generateConfig.lua

-- ==CONFIGURATION FILE FOR IMAGE GENERATION==
local option = ... -- the argument is received as varargs, hence the ...
local commonParameters

assert(option, 'Option not specified.')

if option == 0 then
-- Generate images from scratch using the generator only, no encoder involved (generation/generate.lua)
opt = {
    batchSize = 1000,         -- number of samples to produce (it should be multiple of Y attribute vector length (ny))
    noisetype = 'normal',     -- type of noise distribution (uniform / normal).
    net = 'checkpoints/',     -- path to the generator network
    imsize = 1,               -- used to produce larger images. 1 = 64px. 2 = 80px, 3 = 96px, ...
    noisemode = 'random',     -- random / line / linefull1d / linefull
    name = 'gen_samples',     -- name of the file saved
    gpu = 1,                  -- gpu mode. 0 = CPU, 1 = GPU
    display = 1,              -- Display image: 0 = false, 1 = true
    nz = 100,                 -- Z latent vector length 
    -- Conditional GAN parameters
    dataset = 'celebA'        -- dataset specification: mnist | celebA. It is necessary to know how to sample Y.          
}

else

  -- Commom parameters for options 1 to 3
  commonParameters = {
      decNet = 'checkpoints/',  -- path to the generator network
      encZnet = 'checkpoints/', -- path to encoder Z network
      encYnet = 'checkpoints/', -- path to encoder Y network
      loadSize = {3, 64, 64},   -- image dimensions CxHxW  used as input (output) by the encoders (generator).
      gpu = 1,                  -- gpu mode. 0 = CPU, 1 = GPU
      nz = 100,                 -- Z latent vector length 
  }
  
  if option == 1 then
  -- Reconstruct images with encoder + generator and obtain variations on them (generation/reconstructWithVariations.lua)
  opt = {
      nImages = 50,             -- number of samples to produce (only valid if loadOption != 1)
      loadOption = 2,           -- 1 = load input image, 2 = load multiple input images
      loadPath = '', -- loadOption == 1: path to single image, loadOption==2: path to folder with images
      name = 'encoder_disentangle',
      -- Conditional GAN parameters
      dataset = 'celebA',       -- dataset specification: mnist | celebA. It is necessary to know how to sample Y. 
      threshold = true,         -- (celebA only) true= threshold original encoded Y to binary 
  }
  
  elseif option == 2 then
  -- Attribute transfer: given two images, swap their attribute information Y (generation/attributeTransfer.lua)
  opt = {
      im1Path = '', -- path to image 1
      im2Path = '', -- path to image 2
  }
  
  elseif option == 3 then
  -- Interpolate between two input images (generation/interpolate.lua)
  opt = {
      im1Path = '', -- path to image 1
      im2Path = '', -- path to image 2
      nInterpolations = 4,
  }
  
  else
  
    error("Option not recognized.") 
  
  end
  
  -- Merge tables
  for k,v in pairs(commonParameters) do opt[k] = v end

end