search

HSinger04 / VOGUE-Reimplementation

A reimplementation of the VOGUE paper for the IANNwTF WS20/21 course at the University of Osnabrück
1 stars 0 forks source link

Perceptual loss #9

Closed HSinger04 closed 3 years ago

HSinger04 commented 3 years ago
  1. Just try out L2-norm between unit-scaled activations of last layer between two networks or sth.
HSinger04 commented 3 years ago

VGG-16:

  1. Use conv1_2, conv2_2, ..., conv5_2 (See: https://rn-unison.github.io/articulos/style_transfer.pdf, https://arxiv.org/pdf/1707.09405.pdf
  2. Distance calculation (https://arxiv.org/pdf/1801.03924.pdf, page 6):
  3. Unit-normalize in channel (i.e. the values in channel c get between -1 and 1)
  4. Follow eq. 1 from https://arxiv.org/pdf/1801.03924.pdf, page 6: ignore w_l, just take l2 norm of channelvectors over height and width