Network training, GPU out of memory
3 visualizzazioni (ultimi 30 giorni)
Mostra commenti meno recenti
I am trying to train a 3d convolutional network, but am running into memory issues with my gpu (GTX 3080, 10gb VRAM). My training dataset is 77.5 GB of 16x8x800 image sequences formatted as .mat files. I've already tried loweing the minibatchsize as low as it will go. I am currently trying to train with the cpu but it is painfully slow. Does anyone know how I can make this so that my GPU can run this?
clc
clear
imds = imageDatastore('filelocation','FileExtensions','.mat','LabelSource','none','ReadFcn',@matRead);
labels = load('filelocationLabels');
importLabels = labels.Labels;
imds.Labels = importLabels(1:length(dir('filelocationLabels'))-2);
[imds2,imds1] = splitEachLabel(imds,500,'randomized');
layers = [
image3dInputLayer([16 8 800 1],"Name","image3dinput","Normalization","none")
convolution3dLayer([3 3 3],32,"Name","conv3d")
convolution3dLayer([3 3 3],64,"Name","conv3d_1","Padding",[1 1 1;1 1 1])
maxPooling3dLayer([3 3 3],"Name","maxpool3d_1")
convolution3dLayer([3 3 3],64,"Name","conv3d_2","Padding",[1 1 1;1 1 1])
maxPooling3dLayer([3 3 3],"Name","maxpool3d_2")
fullyConnectedLayer(512,"Name","fc_1")
reluLayer("Name","relu_1")
fullyConnectedLayer(256,"Name","fc_2")
reluLayer("Name","relu_2")
fullyConnectedLayer(128,"Name","fc_3")
reluLayer("Name","relu_3")
fullyConnectedLayer(10,"Name","fc_4")
softmaxLayer("Name","softmax")
classificationLayer("Name","classoutput")];
options = trainingOptions('adam', ...
'InitialLearnRate',0.001, ...
'ExecutionEnvironment','gpu', ...
'ValidationData',imds2, ...
'ValidationFrequency',50, ...
'MiniBatchSize',1, ...
'MaxEpochs',8, ...
'GradientThreshold',2, ...
'Verbose',1, ...
'VerboseFrequency',1000,...
'Shuffle','every-epoch', ...
'Plots','training-progress');
HandyNet = trainNetwork(imds,layers,options);
%%
function data = matRead(filename)
inp = load(filename);
f = fields(inp);
data = inp.(f{1});
end
1 Commento
Abolfazl Chaman Motlagh
il 19 Ott 2021
your network is very heavy. so try to reduce number of filters in conv3d layers, i guess. you have 32 , 64 and 64 filters respectively. each are 3x3x3. for example start with 10 to see if it works.
also use stride in conv.
and also use bigger poolSize in maxpooling layer.
all above reduce data during training.
Risposte (0)
Vedere anche
Categorie
Scopri di più su Image Data Workflows in Help Center e File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!