Work Around for Convolution1DLayer

I am trying to do code generation of a trained deep learning network however it uses a convolution1dlayer. It seems this is not currently supported by matlab. What are some possible solutions to this problem?

Answer 1

[Aditya]

Hi Kevin,

Please do refer to this MATLAB answer post by MathWorks Support Team on the same issue: Is code generation supported for "convolution1DLayer"? - MATLAB Answers - MATLAB Central (mathworks.com)

Hope this helps!

Answer 2

[Ram Kokku]

Hi @Kevin Monahan,

Convolution1DLayer supports code generation in the latest (R2024a) version of MATLAB. See the extended capabilities section : here - https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.convolution1dlayer.html

Answer 3

[Steven Lord]

Are you certain that you're using release R2024a? It appears from the Release Notes that support for generating code from this layer was added in release R2024a.

[Kevin Monahan]

I am. The error specifically saps mlkdnn does not support this. It also does not work with Cuda. I am running 2024a.

[Ram Kokku]

@Kevin Monahan, use TargetLibrary=none (https://www.mathworks.com/help/coder/ref/coder.deeplearningconfig.html). Both MATLAB Coder and GPU Coder support TargetLibrary=none. so, you should be able to generate both C/C++ and CUDA. For Better performance, you may need to explicit enable hardware features like instructionset (AVX2, AVX512) and multi-threading. See https://www.mathworks.com/help/coder/ug/optimize-generic-c-cpp-code-performance.html for more details.

[森]

I meet the same problem when generating coder containg Convolution1DLayer using TargetLibrary=none, and I also use the version 2024a. So this means actualy version 2024a stiil does not support the code generation for Convolution1DLayer?

Answer 4

[Matt J]

Why not just use a convolution2dLayer? A 1D input is just a special case of a 2D input.

[Kevin Monahan]

I did not make the model. It was given to me pretrained. I need to take it and move it out of Matlab for future research.

[Matt J]

Why does that matter? It was given to you, but surely you can use replaceLayer to swap the 1D convolutional layers for 2D ones.