How to declare the weight and bias values for a convolution layer?

Question

Fahmi Akmal Dzulkifli 2023-1-18

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1896240-how-to-declare-the-weight-and-bias-values-for-a-convolution-layer

编辑： Manish 2024-8-27

在 MATLAB Online 中打开

Greetings,

I want to add a convolution layer in the existing squeezenet network. However the errors shown like this:

Error using assembleNetwork (line 47)

Invalid network.

Error in trainyolov3 (line 80)

newbaseNetwork = assembleNetwork(lgraph); % for tiny-yolov3-coco

Caused by:

Layer 'add_conv': Empty Weights property. Specify a nonempty value for the Weights property.

Layer 'add_conv': Empty Bias property. Specify a nonempty value for the Bias property.

Therefore, I want to ask on how to declare the weights and bias value for that layer? My code is as shown below:

lgraph = disconnectLayers(lgraph,'fire8-concat','fire9-squeeze1x1');
layer = [
    maxPooling2dLayer([3 3],"Name","pool6","Padding","same","Stride",[2 2])
    convolution2dLayer([1 1],512,"Name","add_conv","Padding",[1 1 1 1],"Stride",[2 2])
    reluLayer("Name","relu_add_conv")];
lgraph = addLayers(lgraph, layer);
lgraph = connectLayers(lgraph,'fire8-concat','pool6');
lgraph = connectLayers(lgraph,'relu_add_conv','fire9-squeeze1x1');
newbaseNetwork = assembleNetwork(lgraph); 

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Manish 2024-8-27

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1896240-how-to-declare-the-weight-and-bias-values-for-a-convolution-layer#answer_1505511

编辑：Manish 2024-8-27

在 MATLAB Online 中打开

Hi,

It seems the error you're encountering is due to the weights and biases not being initialized.

You can verify this by inspecting the layer properties using “layer(2).Weights.” In the code provided, the “Weights” and “Biases” are uninitialized.

To resolve this issue, you can explicitly initialize the weights and biases and apply them when defining the “add_conv” layer.

Refer to the code snippet below that demonstrates how to initialize weights and biases:

% Load the network and convert to layer graph
net = squeezenet('Weights', 'imagenet');
lgraph = layerGraph(net);
% Disconnect the specified layers
lgraph = disconnectLayers(lgraph, "fire8-concat", "fire9-squeeze1x1");
% Define the input channels based on the architecture
inputChannels = 512; % This should match the number of output channels from the previous layer
% Initialize weights and biases
weights = randn([1, 1, inputChannels, 512]) * 0.01; % Random initialization
bias = zeros([1, 1, 512]); % Zero initialization
% Define the new layers with initialized weights and biases
layer = [
    maxPooling2dLayer([3 3], "Name", "pool6", "Padding", "same", "Stride", [2 2])
    convolution2dLayer([1 1], 512, "Name", "add_conv", "Padding", [1 1 1 1], "Stride", [2 2], ...
                       'Weights', weights, 'Bias', bias)
    reluLayer("Name", "relu_add_conv")
];
% Add the new layers to the graph
lgraph = addLayers(lgraph, layer);
% Connect the layers
lgraph = connectLayers(lgraph, 'fire8-concat', 'pool6');
lgraph = connectLayers(lgraph, 'relu_add_conv', 'fire9-squeeze1x1');
% Assemble the network
newbaseNetwork = assembleNetwork(lgraph);

For more information on initializing the weights and biases explicitly, you can refer to the “Parameters and Initialization” section in the below mentioned documentation:

https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.convolution2dlayer.html

Hope it helps!