TensorFlow 1.0 compatible implementation
It currently works well on LSUN for vanilla gan loss, and converges slow for wgan loss (too small learning rate) 😅 .
To run WGAN, set
d_learningrate=5e-5
g_learningrate=5e-5
Wloss=True
Adam=False
noise='normal'
and
f_h=4, f_w=4, Cc=64
in both generator and discriminator (same setting as the original torch implementation).
After 1.5 Epoch with Batch Normalization:
After 1 Epoch without Batch Normalization (but with bias; if no bias, the generated image is still sharp but less light):
After 4 Epochs with Batch Normalization:
