Out of Memory on classify(.​..,'Execut​ionEnviron​ment','cpu​') on SUSE Linux

I've got an issue with the classify function running on SUSE Linux. I've got a network which is trained for image analysis. When I try to classify new images (100-500) with the network, it takes a while until Matlab takes all the system memory (48gb) and eventually gets killed by the OS (i.e. the Matlab process just shuts down).

x=classify(net,beadImgs,'ExecutionEnvironment','cpu');

The same program works fine on a Windows machine with 16gb of RAM either with 'gpu' or with 'cpu' Execution environment (I cant test the 'gpu' Option on the Linux system). My guess is that this memory footprint is 'somewhat unintended'. Does anybody have an idea on how to fix this?

EDIT:

Okay I figured out a workaround. Manually setting the batch size fixes the issue for the moment. Nevertheless the different memory footprint on different operating systems is still mysterious.

ver yields:

--------------------------------------------------------------------------------------------------------
MATLAB Version: 9.7.0.1247435 (R2019b) Update 2
MATLAB License Number: #######
Operating System: Linux 4.12.14-lp151.28.32-default #1 SMP Wed Nov 13 07:50:15 UTC 2019 (6e1aaad) x86_64
Java Version: Java 1.8.0_202-b08 with Oracle Corporation Java HotSpot(TM) 64-Bit Server VM mixed mode
--------------------------------------------------------------------------------------------------------
MATLAB                                                Version 9.7         (R2019b)
Simulink                                              Version 10.0        (R2019b)
Bioinformatics Toolbox                                Version 4.13        (R2019b)
Communications Toolbox                                Version 7.2         (R2019b)
Computer Vision Toolbox                               Version 9.1         (R2019b)
Control System Toolbox                                Version 10.7        (R2019b)
Curve Fitting Toolbox                                 Version 3.5.10      (R2019b)
DSP System Toolbox                                    Version 9.9         (R2019b)
Database Toolbox                                      Version 9.2         (R2019b)
Deep Learning Toolbox                                 Version 13.0        (R2019b)
Econometrics Toolbox                                  Version 5.3         (R2019b)
Embedded Coder                                        Version 7.3         (R2019b)
Fixed-Point Designer                                  Version 6.4         (R2019b)
Fuzzy Logic Toolbox                                   Version 2.6         (R2019b)
Global Optimization Toolbox                           Version 4.2         (R2019b)
Image Acquisition Toolbox                             Version 6.1         (R2019b)
Image Processing Toolbox                              Version 11.0        (R2019b)
Instrument Control Toolbox                            Version 4.1         (R2019b)
MATLAB Coder                                          Version 4.3         (R2019b)
MATLAB Compiler                                       Version 7.1         (R2019b)
Mapping Toolbox                                       Version 4.9         (R2019b)
Model Predictive Control Toolbox                      Version 6.3.1       (R2019b)
Optimization Toolbox                                  Version 8.4         (R2019b)
Parallel Computing Toolbox                            Version 7.1         (R2019b)
Partial Differential Equation Toolbox                 Version 3.3         (R2019b)
Signal Processing Toolbox                             Version 8.3         (R2019b)
Simscape                                              Version 4.7         (R2019b)
Simscape Electrical                                   Version 7.2         (R2019b)
Simscape Multibody                                    Version 7.0         (R2019b)
Simulink 3D Animation                                 Version 8.3         (R2019b)
Simulink Coder                                        Version 9.2         (R2019b)
Simulink Control Design                               Version 5.4         (R2019b)
Simulink Design Optimization                          Version 3.7         (R2019b)
Stateflow                                             Version 10.1        (R2019b)
Statistics and Machine Learning Toolbox               Version 11.6        (R2019b)
Symbolic Math Toolbox                                 Version 8.4         (R2019b)
System Identification Toolbox                         Version 9.11        (R2019b)
Vehicle Network Toolbox                               Version 4.3         (R2019b)
Wavelet Toolbox                                       Version 5.3         (R2019b)

[Joss Knight]

What happens when you use 'Acceleration', 'none' in your call to classify. Does that make it work?

[GW]

I'm sorry for the late answer. I finally found the time to test your suggestion now. Indeed that makes it work as well.

Kind regards

Answer 1

[Kaashyap Pappu]

This response is just for future reference:

Modifying the Name-Value pair, ‘MiniBatchSize’, to a lower value can help with this issue. The default value is 128. If a similar issue is present during training, you can similarly set the same property in trainNetwork and imageDatastore to a lower value to help with this issue.