Model

Describes the type and characteristics of model.

Options

• filters: the number of filters per conv operations
• clip_values: if true clip output values to [-0.5, 0.5]
• no_layers: number of layers per level
• min_value: the minimum value
• max_value: the maximum value
• kernel_size: the kernel size of the convolution filter
• type: type of mode (resnet, sparse_resnet, gatenet)
• batchnorm: use batch normalization between layers
• activation: convolution activation
• shared_model: if true use a shared model across scales
• local_normalization: integer, if > 0 then applies a local normalization kernel before top level, if == 0 applies global normalization, if < 0 does nothing
• output_multiplier: multiply output with this value to avoid saturation before going to final_activation
• kernel_regularizer: kernel regularization (l1, l2, l1_l2)
• final_activation: final activation at the end of the model
• input_shape: the input shape (minus the batch)
• kernel_initializer: kernel initializer
• add_skip_with_input: if true skip input to the final activation so model learns the diff
• add_residual_between_models: if true the output of the previous scale is mixed with the input of the next scale
• pyramid:
  • 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

"model_denoise": {
    "filters": 16,
    "no_layers": 10,
    "min_value": 0,
    "add_var": false,
    "max_value": 255,
    "kernel_size": 5,
    "type": "resnet",
    "batchnorm": true,
    "activation": "elu",
    "clip_values": true,
    "shared_model": true,
    "output_multiplier": 1.0,
    "local_normalization": -1,
    "kernel_regularizer": "l1",
    "final_activation": "tanh",
    "add_skip_with_input": false,
    "add_residual_between_models": true,
    "input_shape": ["?", "?", 3],
    "kernel_initializer": "glorot_normal",
    "pyramid": {
      "levels": 3,
      "type": "laplacian",
      "xy_max": [1.0, 1.0],
      "kernel_size": [3, 3]
    }
}