Change parameters of network from Deep Network Designer

Hi,

I haven't been able to find the answer to this question in the boards, but perhaps I'm using the wrong terminology.

I designed a 2D Unet using the Deep Network Designer so that I could get a better understanding of how everything links together and the different parameters of each layer. I did the "generate code" option so that I can easily run the .mlx file and get my network.

Now I want a different Unet that has the same structure, but a different number of filters in the final convolution layer. I can manually edit the network using the Deep Network Designer, but I'd rather do this programmatically, however I get a read only error:

This is the final convolution layer from the Deep Network Designer

>> lgraph.Layers(76)
ans = 
  Convolution2DLayer with properties:

              Name: 'conv_19'
   Hyperparameters
        FilterSize: [3 3]
       NumChannels: 'auto'
        NumFilters: 2
            Stride: [1 1]
    DilationFactor: [1 1]
       PaddingMode: 'same'
       PaddingSize: []
      PaddingValue: 0
   Learnable Parameters
           Weights: []
              Bias: []

This is the error that I get when changing the NumFilters

>> lgraph.Layers(76).NumFilters=6
Unable to set the 'NumFilters' property of class
'Convolution2DLayer' because it is read-only. 

Can anyone offer any suggestions? I feel like I'm missing a simple step.

Thank you!

Answer 1

[Yash Srivastava]

It is my understanding that you are trying to change the properties of a convolution layer in a neural network after exporting it from Deep Network Designer.

The properties of layer cannot be changed once they are created. As a work-around to this you can create a new convolution layer with the desired number of filters and use the “replaceLayer” function to add it to the graph.

Please refer to the https://in.mathworks.com/help/deeplearning/ref/layergraph.replacelayer.html documentation for information on how to use “replaceLayer” function.

[Matt K.]

Thanks for the reply. This makes sense.

Answer 2

[Sina Alizad]

use this trick

1-save to a temp net

2-change props in the tmp net

3-load back and assemble the network

1)

tmp_net = lgraph.saveobj;

2)

tmp_net.Layers(2,1).Weights = w1;

tmp_net.Layers(2,1).Bias = b1;

3)

convnet = lgraph.loadobj(tmp_net);

convnet=assembleNetwork(convnet);