search

Peilun-Li / SG-GAN

Semantic-aware Grad-GAN for Virtual-to-Real Urban Scene Adaption
https://arxiv.org/abs/1801.01726
60 stars 18 forks source link

pretrained model #1

Open dleam opened 5 years ago

dleam commented 5 years ago

Excellent job. Can you please provide a pretrained model for the further research?

Peilun-Li commented 5 years ago

Hi @dleam , thanks for your interest! Here it is: https://1drv.ms/u/s!AmVzKjduHCxdqSQrUlK7nxz2wvZj

dleam commented 5 years ago

Thanks for your quick reply. I will try it.

dleam commented 5 years ago

Hi @Peilun-Li , I have tried your model and the result is excellent. I have another question about your code after reading your paper.

self.DB_fake_sample = self.discriminator(
        self.fake_B_sample, self.mask_B_sample, self.options, reuse=True, name="discriminatorB")
self.DA_fake_sample = self.discriminator(
        self.fake_A_sample, self.mask_A_sample, self.options, reuse=True, name="discriminatorA")

Why do you multiply fake_B and mask_B (fake_A and mask_A) here? " To my understanding, the fake_B and (semantic)mask_B have different structures. How to understand this operation in your code? Thanks~

Peilun-Li commented 5 years ago

Hi @dleam , Per this feed_dict setting (https://github.com/Peilun-Li/SG-GAN/blob/master/model.py#L215-L231) you can find fake_B_sample is adapted from real_A, and mask_B_sample is actually mask_A. Since we wanna keep semantic information from being changed during the adaptation, fake_B_sampe and real_A should share the same semantic information, thus we are multiplying fake_B_sample with mask_A, i.e., mask_B_sample.

dleam commented 5 years ago

Thank you!