search

jhoffman / cycada_release

Code to accompany ICML 2018 paper
BSD 2-Clause "Simplified" License
558 stars 126 forks source link

What's the identity loss? #23

Closed Luodian closed 2 years ago

Luodian commented 5 years ago

@In your code, there's an identity loss, but I didn't find this part in paper~

Is it useful in your final improvement?

https://github.com/jhoffman/pytorch-CycleGAN-and-pix2pix/blob/e484612d83449d05024a7d5fd2e012be53faad85/models/cycle_gan_semantic_model.py#L136

Luodian commented 5 years ago

@jhoffman Hi~Prof. Hoffman, I saw you didn't use the identity loss in your code, is it useful for semantic segmentation? May I know this lambda value in your final result model? image

vicdxxx commented 4 years ago

lambdaG is 10 times less than lambdaD， quote paper(Appendix): We weight the representation loss ten times less than the discriminator loss as a convienience since otherwise the discriminator did not learn a suitable model within a single epoch