MATLAB: How to generate code from a trained LSTM network using MATLAB/Simulink

Question

Is code generation from a trained LSTM network supported, and if so, how? There appear to be different approaches to generate code from a trained LSTM network while working without/with Simulink.

Answer 1

Generate code from a trained LSTM network using MATLAB

In order to generate code from a trained LSTM network using MATLAB, there are different approaches:

1) Generate C/C++ code using MATLAB Coder:

From MATLAB R2020a onwards, you can generate code for ARM Cortex-A CPUs. Refer to the blog post below on how to get started with deep learning models on ARM Cortex-A with MATLAB:

https://community.arm.com/developer/tools-software/tools/b/tools-software-ides-blog/posts/getting-started-with-deep-learning-models-on-arm-cortex-a-with-matlab

Another example is here:

https://www.mathworks.com/help/coder/ug/code-generation-for-lstm-network-on-raspberry-pi.html

From MATLAB R2020b onwards, you can generate C++ code for an LSTM network, a stateful LSTM network, or a bidirectional LSTM network that uses the Intel® MKL-DNN library.

2) Generate CUDA code using GPU Coder:

From R2019b onwards, you can generate code for NVIDIA GPUs. See the following documentation link for more information:

https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.lstmlayer.html#mw_a7799934-4908-476b-809d-1f063ef6a7fc

Generate code from a trained LSTM network using Simulink

In order to generate code from a trained LSTM network using Simulink, the following approach will be possible from R2020b onwards:

Using MATLAB Coder or GPU Coder, you can use* coder.loadDeepLearningNetwork* and other associated functions to run inference in a MATLAB Function block. This will support accelerated simulation and code generation for both CPU or GPU targets.