5,321
edits
SinGAN: Learning a Generative Model from a Single Natural Image: Difference between revisions
SinGAN: Learning a Generative Model from a Single Natural Image (view source)
Revision as of 16:24, 14 January 2020
, 14 January 2020→Architecture
| Line 26: | Line 26: | ||
==Architecture== | ==Architecture== | ||
They build <math>N</math> | They build <math>N</math> PatchGANs, N usually 7-8<br> | ||
Each GAN <math>G_n</math> adds details to the image produced by GAN <math>G_{n+1}</math> below it.<br> | Each GAN <math>G_n</math> adds details to the image produced by GAN <math>G_{n+1}</math> below it.<br> | ||
The final GAN <math>G_0</math> adds only fine details. | The final GAN <math>G_0</math> adds only fine details. | ||
===Generator=== | ===Generator=== | ||
The use N generators which they call a hierarchy of patch-GANs.<br> | The use N generators which they call a hierarchy of patch-GANs.<br> | ||
| Line 77: | Line 78: | ||
The architecture is the same as the generator.<br> | The architecture is the same as the generator.<br> | ||
The patch size is <math>11 \times 11</math> | The patch size is <math>11 \times 11</math> | ||
The GAN used is PatchGAN from [https://arxiv.org/abs/1611.07004 pix2pix].<br> | |||
PatchGAN's discriminator is referred to as a Markovian discriminator because the receptive field is smaller than the size of the image.<br> | |||
"Such a discriminator effectively models the image as a Markov random field, assuming independence between pixels separated by more than a patch diameter." | |||
==Training and Loss Function== | ==Training and Loss Function== | ||