MATLAB Answers

Errors in transfer learning using resnet101

51 views (last 30 days)
KEN SUEMATSU
KEN SUEMATSU on 15 Mar 2021
Answered: Akira Agata on 18 Mar 2021
I would like to use resnet101 to do transfer learning.
When I build the network and use the trainNetwork function as shown below, I get the following error. What is the cause?
Layer 'res2a': unconnected input. The input of each layer must be coupled with the output of another layer.
An unconnected input was detected:
net = resnet101;
layers = net.Layers;
layers = [
layers(1:344)
fullyConnectedLayer(Numberofclasses)
layers(346)
classificationLayer];
options = trainingOptions('sgdm',...
'MiniBatchSize',16,...
'InitialLearnRate', 0.0001, ...
...)
trainNetwork(TrainImage,TrainData,layers,options);

Accepted Answer

Akira Agata
Akira Agata on 18 Mar 2021
Since ResNet-101 is imported as a DAGNetwork object, the following steps will be needed (more details can be found in this Link)
  1. Convert DAGNetwork object to LayerGraph object
  2. Replace the last few layers
  3. Freeze bias/weight of initial layers (optional)
  4. Re-connect all the layers in the original order by using the support function createLgraphUsingConnections
So the MATLAB code will be like this.
net = resnet101;
% 1. Convert DAGNetwork object to LayerGraph object
lgraph = layerGraph(net);
% 2. Replace the last few layers
lgraph = replaceLayer(lgraph,'fc1000',...
fullyConnectedLayer(Numberofclasses,'Name','fcNew'));
lgraph = replaceLayer(lgraph,'ClassificationLayer_predictions',...
classificationLayer('Name','ClassificationNew'));
% 4. Re-connect all the layers in the original order
% by using the support function createLgraphUsingConnections
layers = lgraph.Layers;
connections = lgraph.Connections;
lgraph = createLgraphUsingConnections(layers,connections);
% Train the network
options = trainingOptions('sgdm',...
'MiniBatchSize',16,...
'InitialLearnRate', 0.0001, ...
...)
net = trainNetwork(imdsTrain,lgraph,options);

More Answers (0)

Products


Release

R2020a

Community Treasure Hunt

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

Start Hunting!

Translated by