Skip to content

Latest commit

 

History

History

lapgan

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
gen_img
gen_img
 
 
model
model
 
 
lapgan_model.py
lapgan_model.py
 
 
lapgan_tb.png
lapgan_tb.png
 
 
lapgan_train.py
lapgan_train.py
 
 
readme.md
readme.md
 
 

readme.md

Laplacian Pyramid Generative Adversarial Networks

Loss Function

  • used sce loss for G/D nets.

Architecture Networks

  • Similar as LAPGAN paper.
DIFFS LAPGAN Paper ME
Weight initializer normal dist HE initializer
z noise 100 128
fc units D[600], G[1200] D[512], G[1024]
image size 28x28 for cifar-10 32x32 for cifar-10
Optimizer SGD Adam
LR 2e-2 8e-4
LR Decay exponential decay None
momentum .5 to .8 None

Momentum Increasing by 8e-4.
Learning Rate : 2e-2 factor = 1 + 4e-5
Normal Distribution Initializer : (µ = 0, σ = 0.02)
HE Initializer parameters : (factor = 1, FAN_AVG, uniform)

Tensorboard

result

Elapsed time : 2h 58m 50s with GTX 1060 6GB x 1

Result

Name Global Step 50k Global Step 100k Global Step 150k
LAPGAN img img img

To-Do