HilaManor/Generative-deep-features

Issues

• Init noise for image withh/out correlation between channels

  #13 opened 3 years ago by HilaManor
  0

• לראות הרצאות אופטימייזר

  #22 opened 3 years ago by HilaManor
  0

• Research multi-scale pyramid

  #43 opened 3 years ago by dael963
  0

• In SinGAN code, check if there's a better reconstruction (content) loss than MSE

  #27 opened 3 years ago by dael963
  1

• Is replacing the image resize with diffrent CNN stride at each level is equivalent?

  #40 opened 3 years ago by dael963
  0

• check generation in latent space

  #36 opened 3 years ago by dael963
  0

• Figure out weights effect

  #4 opened 3 years ago by HilaManor
  1

• Figure out true convergence time

  #7 opened 3 years ago by HilaManor
  0

• See the result on a different noise

  #11 opened 3 years ago by HilaManor
  1

• Implement Contextual Loss

  #16 opened 3 years ago by HilaManor
  1

• Validate Neural Style Transfer

  #17 opened 3 years ago by HilaManor
  0

• Implement Wasserstein Distance

  #18 opened 3 years ago by HilaManor
  1

• דוח איפיון

  #21 opened 3 years ago by HilaManor
  0

• Try to change to allow fitting to different target-image resolutions

  #25 opened 3 years ago by HilaManor
  0

• deepcopy vgg - why do we need it?

  #30 opened 3 years ago by dael963
  0

• Noise amplitude - seems stronger than it should be

  #47 opened 3 years ago by dael963
  0

• Find optimal weighted combination of VGG layers for loss (pd\gram)

  #45 opened 3 years ago by dael963
  0

• Generated image is grained, with dot-like artifact at higher pyramid layers

  #46 opened 3 years ago by dael963
  0

• Change that the showed noise we be from a constant example noise for the concat generation

  #51 opened 3 years ago by HilaManor
  0

• Loss vs epoch is noisy (at least for pdl) - check lr, gsteps scheduler, alpha

  #48 opened 3 years ago by dael963
  0

• Add Sampling for the end of each scale

  #50 opened 3 years ago by HilaManor
  0

• training line 35-37 - Load the previous weights

  #39 opened 4 years ago by dael963
  1

• Consider averaging losses over amount of pixels

  #41 opened 4 years ago by dael963
  1

• VGG19 minumum images size is 224*224*3. is it ok to insert smaller image?

  #42 opened 4 years ago by dael963
  1

• Pixels loss - should we use rgb or hsv?

  #52 opened 4 years ago by dael963
  1

• Gsteps isn't working for values greater than 1

  #49 opened 4 years ago by dael963
  0

• Research diffrent CNN architectures for generators

  #44 opened 4 years ago by dael963
  0

• repeat tests with recon Loss (and different alpha for style loss)

  #35 opened 4 years ago by HilaManor
  0

• seperate "functions" module into sub-modules (such as imageprocess \ nnhelper)

  #38 opened 4 years ago by dael963
  0

• Implement more scales

  #37 opened 4 years ago by HilaManor
  0

• Train and then insert the Recon-Z

  #31 opened 4 years ago by HilaManor
  0

• Try without recon and a single Z - should be similar to before

  #32 opened 4 years ago by HilaManor
  1

• Save every 500 iter mid-image (same Z!!!)

  #34 opened 4 years ago by HilaManor
  0

• Input different loss graphs

  #33 opened 4 years ago by HilaManor
  0

• Add Scheduler / Change learning rate for Optimizer

  #14 opened 4 years ago by HilaManor
  0

• Randomize the input noise for the learning process

  #24 opened 4 years ago by HilaManor
  0

• Create normal config file

  #26 opened 4 years ago by HilaManor
  0

• Generate_noise - wht to upsample noise?

  #28 opened 4 years ago by dael963
  0

• in_s - what is it? seems to be 0 all the time

  #29 opened 4 years ago by dael963
  0

• Change to using one SinGAN scale (generator)

  #23 opened 4 years ago by HilaManor
  0

• Fix LabAdmin Issues

  #19 opened 4 years ago by HilaManor
  0

• Understand how to use docker

  #20 opened 4 years ago by HilaManor
  0

• Ask for computer from SIPL

  #15 opened 4 years ago by HilaManor
  0

• Define constant seed for the random image

  #6 opened 4 years ago by dael963
  0

• Training neural style transfer optmize an image, but it doesn't train weights. How do we generate new random image?

  #9 opened 4 years ago by dael963
  2

• Deploy environment

  #8 opened 4 years ago by HilaManor
  1

• Try to use zero images instead of white noise

  #10 opened 4 years ago by dael963
  0

• Add time

  #12 opened 4 years ago by HilaManor
  0

• Figure up what the diffrence between convX_1 and convX

  #3 opened 4 years ago by dael963
  1

• Should we normalized the noise before we insert it as input?

  #5 opened 4 years ago by dael963
  1