I want to train a network and save the weights. Then use a new data set and resume training with the weights from the previous data set.

It seems train always randomizes the weights when called. I can't seem to maintain the weights from the first training set.

Answer 1

[Sonam Gupta]

You can continue the training from weights obtained by previous data set by extracting the layers from the network's "Layers" property, and then passing it to "trainNetwork", as follows:

if true
  % Train a network
  net = trainNetwork(XTrain, YTrain, layers, options);
  % Extract layers from the trained network
  newLayers = net.Layers;
  % Retrain the network, but start from where we left off
  newNet = trainNetwork(XTrain, YTrain, newLayers, options);

'trainNetwork' will always use the weights that are stored in the layers which you pass in for training.

[ronald ogren]

Thank you very much. I did not find anything like that in the documentation

[Grzegorz Klosowski]

I join the thanks. Good job.

[Moh. Saadat]

There is a small caveat to this: check that whether your output 'net' is a LayerGraph or a DAGNetwork. If it is not a LayerGraph, use layerGraph(net) instead of net.Layers.

Source: https://www.mathworks.com/help/deeplearning/ug/resume-training-from-a-checkpoint-network.html