Initialize a Neural Network from layers but do NOT train

I want to initialize a neural network from a set of layers, but I do not want to train it. Basically I just want to create it as a forward model, set the weights and biases randomly, and be able to evaluate it at given data point. The tutorials and documentation are not very clear on how to do this. For example from https://www.mathworks.com/help/nnet/examples/create-simple-deep-learning-network-for-classification.html they have

layers = [imageInputLayer([28 28 1])
          convolution2dLayer(5,20)
          reluLayer
          maxPooling2dLayer(2,'Stride',2)
          fullyConnectedLayer(10)
          softmaxLayer
          classificationLayer()];

But then they define the network through a training function:

convnet = trainNetwork(trainDigitData,layers,options);

I don't want this. I just want something along the lines of

convnet = network(layers);

Then I want to be able manually access all weights and biases, set them as I wish and be able to evaluate the network at a single or multiple input images. Is this possible? Thank you.

Answer 1

[Greg Heath]

Since

 trainedNet = trainNetwork(X, Y, layers, options)

 trains and returns a network, trainedNet,

Just use a random Y.

Then train the resulting net using your own algorithm.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Nikola Kovachki]

Yeah this was my thought too, but it just seemed hacky. I guess it's the only way to go. Thanks!

[jim peyton]

In case MATLAB developers are listening, I second that. It's a real use-case.

I'm working with hand-engineered filters and transfer learning.

Answer 2

[Vla Ser]

https://www.mathworks.com/help/deeplearning/ref/assemblenetwork.html - you can use assemble network to create it from list of layers.

[Royi Avital]

For DAGNetwork I get an error that weights are not initialized. Have you solved this?