The compiled executable file takes longer time when processing images in nvidia Jetson edge device

Dear Developer, I am using Nvidia Jetson64bit and Matlab 2024a. After I compiled the executable file for nvidia Jeton Orin. It takes 1 mins to start to process images although I load the model just once. It seems it loads the model every time it starts to process the images. This does not seem to be the situation with Matlab 2023b. Any idea for the scenerio?

[Chao Luo]

I assume the "model" means deep learning model. How do you load the model in MATLAB code? Do you make it persistent, like this:

persistent net;
if isempty(net)
    net = coder.loadDeepLearningNetwork('yourNetwork.mat');
end

[Liwei]

Yes. it is one of the matlab model (Xception and inceptionV3). I did the same thing. It worked good in last year with matlab 2023b. I have a new edge device, it looks work much slower to load the model at the beginning with Matlab 2024a. Did not have a chance to test the matlab 2023b yet.

[Chao Luo]

I don't see the issue on my side. Have you tried the PIL workflow like this:

hwobj = jetson('gpucoder-xaviernx-1', 'ubuntu', 'ubuntu');
cfg = coder.gpuConfig('dll');
cfg.Hardware = coder.Hardware('NVIDIA Jetson');
cfg.VerificationMode = 'PIL';
input = rand(299,299,3);
codegen -config cfg -args {input} foo
out = foo_pil(input); % you can run it multiple times to check the execution time.

function output = foo(input)
    persistent net;
    if isempty(net)
        net = coder.loadDeepLearningNetwork('inceptionv3');
    end
    output = net.predict(input);
end

It works for me that it does not load the model every time you call foo_sil.

[Liwei]

thanks. it takes more time than Matlab 2023b to process the first image (around 40 seconds vs over 1 mins). the model was retrained in Matlab 2023b. Retraining model in 2024a would help?