DL + GPU Coder: Code generation for SequenceInputLayer is not supported.

Hello guys,

I have a DL model I'd like to use to import to Simulink, with the help of the GPU coder. Is a simple function, and my model is uploaded in this message. The error I get when I try to run the "Check for Issues" option in the GPU Coder app from Matlab, is the following:

??? Error generating code for network n_0. Code generation for SequenceInputLayer SequenceInputLayer is not supported.

The function is the following:

function out=modelPred(data)
out=zeros(1,4);
model=coder.loadDeepLearningNetwork('n.mat','n');
out(1,:)=predict(model, data);
end

The full error log looks like this, see the screenshot:

Do you know any way in which I can solve this issue? Thanks in advance.

Answer 1

[Walter Roberson]

on Windows you need to install Visual Studio 2013 Professional or 2015 Professional or 2017 (any edition) and then use mex -setup to configure .

[Ajpaezm]

Hey there Walter, thanks for replying.

So with 2017 community version would do?

[Walter Roberson]

Yes, 2017 Community should be fine.

[Ajpaezm]

Walter, an additional question, and perhaps I should've start with this, but I noticed CUDA boards are almost strictly from NVIDIA, right?

The graphic card of my laptop is Intel Graphics 4000. And is not CUDA enabled far as I'm concerned.

Is this the only way to generate CUDA code in Matlab?

[Walter Roberson]

CUDA is definitely only NVIDIA.

It is not impossible that at some point Intel might buy NVIDIA (which has revenues about 1/7th those of Intel these days), and if that ever happens, then Intel graphics boards might gain CUDA capabilities. But for now they are rivals.

Intel and a number of other companies follow OpenCL for parallel computing; however it turns out that OpenCL has so much room for variation that Mathworks would end up having to tune their system for each different manufacturer, and probably even each different major product line.

[Hariprasad Ravishankar]

In addition, GPU Coder added support for Sequence Input and LSTM Layers in R2019b.

Please use a supported compiler and R2019b to generate code for LSTM networks using CuDNN target

Supported compilers:

https://www.mathworks.com/help/gpucoder/gs/install-prerequisites.html#mw_aa8b0a39-45ea-4295-b244-52d6e6907bff