i need to utilize fully of my GPUs during network training!

Greetings,

I am currently engaged in training a hybrid neural network architecture, GoogleNet and LSTM. My training dataset comprises a voluminous set of images amounting to approximately 2.5 million samples, and the total size of these images as stored in my computer's memory is approximately 18 gigabytes. During the training process, I am encountering constraints in utilizing all available computational resources offered by my computer's GPU(the gpu load mostly under 70%). In this context, I am contemplating the application of the "imageDatastore" function as a potential solution.

Kindly find appended herewith my code for your reference, and I would highly appreciate any guidance or recommendations that you might offer.

Sincerely,

Abolfazl

reset(gpuDevice(1));
% ---------- Load Data
allImages = imageDatastore(fullfile(parentDir,dataDir),...
    'IncludeSubfolders',true,...
    'LabelSource','foldernames');
rng default
[imgsTrain,imgsValidation] = splitEachLabel(allImages,0.80,'randomized');

% ---------- Hybrid Google Net with LSTM Layer 
net = googlenet;
lgraph = layerGraph(net);
% Modify GoogLeNet Network Parameters
newDropoutLayer = dropoutLayer(0.2,'Name','new_Dropout');
lgraph = replaceLayer(lgraph,'pool5-drop_7x7_s1',newDropoutLayer);

numClasses = numel(categories(allImages.Labels));
newConnectedLayer = fullyConnectedLayer(numClasses,...
    'WeightLearnRateFactor',5,'BiasLearnRateFactor',5);

LSTM_layer = [  flattenLayer;
    lstmLayer(350,OutputMode="last");
    newConnectedLayer];
lgraph = replaceLayer(lgraph,'loss3-classifier',LSTM_layer);
optimizer = 'adam';

newClassLayer = classificationLayer('Name','new_classoutput');
lgraph = replaceLayer(lgraph,'output',newClassLayer);


% ---------- Network Training 
% Set Training Options and Train GoogLeNet
options = trainingOptions(optimizer,...
    'MiniBatchSize',1024,...
    'MaxEpochs',15,...
    'InitialLearnRate',1e-4,...
    'ValidationData',imgsValidation,...
    'ValidationFrequency',10,...
    'Verbose',1,...
    'shuffle','every-epoch',...
    'ExecutionEnvironment','gpu',...
    'BatchNormalizationStatistics','moving',...
    'Plots','training-progress');
rng default

trainedGN = trainNetwork(imgsTrain,lgraph,options);

Answer 1

[Ben]

To use more of the GPU resource per iteration you can increase the minibatch size.

I'll note that the LSTM layer you are adding is not receiving sequence data, so it will simply run one iteration of the LSTM cell, which wouldn't be standard usage.

Answer 2

[Abolfazl Nejatian]

In regard to the minibatch size, it should be noted that the maximum permissible size has been selected due to memory constraints. Any attempt to further increase the minibatch size would result in a "out of memory" error being encountered.

When you say 'sequence data', could you please explain what you mean by that?

regards,

Abolfazl

[Ben]

>Any attempt to further increase the minibatch size would result in a "out of memory" error

In that case there is no way to utilize more GPU resources through trainNetwork.

>When you say 'sequence data', could you please explain what you mean by that?

LSTMs are typically used on sequence data such as time series. In practical terms they will iterate the LSTM "cell" operation over the sequence data, e.g. the T dimension of a dlarray with a dimension with label T. Typically LSTM is used in this way because it has a "state" that it can use at time step N that can hold information about timesteps 1,2,...,N-1. However in your example there is no sequence data, the LSTM "cell" just operates once.

[Abolfazl Nejatian]

Thank you for your information.

GPU utilization did not improve significantly when the pure GoogleNet model was trained. Therefore, the LSTM component wasn't solely responsible for the problems.

Answer 3

[Joss Knight]

It's hard to be sure from the info you provide but it looks like the filesystem is your bottleneck. If you cannot load the next 1024 images before you've finished the previous iteration, your GPU will have to wait.

[Abolfazl Nejatian]

Thank you for your insightful response. I concur with your assessment that the bottleneck in the system appears to be the filesystem, which may cause the GPU to wait if it cannot load the next 1024 images before the previous iteration is completed. Given that i have 250 GB RAM on my system, I am wondering if there are any potential solutions that could help reduce the loading time. Do you have any recommendations or strategies to address this issue?

[Joss Knight]

First use the MATLAB profiler to profile your code and demonstrate that the bottleneck is in the filesystem.

[Abolfazl Nejatian]

Am I correct in understanding that profiling here can be utilized to measure the runtime performance of the training network procedures, and subsequently identify the most time-consuming functions in trainNetwork function?

If this is accurate, could you please explain the process for obtaining the profille for my code?

[Joss Knight]

https://uk.mathworks.com/help/matlab/matlab_prog/profiling-for-improving-performance.html

[Abolfazl Nejatian]

I have included the profiler obtained from a smaller dataset trained for only one epoch.

Total run time: 1129.042 s

here is profiler results

https://drive.google.com/file/d/1BnrvjNgejWcxoZ5gBGc1aQKJkPOjpHcU/view?usp=share_link