dag2dlnetwork

Convert SeriesNetwork and DAGNetwork to dlnetwork

Since R2024a

Syntax

dlnet = dag2dlnetwork(net)

Description

dlnet = dag2dlnetwork(net) converts the specified SeriesNetwork or DAGNetwork object to a dlnetwork object.

Examples

Convert SeriesNetwork Object to dlnetwork

Load a trained SeriesNetwork object into the workspace and inspect it.

load digitsSeriesNet
net

net = 
  SeriesNetwork with properties:

         Layers: [15×1 nnet.cnn.layer.Layer]
     InputNames: {'imageinput'}
    OutputNames: {'classoutput'}

Convert the network to a dlnetwork object.

dlnet = dag2dlnetwork(net)

dlnet = 
  dlnetwork with properties:

         Layers: [14×1 nnet.cnn.layer.Layer]
    Connections: [13×2 table]
     Learnables: [14×3 table]
          State: [6×3 table]
     InputNames: {'imageinput'}
    OutputNames: {'softmax'}
    Initialized: 1

  View summary with summary.

Input Arguments

net — Input network

SeriesNetwork object | DAGNetwork object

Input network, specified as a SeriesNetwork or DAGNetwork object.

Output Arguments

dlnet — Converted network

dlnetwork object

Converted network, returned as a dlnetwork object.

Algorithms

dlnetwork Conversion

The dag2dlnetwork function applies these adjustments to the input network to make it compatible with dlnetwork functionality:

• Remove the output layers — To specify the loss function when you train a neural network, use the trainnet function.

• Replace fully connected layers with convolution layers — For neural networks that require fully connected layers to output data with singleton spatial dimensions, the dag2dlnetwork function replaces the fully connected layer with the equivalent convolutional layer. Otherwise, the dag2dlnetwork function does not replace the layer. In the output dlnetwork object, the fully connected layer outputs data without spatial dimensions.

• For networks that output singleton spatial dimensions, include a flatten layer — For neural networks that output data with singleton spatial dimensions (for example, a SqueezeNet neural network), the dag2dlnetwork function adds a flatten layer at the end of network that removes the singleton spatial dimensions.

• Return an uninitialized network when they contains unsupported layers — If the network contains layers that dlnetwork objects do not support, then the function returns an uninitialized dlnetwork object. To remove or replace unsupported layers in a dlnetwork object, use the removeLayers and replaceLayer functions, respectively.

Version History

Introduced in R2024a