CycleGan about keras-gan HOT 7 CLOSED

In the cycle-gan keras code, why the 'combine model' required using the 'identity mapping' as a loss (as following):

'Identity mapping of images'
img_A_id = self.g_BA(img_A)
img_B_id = self.g_AB(img_B)

Anyone have idea? In the original paper (Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks), they do not have this terms in their loss. Am I correct?

from keras-gan.

@parasjain-12 The dataloader can't find images in the path that you specified in that case. So np.random.random([]) will throw an errow.

@ubersexualShupeng The identity loss in from CycleGAN means that an image from domain A that is translated to domain A should exactly
match the original image. In this case images from A translated to domain A and images from domain B
translated to domain B should reproduce the original images, just as the implementation does. It is used in the paper for painting-to-photograph translation problems.

from keras-gan.

@eriklindernoren Thank you for your kind response! I understand that the 'reconstruct image' should be matched to the 'real image', as the following two lines code:

Translate images back to original domain
reconstr_A = self.g_BA(fake_B)
reconstr_B = self.g_AB(fake_A)

However, I quite confuse with another two line codes as follows:
Identity mapping of images
img_A_id = self.g_BA(img_A)
img_B_id = self.g_AB(img_B)

Why the model loss required both of them?
Thanks.

from keras-gan.

@ubersexualShupeng This is explained in detail in the section Photo generation from paintings in the paper (https://arxiv.org/pdf/1703.10593.pdf).

from keras-gan.

@eriklindernoren OK. Got it. Thank you!

from keras-gan.

I run this code and obtained good results on different datasets but I want to know something
about quantitative parameters we obtained after the end of code.

Acc, adv: recon, and iD? Can you please explain them

from keras-gan.

Hi, erik.

May I ask how to save the trained model of cyclegan?

I tried follow the codes of acgan.py but still failed to save the generate model and the discrimitor model for now.

from keras-gan.