Weight Tying for Layers in a CNN model

3 views (last 30 days)
Junaid Farooq
Junaid Farooq on 16 Jun 2024
Commented: Samuel Somuyiwa on 18 Jun 2024
I am building a Deep Learning network looks like the figure below. I want to tie the weights of : Conv1 to Conv2; Conv3 to Conv6 and Conv4 to Conv5. Is it possible to do so in MATLAB? If yes, please help me with the code.
  2 Comments
Junaid Farooq
Junaid Farooq on 16 Jun 2024
I tried this:
imageSize = [31 31 1];
encoderLayer1 = [
dlhdl.layer.sliceLayer(Name="slice",Groups=2,GroupID=2)
convolution2dLayer(3,32,"Padding",1,"WeightsInitializer","he","Name","conv1")
convolution2dLayer(3,32,"Padding",1,"WeightsInitializer","he","Name","conv3")
additionLayer(2,"Name","add");
];
encoderLayer2 = [
convolution2dLayer(3,32,"Padding",1,"WeightsInitializer","he","Name","conv2")];
layers = [imageInputLayer(inputSize, Normalization="none")
weightTyingEncoderLayer1(encoderLayer1,encoderLayer2)];
net = dlnetwork(layers);
Junaid Farooq
Junaid Farooq on 16 Jun 2024
It gives following error:
Error using dlnetwork/initialize (line 558)
Invalid network.
Error in dlnetwork (line 167)
net = initialize(net, dlX{:});
Caused by:
Layer 'recursiveNet': Error using the predict function in layer weightTyingEncoderLayer1. The
function threw an error and could not be executed.
Error using convolution2dLayer (line 1)
Invalid argument at position 3. Function requires exactly 2 positional input(s).
Error in weightTyingEncoderLayer1/predict (line 36)
Y1 = convolution2dLayer(Y,this.Encoder.Learnables.Value{1}',this.Conv1);

Sign in to comment.

Answers (2)

Sanjana
Sanjana on 17 Jun 2024
Hi Junaid,
Weight Tying is supported in MATLAB using Nested Layer. I have observed in your code for the "WeightTyingEncoderLayer1.m", that you have not declared and defined the "intialize" function, which is necessary initializing the learnables and parameters of the layers whose weights need to be tied, based on the "numChannels" of the layers chosen for this purpose.
The Error using "convolution2dLayer", is due to the reason that "this.Conv1" is not initialized.
To resolve the errors and implement Weight Tying Using Nested Layer, refer to the following example,
Also, refer to the following link for more information on creating custom "Nested Layer", and specifically refer to the "Custom Layer Template" section to understand how to create a custom nested layer in MATLAB.
  1 Comment
Junaid Farooq
Junaid Farooq on 17 Jun 2024
Thank you Sanjana. I already read the articles provided in links. I would appreciate if you can help me out by writing a code for the diagram given above.

Sign in to comment.


Samuel Somuyiwa
Samuel Somuyiwa on 17 Jun 2024
See attached weightTyingAutoEncoder layer example. The layer follows on from the example in the link that Sanjana shared earlier.
See an example of how to use the layer below. Note that I couldn't understand what dlhdl.layer.sliceLayer in your example was meant to do or whether it was used correctly in that example. So, I have not included it in the example below
inputSize = [32 32 32];
layers = [
imageInputLayer(inputSize,Normalization="none")
weightTyingAutoEncoderLayer
];
net = dlnetwork(layers);
%%
X = dlarray(rand(inputSize),'SSCB');
Y = predict(net, X);
  2 Comments
Junaid Farooq
Junaid Farooq on 17 Jun 2024
Thanks. As per your code will Conv1 and Conv3 get the input from Input Layer.
If we add Relu in between Convolution layers will the values of below code change:
decoder.Learnables.Value{2} = dlarray(this.DecoderBias1);
decoder.Learnables.Value{4} = dlarray(this.DecoderBias2);
decoder.Learnables.Value{6} = dlarray(this.DecoderBias3);
Samuel Somuyiwa
Samuel Somuyiwa on 18 Jun 2024
Yes, the convolution layers get the input from the input layer.
No, adding RELU layer between the convolution layers will not change the values of the code.

Sign in to comment.

Categories

Find more on Image Data Workflows in Help Center and File Exchange

Products


Release

R2024a

Community Treasure Hunt

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

Start Hunting!