search

tamarott / SinGAN

Official pytorch implementation of the paper: "SinGAN: Learning a Generative Model from a Single Natural Image"
https://tamarott.github.io/SinGAN.htm
Other
3.3k stars 610 forks source link

question about the loss #69

Open XiaoHao-Chen opened 4 years ago

XiaoHao-Chen commented 4 years ago

Why is the output of patch-d the average value of all patches, rather than the cross entropy of biclassification calculated by BCELoss, like the common patch-gan?

tamarott commented 4 years ago

This is a WGAN, so no cross-entropy loss is used. You can find more details regarding the training scheme of WGAN at the paper "Improved Training of Wasserstein GANs" by Gulrajani et. al: https://arxiv.org/pdf/1704.00028.pdf (Generally, we follow this)