Issue with LSTM OutputMode='last' in MATLAB Deep Learning HDL Toolbox

Hi all!

I am profiling the performance of the MATLAB Deep Learning HDL Toolbox for an LSTM-based model trained on the IMDb dataset.

I leave below the code I am running:

% Initialize model (code generated by the Deep Network Designer App after importing the tensorflow model)
params = load("params_2025_01_22__11_31_51.mat");

net = dlnetwork;

tempNet = [
    sequenceInputLayer(1,"Name","embedding_input","MinLength", 235)
    lstmLayer(32,"Name","lstm","outputMode", "last", "BiasInitializer","narrow-normal","InputWeightsInitializer","narrow-normal","RecurrentWeightsInitializer","narrow-normal")
    fullyConnectedLayer(1,"Name","dense","Bias",params.dense.Bias,"Weights",params.dense.Weights)
    sigmoidLayer("Name","dense_sigmoid")
    ];

net = addLayers(net,tempNet);

% clean up helper variable
clear tempNet;

net = initialize(net);

% Initialize dlhdl processor with default configuration
hPC = dlhdl.ProcessorConfig

% Optimize configuration for the target network
optimizeConfigurationForNetwork(hPC, net)
hPC.estimateResources
hPC.estimatePerformance(net)

I noticed that the "outputMode='last'" property of the LSTM layer generates the following error when customizing the processor for the target network:

"Error using dnnfpga.macros.verifyRNNLayerParameters

The layer 'lstm' has an OutputMode of 'last' which is not supported.

Error in dnnfpga.macros.createLSTMLayerNet

Error in dnnfpga.macros.Macros.createNet

Error in dnnfpga.compiler.optimizations.preprocessNetworkForDLP

Error in dnnfpga.compiler.cnn5ProcessorFrontend/runFrontend

Error in dnnfpga.compiler.cnn5ProcessorFrontend/doit

Error in dnnfpga.config.CNN5ProcessorConfig/optimizeConfigurationForNetwork

Error in lstm_imdb_test (line 26)

optimizeConfigurationForNetwork(hPC, net)"

I temporarly switched this property to 'sequence' and it seems working.

Please, can you help me find any explanation for this problem?

Thanks in advance for your help!

Answer 1

[Wang Chen]

Hi Tommaso,

The "OutputMode" property = "last" is not supported in Deep Learning HDL Toolbox yet.

https://www.mathworks.com/help/deep-learning-hdl/ug/supported-networks-layers-boards-and-tools.html

"When generating code for a network using this layer, these limitations apply:

• The input must be of single data type.
• The OutputMode property must be set to sequence."

As a work-around, you could use "OutputMode" property = "sequence", and pick out last time step's output from the complete output sequences.

Thanks,

Wang