Clear Filters
Clear Filters

loading and training an existing network.

37 views (last 30 days)
Mark Hubelbank
Mark Hubelbank on 7 Oct 2024 at 20:51
Edited: Matt J on 8 Oct 2024 at 0:58
I am trying define a network, then train it in multiple sessions. The problem is that I can't get the load or read of the network to work in the second session. The code is:
layers = [ ...
sequenceInputLayer(270)
bilstmLayer(numHiddenUnits,OutputMode="last")
fullyConnectedLayer(numClasses)
softmaxLayer
classificationLayer]
options = trainingOptions("adam", ...
InitialLearnRate=0.002,...
MaxEpochs=15, ...
Shuffle="never", ...
GradientThreshold=1, ...
Verbose=false, ...
ExecutionEnvironment="gpu", ...
Plots="training-progress");
clabels=categorical(labels);
numLables=numel(clabels)
load("savednet.mat","layers");
net = trainNetwork(data,clabels,layers,options);
save("savednet","net");
I have tried many variations of the load command and it always gives an error on the second argument:
Warning: Variable 'layers' not found.
Exactly what should that look like and then how should it be used as input to the trainNetwork routine?
  7 Comments
Mark Hubelbank
Mark Hubelbank on 7 Oct 2024 at 23:28
Moved: Walter Roberson on 7 Oct 2024 at 23:33
Perhaps I don't understand how one can train in stages then. The idea is that the training will be continued in the second and subsequent sessions. Sort of a continuing transfer learning. The idea is that over time the network keeps improving. perhaps I should be using trainnet instead of trainnetwork. Then it would appear the call is:
load(filename,"net1","layers");
net=trainnet(data,clabels,net1,"crossentropy",options);
Is this the correct direction?
Walter Roberson
Walter Roberson on 7 Oct 2024 at 23:36
Probably
net1 = trainnet(data,clabels,net1,"crossentropy",options);

Sign in to comment.

Accepted Answer

Matt J
Matt J on 7 Oct 2024 at 23:54
previous = load("savednet","net","layers");
net = trainNetwork(data,clabels,previous.net,options);
  1 Comment
Matt J
Matt J on 7 Oct 2024 at 23:55
Edited: Matt J on 8 Oct 2024 at 0:58
perhaps I should be using trainnet instead of trainnetwork.
It would be better, since trainnet is newer and has more flexibility. However, it won't make a difference as far as how to resume the training of a pre-existing network..

Sign in to comment.

More Answers (0)

Categories

Find more on Image Data Workflows in Help Center and File Exchange

Products


Release

R2024b

Community Treasure Hunt

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

Start Hunting!