Error while running importKerasNetwork

Hello,

I have a Matlab script (version 2020b) in which I import a neural network from Python using the next command:

importKerasNetwork(modelName)

Before the script was working well. After reinstalling my Matlab (same version, 2020b), I get the following error:

SWITCH expression must be a scalar or a character vector.

Error in nnet.internal.cnn.keras.LayerSpec>DAGOutputLayer (line 166)

switch Loss

Error in nnet.internal.cnn.keras.LayerSpec/maybeAppendOutputLayer (line 148)

NNTLayers{end+1} = DAGOutputLayer(OutputTensors{pos}, TrainingConfig, isRNN);

Error in nnet.internal.cnn.keras.LayerSpec/translateDAGLayer (line 64)

NNTLayers = maybeAppendOutputLayer(this, NNTLayers, OutputTensors, TrainingConfig, isRNN);

Error in

nnet.internal.cnn.keras.AssembledModel>@(LSpec)translateDAGLayer(LSpec,AM.OutputTensors,AM.TrainingConfig,TranslateWeights,TranslateTrainingParams,UserImageInputSize,isRNN(AM))

(line 226)

NNTLayerGroups = cellfun(@(LSpec)translateDAGLayer(LSpec, AM.OutputTensors, AM.TrainingConfig, TranslateWeights,...

Error in nnet.internal.cnn.keras.AssembledModel/translateDAGModel (line 226)

NNTLayerGroups = cellfun(@(LSpec)translateDAGLayer(LSpec, AM.OutputTensors, AM.TrainingConfig, TranslateWeights,...

Error in nnet.internal.cnn.keras.AssembledModel/translateAssembledModel (line 57)

Layers = translateDAGModel(this, TranslateWeights, TranslateTrainingParams, UserImageInputSize);

Error in nnet.internal.cnn.keras.importKerasNetwork (line 30)

LayersOrGraph = translateAssembledModel(AM, TrainingConfig, ImportWeights, TrainOptionsRequested, UserImageInputSize);

Error in importKerasNetwork (line 91)

Network = nnet.internal.cnn.keras.importKerasNetwork(modelfile, varargin{:});

The error is given in a line of a scrpit that I haven't programmed. Could anyone help me?

Thank you!

Answer 1

[David Willingham]

Hi,

This has been observed before. I don't believe it's due to reinstalling MATLAB.

Please check if the following work around will work:

==

If one provides the loss value using a Python list (shown below), this issue will occur:

model.compile(loss=['categorical_crossentropy'],

optimizer=keras.optimizers.Adadelta(),

metrics=['accuracy'])

Although Keras accepts a list for the loss input, this type is not mentioned as one of the accepted type in Keras documentation https://keras.io/losses/.

If the loss is specified this way the problem should go away:

model.compile(loss='mean_squared_error')

==

[Irati Ibañez Hidalgo]

My keras model has the loss defined like:

model.compile(

optimizer=optimizer,

loss=['mean_squared_error'],

metrics = ['mse'])ç

Is it the problem?

Anyway, what I don't understand is why before the reinstallation the script was working with the same neurla network that I'm trying to upload now and now it isn't working.

I changed the line 166 in this file nnet.internal.cnn.keras.LayerSpec > DAGOutputLayer to:

switch Loss{1}

And after that it seems to work correctly.