How Can I change the Layer Names of my Network So I can Import It?

I'm trying to import a network that I trained myself using this code. This code is run twice. It fails on the second attempt:

    classes = ["Parasitized" "Uninfected"];

    malaria_alexnet = 'C:\<directory>\malaria_alexnet_step2_accuracy_0.95952.onnx';
    malaria_alexnet = importONNXNetwork(malaria_alexnet,'OutputLayerType','classification','Classes',classes);

    lgraph = layerGraph(malaria_alexnet.Layers);
    malaria_alexnet = trainNetwork(augimdsTrain,lgraph,options);

The reason that I use **lgraph = layerGraph(malaria_alexnet.Layers);** is because I don't want to drop my weights.

However, when I import this network the second time, I get this error:

Error using nnet.internal.cnn.util.validateLayersForLayerGraph>iAssertUniqueAndNonEmptyLayerNames (line 56)
Layer names in layer array must be different from the names of layers in layer graph.

Error in nnet.internal.cnn.util.validateLayersForLayerGraph (line 33)
iAssertUniqueAndNonEmptyLayerNames(larray, existingLayers);

Error in nnet.cnn.LayerGraph>iValidateLayers (line 551)
larray = nnet.internal.cnn.util.validateLayersForLayerGraph(larray, existingLayers);

Error in nnet.cnn.LayerGraph/addLayers (line 190)
            larray = iValidateLayers(larray, existingLayers);

Error in nnet.internal.cnn.onnx.translateONNX>makeConnections (line 301)
    NNTLayerGraph = addLayers(NNTLayerGraph, NNTLayers{L});

Error in nnet.internal.cnn.onnx.translateONNX (line 131)
    NNTLayerGraphOrArray = makeConnections(NNTLayers, thisGraph, initializerNames);

Error in nnet.internal.cnn.onnx.importONNXNetwork (line 8)
LayersOrGraph   = nnet.internal.cnn.onnx.translateONNX(modelProto, OutputLayerType, UserImageInputSize, true);

Error in importONNXNetwork (line 52)
Network = nnet.internal.cnn.onnx.importONNXNetwork(modelfile, varargin{:});

Error in sgd_with_restarts_imported_network (line 17)
malaria_alexnet = importONNXNetwork(malaria_alexnet,'OutputLayerType','classification','Classes',classes);

How can I import my .onnx file or change the layer names to something different than what is found in layer graph?

Answer 1

[Sai Bhargav Avula]

Hi,

For the above mentioned purpose you can try using the importONNXLayers. Set the 'ImportWeights' name-value argument to true to retain the weights. This way you can play around with the layers, instead of importing it twice.

Hope this helps!