Home

Awesome

BEGAN: Boundary Equilibrium Generative Adversarial Networks

Implementation of Google Brain's BEGAN: Boundary Equilibrium Generative Adversarial Networks in Tensorflow.
BEGAN is the state of the art when it comes to generate realistic faces.

<p> <img src="Result/128x128.bmp" width="500" height="500" /> <img src="Result/64x64.bmp" width="250" height="250"/> </p>

Figure1a. 128x128 img and 64x64 img. 128x128 img is very impressive. You can see SET OF TEETH

<p> <img src="Result/gamma_0.3.bmp" width="250" height="250" /> <img src="Result/gamma_0.4.bmp" width="250" height="250" /> <img src="Result/gamma_0.5.bmp" width="250" height="250" /> </p>

Figure1b. This is random result from my train model. From gamma 0.3 to 0.5. No cherry picking. gamma 0.3, nice but bias to women's face. gamma 0.4, Best. gamma 0.5, good texture but hole problem.

<p> <img src="Result/result.gif" width="300" height="300" /> </p>

Figure1c. From scratch to 200k iter

Implementation detail

This train model is 64x64. 128x128 will be update. Different with original paper is train loss update method, learning rate decay. First, paper's loss update way is Loss_G and Loss_D simultaneously. But when I tried that way, models are mode collapse. So, This code use altenative way. Second, learning rate decay is 0.95 every 2000 iter. This parameter is just train experienc. You can change or see the paper.

Train progress

If you want to see the train progress download this dropbox folder and run "tensorboard --logdir='./'". I uploaded two trained model(64x64 and 128x128)

<p> <img src="Result/kt.jpg" width="810" height="270" /> </p>

Figure2. Kt graph. When you train model, reference this result. It doesn't reach to 1.0. In my case, it's converge to 0.08

<p> <img src="Result/m_global.jpg" width="810" height="270" /> </p>

Figure3. Convergence measure(M_global). Similar with paper's graph

<p> <img src="Result/gamma_0.4.bmp" width="270" height="270" /> <img src="Result/decoder.bmp" width="270" height="270" /> </p>

Figure4. Compare with Generator output and Decoder output.

Usage

Recommend to download trained model this dropbox folder.

Make Train Data

  1. Download celebA dataset (img_align_celeba.zip) and unzip to 'Data/celeba/raw'
  2. Run ' python ./Data/celeba/face_detect.py '

Train (refer the main.py began_cmd)

ex) 64x64 img | Nz,Nh 128 | gamma 0.4
python3 main.py -f 1 -p "began" -trd "celeba" -tro "crop" -trs 64  -z 128 -em 128 -fn 64  -b 16 -lr 1e-4 -gm 0.4 -g "0"

ex) 128x128 img | Nz,Nh 64 | gamma 0.7
python3 main.py -f 1 -p "began" -trd "celeba" -tro "crop" -trs 128 -z 64  -em 64  -fn 128 -b 16 -lr 1e-4 -gm 0.7 -g "0"

Test (refer the main.py and began_cmd)

ex) 64x64 img | Nz,Nh 128 | gamma 0.4
python3 main.py -f 0 -p "began" -trd "celeba" -tro "crop" -trs 64  -z 128 -em 128 -fn 64  -b 16 -lr 1e-4 -gm 0.4 -g "0"

ex) 128x128 img | Nz,Nh 64 | gamma 0.7
python3 main.py -f 0 -p "began" -trd "celeba" -tro "crop" -trs 128 -z 64  -em 64  -fn 128 -b 16 -lr 1e-4 -gm 0.7 -g "0"

Requirements

Author

Heumi / ckhfight@gmail.com