ConstantArrayLayer

represents a layer that has no input and produces as output a constant array.

includes options for initial value of the array or output size.

Details and Options

ConstantArrayLayer is typically used in a NetGraph and allows a learned array to be used as an ordinary input to other layers. Without ConstantArrayLayer, all layer inputs must ultimately be derived from inputs to the entire net.
The array stored by ConstantArrayLayer will be learned during training.
The constant array stored in ConstantArrayLayer has the same dimensions as the output of ConstantArrayLayer.
ConstantArrayLayer["Array"->array] can be used to give an explicit value of the constant array.
When the constant array is not explicitly specified or is given as Automatic, it is added automatically when NetInitialize or NetTrain is used.
ConstantArrayLayer["Output"->{d₁,d₂,…}] can be used to fix the output size of the layer, which also fixes the size of the constant array.
The following training parameter can also be included:
LearningRateMultipliers Automatic learning rate multipliers for the array
ConstantArrayLayer[…][] returns the constant array.
The constant array can also be obtained from an initialized ConstantArrayLayer using NetExtract[layer,"Array"].
Options[ConstantArrayLayer] gives the list of default options to construct the layer. Options[ConstantArrayLayer[…]] gives the list of default options to evaluate the layer on some data.
Information[ConstantArrayLayer[…]] gives a report about the layer.
Information[ConstantArrayLayer[…],prop] gives the value of the property prop of ConstantArrayLayer[…]. Possible properties are the same as for NetGraph.

open allclose all

Create an initialized ConstantArrayLayer with output dimensions specified:

Extract the initialized array:

Create a ConstantArrayLayer with a specific initial value for the array:

Embed the layer in a graph that raises the input to a power of the array elementwise:

ConstantArrayLayer cannot accept symbolic arrays:

Introduced in 2017

(11.1)

Updated in 2020

(12.1)

Top