about the "is_training" argument...

[johnnyasd12]

Hi, I wonder how is is_training argument work? In the code, it seems it's for setting batch_norm layer behavior, but after training, if I need to generate image with is_training=False, then it generate bad results.... It's weird that if I set is_training = True then I can get better results.

UPDATE: the generated result will be bad when i only reconstruct one image when is_training = True, if i reconstruct many images (say 100 images) then it will be better result.

Did I misunderstand something?

Thank you!