th> <th><img src="example/v1/ep150/gen_000.jpg"></th> <th><img src="example/v1/ep150/gen_001.jpg"></th> <th><img src="example/v1/ep150/gen_002.jpg"></th> <th><img src="example/v1/ep150/gen_003.jpg"></th> <th><img src="example/v1/ep150/gen_004.jpg"></th> </tr> <tr> <th>Epoch 175</th> <th><img src="example/v1/ep175/gen_000.jpg"></th> <th><img src="example/v1/ep175/gen_001.jpg"></th> <th><img src="example/v1/ep175/gen_002.jpg"></th> <th><img src="example/v1/ep175/gen_003.jpg"></th> <th><img src="example/v1/ep175/gen_004.jpg"></th> </tr> <tr> <th>Epoch 200</th> <th><img src="example/v1/ep200/gen_000.jpg"></th> <th><img src="example/v1/ep200/gen_001.jpg"></th> <th><img src="example/v1/ep200/gen_002.jpg"></th> <th><img src="example/v1/ep200/gen_003.jpg"></th> <th><img src="example/v1/ep200/gen_004.jpg"></th> </tr> <tr> <th>Epoch 215</th> <th><img src="example/v1/ep215/gen_000.jpg"></th> <th><img src="example/v1/ep215/gen_001.jpg"></th> <th><img src="example/v1/ep215/gen_002.jpg"></th> <th><img src="example/v1/ep215/gen_003.jpg"></th> <th><img src="example/v1/ep215/gen_004.jpg"></th> </tr> </table>

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: https://github.com/fchollet/keras/issues/5935