Load multiple Files from fileDatastore in minibatchqueue

조회 수: 5 (최근 30일)
Simon
Simon 2025년 3월 12일
댓글: Joss Knight 2025년 3월 14일
When training my neural network, I noticed that approximately 50% of the runtime is spent on loading data. Currently, the data loading process from my minibatchqueue is sequential and involves numerous function calls. Specifically, each minibatch requires 256 individual function calls, with each file loaded one after another, causing significant delays.
My goal is to parallelize or otherwise optimize this data-loading process to significantly reduce the runtime.
I'm looking for recommendations or best practices to adress the issue.
Any suggestions or ideas for enhancing data-loading efficiency would be greatly appreciated.
the path is similar to this: "C:/user/me/data/s*/spec.mat"
% Create Datastore
fdsNetInput = fileDatastore(path, ReadFcn=@loadDatastoreData, FileExtensions=".mat");
% add labels and create trainSet (not the core of my Question)
fullData = combine(fdsNetInput, fdsLabel);
trainMask = trainMask(randperm(N));
trainData = subset(fullData, trainMask);
my ReanFcn:
function [netInput] = loadDatastoreData(file)
% Load Spectrum
netInputSpectrum = load(file, "spectrum"); %<-------------------------- This Line takes a lot of time
% scale Spectrum to the right size
netInputSpectrum = scaleData(netInputSpectrum, "VGGish", false);
% ouput Spectrum with the propper dims
netInput = dlarray(netInputSpectrum, 'SSC');
end
then I create a minibatchqueue with the fileDatastore
% create train minibatchqueue
trainMBQ = minibatchqueue(trainData,...
MiniBatchFormat = ["SSCB", "CB"], ...
MiniBatchSize = 256, ...
PartialMiniBatch="discard", ...
OutputEnvironment="gpu");
Read the Data from the minibatchqueue to process it.
% Read mini-batch of data.
[input,target] = next(trainMBQ);

채택된 답변

Joss Knight
Joss Knight 2025년 3월 12일
Does the training option PreprocessingEnvironment set to 'background' not work for you? Or is this a custom training loop?
  댓글 수: 2
Simon
Simon 2025년 3월 13일
Thank you for your fast Answer.
it is a custom training loop and I added PreprocessingEnvironment="parallel" in the Minibatchqueue, witch inproved the runtime a lot.
trainMBQ = minibatchqueue(trainData,...
MiniBatchFormat = ["SSCB", "CB"], ...
MiniBatchSize = minibatchSize, ...
PartialMiniBatch="discard", ...
PreprocessingEnvironment="parallel", ...
OutputEnvironment="gpu");
Joss Knight
Joss Knight 2025년 3월 14일
Great! "background" should hopefully also work and be faster.

댓글을 달려면 로그인하십시오.

추가 답변 (0개)

카테고리

Help CenterFile Exchange에서 Sequence and Numeric Feature Data Workflows에 대해 자세히 알아보기

제품


릴리스

R2024b

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by