Operations serialization

This page describes the operation options.

Every operation necessary has _name and _type parameters. The rest of operation parameters depends on its type.

Conv2D

- _name: arbitrary operation name
  _type: conv2d        # fixed string
  kernel_size: 3       # size of convolution kernel
  input_channels: 3    # number of input feature channels
  output_channels: 16  # number of output feature channels
  stride: 2            # stride (defaults to 1)
  padding: valid       # paddling, string, "valid" or "same" (defaults to "valid")
  use_bias: true       # bias addition, "true" or "false" (defaults to "true")
  groups: 1            # number of groups for grouped convolution (defaults to 1)
  activation: default  # activation function

For activation functions see Activation functions serialization.

Dense

- _name: arbitrary operation name
  _type: dense      # fixed string
  use_bias: true    # bias addition, "true" or "false"

ImageSampler

- _name: arbitrary operation name
  _type: image_sampler  # fixed string
  output_width: 224     # resulting image width in pixels
  output_height: 224    # resulting image width in pixels
  center_crop: true     # center crop enabled, "true" or "false" (defaults to "true")
  linear_interp: true   # linear interpolation of input pixels enabled, "true" or "false" (defaults to "true")

Pooling2D

- _name: arbitrary operation name
  _type: pooling2d     # fixed string
  operator: max        # "max" or "average"
  size: 3              # pooling size
  stride: 2            # stride (defaults to 1)
  padding: valid       # paddling, string, "valid" or "same" (defaults to "valid")

Softmax

Softmax operation has no parameters (except its name and the type).

- _name: arbitrary operation name
  _type: softmax    # fixed string