Restart Training in Semantic Segmentation of Multispectral Images Using Deep Learning

Hi,

I went through the example "Semantic Segmentation of Multispectral Images Using Deep Learning." I'd like to restart training to improve the network I trained and, according to MATLAB's help, one needs to provide the net.Layers data to trainNetwork.

The command I used to restart training is

[net,info] = trainNetwork(dsTrain,net.Layers,options);

Launching the code I get the following error:

Error using trainNetwork
Invalid network.

Error in Deep_Learning_Testing (line 2357)
[net,info] = trainNetwork(dsTrain,net.Layers,options);

Caused by:
Layer 'Decoder-Section-1-DepthConcatenation': Unconnected input. Each layer input must be connected to the output of another layer.
Layer 'Decoder-Section-2-DepthConcatenation': Unconnected input. Each layer input must be connected to the output of another layer.
Layer 'Decoder-Section-3-DepthConcatenation': Unconnected input. Each layer input must be connected to the output of another layer.
Layer 'Decoder-Section-4-DepthConcatenation': Unconnected input. Each layer input must be connected to the output of another layer.

Any idea why this error pops up?

Thanks,

Jacopo

Answer 1

[Parth Parikh]

Hi,

Mostly the error is due to DAGNet object is being passed to trainedNetwork, the input should be a layergraph object.

Instead of this: [net,info] = trainNetwork(dsTrain,net.Layers,options);

It should be: [net,info] = trainNetwork(dsTrain,layerGraph(net),options);

If you would like to know more about the hyperspectral/multispectral features than take a look at Hyperspectral Imaging library:

Image Processing Toolbox Hyperspectral Imaging Library - File Exchange - MATLAB Central (mathworks.com)

Doc link: Hyperspectral Image Processing - MATLAB & Simulink (mathworks.com)