Download and install the Deep Learning Toolbox Converter for TensorFlow Models support package.

TypeimportKerasLayersat the command line.

importKerasLayers

If the Deep Learning Toolbox Converter for TensorFlow Models support package is not installed, then the function provides a link to the required support package in the Add-On Explorer. To install the support package, click the link, and then clickInstall. Check that the installation is successful by importing the layers from the model file'digitsDAGnet.h5'at the command line. If the required support package is installed, then the function returns aLayerGraphobject.

modelfile ='digitsDAGnet.h5'; net = importKerasLayers(modelfile)

net = LayerGraph with properties: Layers: [13x1 nnet.cnn.layer.Layer] Connections: [13x2 table] InputNames: {'input_1'} OutputNames: {'ClassificationLayer_activation_1'}

Import the network layers from the model filedigitsDAGnet.h5.

modelfile ='digitsDAGnet.h5'; layers = importKerasLayers(modelfile)

layers = LayerGraph with properties: Layers: [13x1 nnet.cnn.layer.Layer] Connections: [13x2 table] InputNames: {'input_1'} OutputNames: {'ClassificationLayer_activation_1'}

Plot the network architecture.

plot(layers)

Specify the network file to import.

modelfile ='digitsDAGnet.h5';

Import network layers.

layers = importKerasLayers(modelfile)

layers = LayerGraph with properties: Layers: [13x1 nnet.cnn.layer.Layer] Connections: [13x2 table] InputNames: {'input_1'} OutputNames: {'ClassificationLayer_activation_1'}

Load a data set for training a classifier to recognize new digits.

folder = fullfile(toolboxdir('nnet'),'nndemos','nndatasets','DigitDataset'); imds = imageDatastore(folder,...'IncludeSubfolders',true,...'LabelSource','foldernames');

Partition the dataset into training and test sets.

numTrainFiles = 750; [imdsTrain,imdsTest] = splitEachLabel(imds,numTrainFiles,'randomize');

Set the training options.

options = trainingOptions('sgdm',...'MaxEpochs',10,...'InitialLearnRate',0.001);

Train network using training data.

net = trainNetwork(imdsTrain,layers,options);

Training on single CPU. |========================================================================================| | Epoch | Iteration | Time Elapsed | Mini-batch | Mini-batch | Base Learning | | | | (hh:mm:ss) | Accuracy | Loss | Rate | |========================================================================================| | 1 | 1 | 00:00:01 | 15.62% | 12.6982 | 0.0010 | | 1 | 50 | 00:00:17 | 63.28% | 1.2109 | 0.0010 | | 2 | 100 | 00:00:34 | 85.16% | 0.4193 | 0.0010 | | 3 | 150 | 00:00:51 | 96.88% | 0.1749 | 0.0010 | | 4 | 200 | 00:01:06 | 99.22% | 0.0457 | 0.0010 | | 5 | 250 | 00:01:19 | 100.00% | 0.0374 | 0.0010 | | 6 | 300 | 00:01:34 | 96.88% | 0.1223 | 0.0010 | | 7 | 350 | 00:01:50 | 100.00% | 0.0087 | 0.0010 | | 7 | 400 | 00:02:04 | 100.00% | 0.0166 | 0.0010 | | 8 | 450 | 00:02:20 | 100.00% | 0.0098 | 0.0010 | | 9 | 500 | 00:02:38 | 100.00% | 0.0047 | 0.0010 | | 10 | 550 | 00:03:03 | 100.00% | 0.0031 | 0.0010 | | 10 | 580 | 00:03:13 | 100.00% | 0.0059 | 0.0010 | |========================================================================================| Training finished: Max epochs completed.

Run the trained network on the test set that was not used to train the network and predict the image labels (digits).

YPred = classify(net,imdsTest); YTest = imdsTest.Labels;

Calculate the accuracy.

accuracy = sum(YPred == YTest)/numel(YTest)

accuracy = 0.9856

Specify the network file to import layers and weights from.

modelfile ='digitsDAGnet.h5';

Import the network architecture and weights from the files you specified. To import the layer weights, specify'ImportWeights'to betrue. The function also imports the layers with their weights from the same HDF5 file.

layers = importKerasLayers(modelfile,'ImportWeights',true)

layers = LayerGraph with properties: Layers: [13x1 nnet.cnn.layer.Layer] Connections: [13x2 table] InputNames: {'input_1'} OutputNames: {'ClassificationLayer_activation_1'}

View the size of the weights in the second layer.

weights = layers.Layers(2).Weights; size(weights)

ans =1×47 7 1 20

The function has imported the weights so the layer weights are non-empty.

Specify the network file to import layers from and the file containing weights.

modelfile ='digitsDAGnet.json'; weights ='digitsDAGnet.weights.h5';

Import the network architecture and weights from the files you specified. The .json file does not include an output layer. Specify the output layer, so that importKerasLayers adds an output layer at the end of the networks architecture.

layers = importKerasLayers(modelfile,...'ImportWeights',true,...'WeightFile',weights,...'OutputLayerType','classification')

layers = LayerGraph with properties: Layers: [13x1 nnet.cnn.layer.Layer] Connections: [13x2 table] InputNames: {'input_1'} OutputNames: {'ClassificationLayer_activation_1'}

This example shows how to import the layers from a pretrained Keras network, replace the unsupported layers with custom layers, and assemble the layers into a network ready for prediction.

filename ='digitsDAGnetwithnoise.h5'; lgraph = importKerasLayers(filename,'ImportWeights',true);

Warning: Unable to import some Keras layers, because they are not supported by the Deep Learning Toolbox. They have been replaced by placeholder layers. To find these layers, call the function findPlaceholderLayers on the returned object.

figure plot(lgraph) title("Imported Network")

placeholderLayers = findPlaceholderLayers(lgraph)

placeholderLayers = 2x1 PlaceholderLayer array with layers: 1 'gaussian_noise_1' PLACEHOLDER LAYER Placeholder for 'GaussianNoise' Keras layer 2 'gaussian_noise_2' PLACEHOLDER LAYER Placeholder for 'GaussianNoise' Keras layer

placeholderLayers.KerasConfiguration

ans =struct with fields:trainable: 1 name: 'gaussian_noise_1' stddev: 1.5000

ans =struct with fields:trainable: 1 name: 'gaussian_noise_2' stddev: 0.7000

gnLayer1 = gaussianNoiseLayer(1.5,'new_gaussian_noise_1'); gnLayer2 = gaussianNoiseLayer(0.7,'new_gaussian_noise_2');

lgraph = replaceLayer(lgraph,'gaussian_noise_1',gnLayer1); lgraph = replaceLayer(lgraph,'gaussian_noise_2',gnLayer2);

figure plot(lgraph) title("Network with Replaced Layers")

lgraph.Layers

ans = 15x1 Layer array with layers: 1 'input_1' Image Input 28x28x1 images 2 'conv2d_1' Convolution 20 7x7x1 convolutions with stride [1 1] and padding 'same' 3 'conv2d_1_relu' ReLU ReLU 4 'conv2d_2' Convolution 20 3x3x1 convolutions with stride [1 1] and padding 'same' 5 'conv2d_2_relu' ReLU ReLU 6 'new_gaussian_noise_1' Gaussian Noise Gaussian noise with standard deviation 1.5 7 'new_gaussian_noise_2' Gaussian Noise Gaussian noise with standard deviation 0.7 8 'max_pooling2d_1' Max Pooling 2x2 max pooling with stride [2 2] and padding 'same' 9 'max_pooling2d_2' Max Pooling 2x2 max pooling with stride [2 2] and padding 'same' 10 'flatten_1' Keras Flatten Flatten activations into 1-D assuming C-style (row-major) order 11 'flatten_2' Keras Flatten Flatten activations into 1-D assuming C-style (row-major) order 12 'concatenate_1' Depth concatenation Depth concatenation of 2 inputs 13 'dense_1' Fully Connected 10 fully connected layer 14 'activation_1' Softmax softmax 15 'ClassificationLayer_activation_1' Classification Output crossentropyex

cLayer = lgraph.Layers(end)

cLayer = ClassificationOutputLayer with properties: Name: 'ClassificationLayer_activation_1' Classes: 'auto' ClassWeights: 'none' OutputSize: 'auto' Hyperparameters LossFunction: 'crossentropyex'

cLayer.Classes = string(0:9)

cLayer = ClassificationOutputLayer with properties: Name: 'ClassificationLayer_activation_1' Classes: [0 1 2 3 4 5 6 7 8 9] ClassWeights: 'none' OutputSize: 10 Hyperparameters LossFunction: 'crossentropyex'

lgraph = replaceLayer(lgraph,'ClassificationLayer_activation_1',cLayer);

net = assembleNetwork(lgraph)

net = DAGNetwork with properties: Layers: [15x1 nnet.cnn.layer.Layer] Connections: [15x2 table] InputNames: {'input_1'} OutputNames: {'ClassificationLayer_activation_1'}

Import layers from a Keras network that has parametric rectified linear unit (PReLU) layers.

A PReLU layer performs a threshold operation, where for each channel, any input value less than zero is multiplied by a scalar. The PReLU operation is given by

where ${\mathit{x}}_{\mathit{i}}$is the input of the nonlinear activation $\mathit{f}$on channel $$i$$, and $$a_i$$is the scaling parameter controlling the slope of the negative part. The subscript $$i$$in $$a_i$$indicates that the parameter can be a vector and the nonlinear activation can vary on different channels.

importKerasNetworkandimportKerasLayerscan import a network that includes PReLU layers. These functions support both scalar-valued and vector-valued scaling parameters. If a scaling parameter is a vector, then the functions replace the vector with the average of the vector elements. You can modify a PReLU layer to have a vector-valued scaling parameter after import.

Specify the network file to import.

modelfile =“digitsDAGnetwithPReLU.h5';

digitsDAGnetwithPReLUincludes two PReLU layers. One has a scalar-valued scaling parameter, and the other has a vector-valued scaling parameter.

Import the network architecture and weights frommodelfile.

layers = importKerasLayers(modelfile,'ImportWeights',true);

Warning: Layer 'p_re_lu_1' is a PReLU layer with a vector-valued parameter. The function replaces the parameter with the average of the vector elements. You can change the parameter back to a vector after import.

TheimportKerasLayersfunction displays a warning for the PReLu layerp_re_lu_1. The function replaces the vector-valued scaling parameter ofp_re_lu_1with the average of the vector elements. You can change the parameter back to a vector. First, find the index of the PReLU layer by viewing theLayersproperty.

layers.Layers

ans = 13x1 Layer array with layers: 1 'input_1' Image Input 28x28x1 images 2 'conv2d_1' Convolution 20 7x7x1 convolutions with stride [1 1] and padding 'same' 3 'conv2d_2' Convolution 20 3x3x1 convolutions with stride [1 1] and padding 'same' 4 'p_re_lu_1' PReLU PReLU layer 5 'p_re_lu_2' PReLU PReLU layer 6 'max_pooling2d_1' Max Pooling 2x2 max pooling with stride [2 2] and padding 'same' 7 'max_pooling2d_2' Max Pooling 2x2 max pooling with stride [2 2] and padding 'same' 8 'flatten_1' Keras Flatten Flatten activations into 1-D assuming C-style (row-major) order 9 'flatten_2' Keras Flatten Flatten activations into 1-D assuming C-style (row-major) order 10 'concatenate_1' Depth concatenation Depth concatenation of 2 inputs 11 'dense_1' Fully Connected 10 fully connected layer 12 'dense_1_softmax' Softmax softmax 13 'ClassificationLayer_dense_1' Classification Output crossentropyex

layershas two PReLU layers. Extract the fourth layerp_re_lu_1,最初向量值扩展标准ameter for a channel dimension.

tempLayer = layers.Layers(4)

tempLayer = PreluLayer with properties: Name: 'p_re_lu_1' RawAlpha: [20x1 single] Learnable Parameters Alpha: 0.0044 State Parameters No properties. Show all properties

TheRawAlphaproperty contains the vector-valued scaling parameter, and theAlphaproperty contains a scalar that is an element average of the vector values. ReshapeRawAlphato place the vector values in the third dimension, which corresponds to the channel dimension. Then, replaceAlphawith the reshapedRawAlphavalues.

tempLayer.Alpha = reshape(tempLayer.RawAlpha,[1,1,numel(tempLayer.RawAlpha)])

tempLayer = PreluLayer with properties: Name: 'p_re_lu_1' RawAlpha: [20x1 single] Learnable Parameters Alpha: [1x1x20 single] State Parameters No properties. Show all properties

Replace thep_re_lu_1layer inlayerswithtempLayer.

layers = replaceLayer(layers,'p_re_lu_1', tempLayer); layers.Layers(4)

ans = PreluLayer with properties: Name: 'p_re_lu_1' RawAlpha: [20x1 single] Learnable Parameters Alpha: [1x1x20 single] State Parameters No properties. Show all properties

Now thep_re_lu_1layer has a vector-valued scaling parameter.