MATLAB: How to compile Deep learning Neural Network function

convnetmodelpre-trainedpretrainedtrainnetwork

I am trying to create an executable with MATLAB compiler for neural network toolbox example here:

https://www.mathworks.com/help/releases/R2017b/nnet/examples/create-simple-deep-learning-network-for-classification.html

I get this error:

error: some functionality cannot be deployed

Best Answer

Starting R2016b MATLAB release:

You should be able to compile 'trainNetwork' and most command line functions (from both classical and deep learning networks) starting in R2016b.

Functions that cannot be compiled include the deep learning training "plot" function and all user interfaces.

Please refer the 'Neural Network Toolbox' product in this link for information on this:

https://www.mathworks.com/products/compiler/compiler_support.html

Prior to R2016b release:

You can only compile a pre-trained network and use classify function to classify the test data.

So in order to compile the doc example below,

https://www.mathworks.com/help/releases/R2017b/nnet/examples/create-simple-deep-learning-network-for-classification.html

in releases prior to R2016b, please follow the steps below:

1. Run the example code in MATLAB

2. This will create the 'convnet' struct variable in your workspace. This is the pretrained network object. Save this to a mat file like below:

save 'model.mat' convnet

3. Also save the testImageData variable in the workspace to a mat file:

save 'testDigitData.mat' testDigitData

4. Then you can create a MATLAB function like below to be compiled into an executable that used the pretrained network to classify the test data.

function accuracy = runModelFromMATLAB()\nload('model.mat');\nload('testDigitData.mat')\n\nYTest = classify(convnet,testDigitData);\nTTest = testDigitData.Labels;\naccuracy = sum(YTest == TTest)/numel(TTest)\nend

5. Then create executable for this function with MATLAB compiler

Related Solutions

MATLAB: Transfer learning layers in alexnet and googlenet

Here is a page where you can compare pretrained networks: https://www.mathworks.com/help/deeplearning/ug/pretrained-convolutional-neural-networks.html

MATLAB: How to compile a pre-trained neural network into a standalone application using the MATLAB Compiler 4.10 (R2009a)

The following script sets up the creation of a pre-trained network and stores it in a MAT-file ('pre_trained_net.mat').

clc
clear all
close all
P = [1 2 3 4 5];
T = [1 2 3 4 3];
net = newff(P,T,5);
net.trainParam.epochs = 50;
net = train(net,P,T);
save pre_trained_net net

The following function utilizing the pre-trained network can be compiled into a stand-alone application:

function compiled_network(input)
% Include Neural Network Toolbox dependencies.
% web([docroot '/toolbox/compiler/function.html'])
%#function network
% Load data.
load pre_trained_net
if ischar(input)
      input=str2num(input);
end
output = sim(net,input)
plot(input,output,'o')

The following command may be used to compile the above function into a stand-alone application:

mcc -m compiled_network

The compiled application may be called from the system command line as follows to generate output for new input:

compiled_network([1,2,3,4,5])

For your convenience, find attached 'pre_trained_net.mat' and 'compiled_network.exe'. The executable was compiled using MATLAB Compiler 4.10 (R2009a) on Windows XP. These files can always be created as described earlier.

Best Answer

Related Solutions

MATLAB: Transfer learning layers in alexnet and googlenet

MATLAB: How to compile a pre-trained neural network into a standalone application using the MATLAB Compiler 4.10 (R2009a)

Related Question