How to replace imageInputLayer with sequenceInputLayer in imported Tensorflow Keras network

Hi,

I've imported a pre-trained network from tensorflow keras on MATLAB using importKerasLayers (importKerasNetwork didn't work as I've got 3 inputs). This network is made for sequence to sequence time series regression, hence I need a sequenceInputLayer in front of the network, however when I import the network an imageInputLayer is put in front of the network. Of course that layer is unsuitable for my use case, hence I tried replacing the imageInputLayer with a sequenceInputLayer but the error below is shown. The name of the imageInputLayer is 'lstm_10_input'.

Error using nnet.internal.cnn.util.validateLayersForLayerGraph>iAssertUniqueAndNonEmptyLayerNames (line 45)

Layer names in layer array must be nonempty.

Error in nnet.internal.cnn.util.validateLayersForLayerGraph (line 33)

iAssertUniqueAndNonEmptyLayerNames(larray, existingLayers);

Error in nnet.cnn.LayerGraph>iValidateLayers (line 556)

larray = nnet.internal.cnn.util.validateLayersForLayerGraph(larray, existingLayers);

Error in nnet.cnn.LayerGraph/replaceLayer (line 420)

larray = iValidateLayers(larray, existingLayers);

Error in keras_import (line 8)

lgraph = replaceLayer(lgraph,'lstm_10_input',inputLayer);

My code is as follows:

net = importKerasLayers('model.h5', 'ImportWeights', true);

lgraph = layerGraph(net);

inputLayer = sequenceInputLayer(3);

lgraph = replaceLayer(lgraph,'lstm_10_input',inputLayer); %This is where the error occurs

Answer 1

[Raunak Gupta]

Hi Lucas,

From the error message it is evident that the inputLayer in your code must contain a valid name (any character array) so that it can be replaced with ‘lstm_10_input’ layer. Giving a name to inputLayer will clear out the error message.

inputLayer = sequenceInputLayer(3,'Name','sequence_10_input'); 
lgraph = replaceLayer(lgraph,'lstm_10_input',inputLayer); 

[Paromita Kundu]

This is my code.I am getting the same error.What could be the reason?

layers = [

sequenceInputLayer(inputSize,'Name','input')

sequenceFoldingLayer('Name','fold')

convolution2dLayer(filterSize,numFilters,'Padding','same')

batchNormalizationLayer('Name','BN_1')

reluLayer('Name','relu_1')

maxPooling2dLayer(2,'Stride',2)

convolution2dLayer(filterSize,numFilters,'Padding','same')

batchNormalizationLayer('Name','BN_1')

reluLayer('Name','relu_1')

maxPooling2dLayer([4 2],'Stride',[4 2])

convolution2dLayer(filterSize,2*numFilters,'Padding',"same")

batchNormalizationLayer('Name','BN_1')

reluLayer('Name','relu_1')

maxPooling2dLayer([4 2],'Stride',[4 2])

convolution2dLayer(filterSize,2*numFilters,'Padding',"same")

batchNormalizationLayer('Name','BN_1')

reluLayer('Name','relu_1')

maxPooling2dLayer([4 2],'Stride',[4 2])

sequenceUnfoldingLayer('Name','unfold')

flattenLayer

lstmLayer(numHiddenUnits,'OutputMode',"last")

fullyConnectedLayer(4)

softmaxLayer

classificationLayer

];

lgraph = layerGraph(layers)

Error using layerGraph (line 62)

Layer names in layer array must be nonempty.

Error in two_cnn_lstm_ (line 58)

lgraph = layerGraph(layers);