Loss

Describes how the loss function is composed.

Options

• hinge: allow this much error before counting (per pixel)
• mae_multiplier: when calculating total loss multiply MAE loss by this multiplier
• mae_delta: if true add emphasis on the deltas magnitude this multiplier
• regularization: when calculating total loss multiple regularization loss by this multiplier
• pyramid: when calculating loss use a gaussian pyramid to get loss at different scales
  • levels: how many multiscale levels to build.
  • type: type of laplacian (laplacian, gaussian, gaussian_learnable, none)
  • xy_max: expansion of the normal distribution
  • kernel_size: size of each level's convolution kernel

Example

"loss": {
   "hinge": 2.5,
   "mae_multiplier": 1.0,
   "regularization": 0.01,
   "nae_multiplier": 10.0,
   "input_shape": ["?", "?", 3],
   "pyramid": {
     "levels": 4,
     "type": "gaussian",
     "xy_max": [1.0, 1.0],
     "kernel_size": [5, 5]
   }
}