Matlab Coder for DeepLearning

The DL layers supported by Matlab Coder have been listed and updated contineously in the link below:

https://ww2.mathworks.cn/help/deeplearning/ug/list-of-deep-learning-layers.html?searchHighlight=sequencefoldinglayer&s_tid=srchtitle_support_results_7_sequencefoldinglayer

Many layers have been supported by Generc C/C++ nowadays.

My question is:

When is it possible for the Matlab Coder to support SequenceFolding Layer & SequenceUnfolding Layer with Generic C/C++? Is there a plan for these two layers? Or they will just be skipped?

[Sergio Matiz Romero]

Hi @Yang Liu,

Thank you for reaching out. Does your application require that you insert the folding and unfolding layers explicitly for a particular reason? Notice that you can completely avoid the insertion of these layers if you use dlnetwork instead of DAG networks. For instance, you can construct a convolution + LSTM network as:

layers = [ 
    sequenceInputLayer([8 8 3], 'name','seq')
    convolution2dLayer(3,3, 'Name', 'convolution1','Padding','same')
    lstmLayer(20,'name','lstm')
    fullyConnectedLayer(10,'Name','fc')
    ];

dlnet = dlnetwork(layers);

and the above network does support generic C/C++ code generation. On the other hand, when using DAG networks, you would need to insert sequence folding/unfolding layers (around convolution), which are not currently supported for generic C/C++ code generation.

Can the network you are working with be expressed as a dlnetwork to avoid the use of the unsupported layers? If so, I would recommend using a dlnetwork since it will soon become the recommended workflow. Otherwise, please let me know more about your use case to be able to further assist you

[Yang Liu]

Dear Sergio,

Thanks a lot for your detailed explaination and suggestion!

My network is actually FC(Fully Connect)+LSTM for classification. My layers are constructed as below (sorry, a bit lenthy). Seveal layers of FCs connect with concadenated LSTMs.

Maybe I'm out-dated about the officially recommended operations (yes, I use the DAG network currently), but can "dlnetwork" also waive the sequence folding/unfoldng layer, when applying to FC network?

You code seems neat and concise!

layers = [ ...
    sequenceInputLayer(inputSize,'Name','Tgt_Seq_Input')
    
    sequenceFoldingLayer('Name','SeqFold')
    
    fullyConnectedLayer(32,'Name','FC_Out_32')
    batchNormalizationLayer('Name','BN_FC_Out_32')
    reluLayer('Name','Relu_FC_Out_32')
    
    fullyConnectedLayer(16,'Name','FC_Out_16')
    batchNormalizationLayer('Name','BN_FC_Out_16')
    reluLayer('Name','Relu_FC_Out_16')
    
    fullyConnectedLayer(8,'Name','FC_Out_8')
    batchNormalizationLayer('Name','BN_FC_Out_8')
    reluLayer('Name','Relu_FC_Out_8')
    
    fullyConnectedLayer(4,'Name','FC_Out_4')
    batchNormalizationLayer('Name','BN_FC_Out_4')
    reluLayer('Name','Relu_FC_Out_4')

    sequenceUnfoldingLayer('Name','SeqUnfold')
    
    lstmLayer(numHiddenUnits,'OutputMode','sequence','Name','lstm1')
    lstmLayer(numHiddenUnits,'OutputMode','sequence','Name','lstm2')
    lstmLayer(numHiddenUnits,'OutputMode','sequence','Name','lstm3')
    lstmLayer(numHiddenUnits,'OutputMode','last','Name','lstm4')

    fullyConnectedLayer(numClasses, 'Name','FC_Out_2')
    softmaxLayer('Name','softmax')
    classificationLayer('Name','classification')
    ];

lgraph = layerGraph(layers);
lgraph_1 = connectLayers(lgraph,'SeqFold/miniBatchSize','SeqUnfold/miniBatchSize');
figure
plot(lgraph_1)

options = trainingOptions('sgdm', ...
    'ExecutionEnvironment','cpu', ...
    'GradientThreshold',1, ...
    'InitialLearnRate',0.5,...
    'LearnRateSchedule','piecewise', ...
    'LearnRateDropFactor',0.5, ...
    'LearnRateDropPeriod',10, ...
    'Momentum',0.15,...
    'MaxEpochs',maxEpochs, ...
    'MiniBatchSize',miniBatchSize, ...
    'Shuffle','every-epoch', ...
    'ValidationData',{TS_Train_Comb_Cell_Validation,TS_Train_Comb_Cell_Validation_Label}, ...
    'ValidationFrequency',300, ...
    'Verbose',1, ...
    'Plots','training-progress');

% net = trainNetwork(trainingDatastore,lgraph_1,options);
net = trainNetwork(TS_Train_Comb_Cell_Train,TS_Train_Comb_Cell_Train_Label,lgraph_1,options);