layerGraph

Graph of network layers for deep learning

Description

A layer graph specifies the architecture of a deep learning network with a more complex graph structure in which layers can have inputs from multiple layers and outputs to multiple layers. Networks with this structure are called directed acyclic graph (DAG) networks. After you create a layerGraph object, you can use the object functions to plot the graph and modify it by adding, removing, connecting, and disconnecting layers. To train the network, use the layer graph as input to the trainNetwork function or convert it to a dlnetwork and train it using a custom training loop.

Creation

Syntax

lgraph = layerGraph

lgraph = layerGraph(layers)

lgraph = layerGraph(net)

Description

lgraph = layerGraph creates an empty layer graph that contains no layers. You can add layers to the empty graph by using the addLayers function.

lgraph = layerGraph(layers) creates a layer graph from an array of network layers and sets the Layers property. The layers in lgraph are connected in the same sequential order as in layers.

lgraph = layerGraph(net) extracts the layer graph of a SeriesNetwork, DAGNetwork, or dlnetwork object. For example, you can extract the layer graph of a pretrained network to perform transfer learning.

Input Arguments

net — Deep learning network

SeriesNetwork object | DAGNetwork object | dlnetwork object

Deep learning network, specified as a SeriesNetwork, DAGNetwork, or dlnetwork object.

Properties

Layers — Network layers

Layer array

This property is read-only.

Network layers, specified as a Layer array.

Connections — Layer connections

table

This property is read-only.

Layer connections, specified as a table with two columns.

Each table row represents a connection in the layer graph. The first column, Source, specifies the source of each connection. The second column, Destination, specifies the destination of each connection. The connection sources and destinations are either layer names or have the form 'layerName/IOName', where 'IOName' is the name of the layer input or output.

Data Types: table

InputNames — Network input layer names

cell array of character vectors

This property is read-only.

Network input layer names, specified as a cell array of character vectors.

Data Types: cell

OutputNames — Network output layer names

cell array

Network output layer names, specified as a cell array of character vectors.

Data Types: cell

Object Functions

addLayers	Add layers to layer graph or network
removeLayers	Remove layers from layer graph or network
replaceLayer	Replace layer in layer graph or network
connectLayers	Connect layers in layer graph or network
disconnectLayers	Disconnect layers in layer graph or network
plot	Plot neural network architecture

Examples

Add Layers to Layer Graph

Create an empty layer graph and an array of layers. Add the layers to the layer graph and plot the graph. addLayers connects the layers sequentially.

lgraph = layerGraph;

layers = [
    imageInputLayer([32 32 3],'Name','input')  
    convolution2dLayer(3,16,'Padding','same','Name','conv_1')
    batchNormalizationLayer('Name','BN_1')
    reluLayer('Name','relu_1')];

lgraph = addLayers(lgraph,layers);
figure
plot(lgraph)

Create Layer Graph from an Array of Layers

Create an array of layers.

layers = [
    imageInputLayer([28 28 1],'Name','input')  
    convolution2dLayer(3,16,'Padding','same','Name','conv_1')
    batchNormalizationLayer('Name','BN_1')
    reluLayer('Name','relu_1')];

Create a layer graph from the layer array. layerGraph connects all the layers in layers sequentially. Plot the layer graph.

lgraph = layerGraph(layers);
figure
plot(lgraph)

Extract Layer Graph of DAG Network

Load a pretrained SqueezeNet network. You can use this trained network for classification and prediction.

net = squeezenet;

To modify the network structure, first extract the structure of the DAG network by using layerGraph. You can then use the object functions of LayerGraph to modify the network architecture.

lgraph = layerGraph(net)

lgraph = 
  LayerGraph with properties:

     InputNames: {'data'}
    OutputNames: {'ClassificationLayer_predictions'}
         Layers: [68x1 nnet.cnn.layer.Layer]
    Connections: [75x2 table]

Create Simple DAG Network

Create a simple directed acyclic graph (DAG) network for deep learning. Train the network to classify images of digits. The simple network in this example consists of:

• A main branch with layers connected sequentially.

• A shortcut connection containing a single 1-by-1 convolutional layer. Shortcut connections enable the parameter gradients to flow more easily from the output layer to the earlier layers of the network.

Create the main branch of the network as a layer array. The addition layer sums multiple inputs element-wise. Specify the number of inputs for the addition layer to sum. To easily add connections later, specify names for the first ReLU layer and the addition layer.

layers = [
    imageInputLayer([28 28 1])
    
    convolution2dLayer(5,16,'Padding','same')
    batchNormalizationLayer
    reluLayer('Name','relu_1')
    
    convolution2dLayer(3,32,'Padding','same','Stride',2)
    batchNormalizationLayer
    reluLayer
    convolution2dLayer(3,32,'Padding','same')
    batchNormalizationLayer
    reluLayer
    
    additionLayer(2,'Name','add')
    
    averagePooling2dLayer(2,'Stride',2)
    fullyConnectedLayer(10)
    softmaxLayer
    classificationLayer];

Create a layer graph from the layer array. layerGraph connects all the layers in layers sequentially. Plot the layer graph.

lgraph = layerGraph(layers);
figure
plot(lgraph)

Create the 1-by-1 convolutional layer and add it to the layer graph. Specify the number of convolutional filters and the stride so that the activation size matches the activation size of the third ReLU layer. This arrangement enables the addition layer to add the outputs of the third ReLU layer and the 1-by-1 convolutional layer. To check that the layer is in the graph, plot the layer graph.

skipConv = convolution2dLayer(1,32,'Stride',2,'Name','skipConv');
lgraph = addLayers(lgraph,skipConv);
figure
plot(lgraph)

Create the shortcut connection from the 'relu_1' layer to the 'add' layer. Because you specified two as the number of inputs to the addition layer when you created it, the layer has two inputs named 'in1' and 'in2'. The third ReLU layer is already connected to the 'in1' input. Connect the 'relu_1' layer to the 'skipConv' layer and the 'skipConv' layer to the 'in2' input of the 'add' layer. The addition layer now sums the outputs of the third ReLU layer and the 'skipConv' layer. To check that the layers are connected correctly, plot the layer graph.

lgraph = connectLayers(lgraph,'relu_1','skipConv');
lgraph = connectLayers(lgraph,'skipConv','add/in2');
figure
plot(lgraph);

Load the training and validation data, which consists of 28-by-28 grayscale images of digits.

[XTrain,YTrain] = digitTrain4DArrayData;
[XValidation,YValidation] = digitTest4DArrayData;

Specify training options and train the network. trainNetwork validates the network using the validation data every ValidationFrequency iterations.

options = trainingOptions('sgdm', ...
    'MaxEpochs',8, ...
    'Shuffle','every-epoch', ...
    'ValidationData',{XValidation,YValidation}, ...
    'ValidationFrequency',30, ...
    'Verbose',false, ...
    'Plots','training-progress');
net = trainNetwork(XTrain,YTrain,lgraph,options);

Display the properties of the trained network. The network is a DAGNetwork object.

net

net = 
  DAGNetwork with properties:

         Layers: [16x1 nnet.cnn.layer.Layer]
    Connections: [16x2 table]
     InputNames: {'imageinput'}
    OutputNames: {'classoutput'}

Classify the validation images and calculate the accuracy. The network is very accurate.

YPredicted = classify(net,XValidation);
accuracy = mean(YPredicted == YValidation)

accuracy = 0.9934

Limitations

• Layer graph objects contain no quantization information. Extracting the layer graph from a quantized network and then reassembling the network using assembleNetwork or dlnetwork removes quantization information from the network.

Version History

Introduced in R2017b