Question

我已定义了一些图层，我想手动设置此网络的权重和偏差。为此，我在Matlab中编写了一个代码，如下所示：

%% Defining Layers
layers = [imageInputLayer([28 28 1],'Normalization','none');
convolution2dLayer(5,20);
reluLayer();
maxPooling2dLayer(2,'Stride',2);
fullyConnectedLayer(10);
softmaxLayer();];
%% Set Manually Weights and Biases
layers(2,1).Weights = Conv1_Weights ;
layers(2,1).Bias = CONV1_Biases  ;
layers(5,1).Weights = FC1_Weights ;
layers(5,1).Bias = FC2_Bias ;
net = SeriesNetwork(layers);
%% 
X = imread(____)
X_Single = single(X);
X_Single_Normalize = (X_Single-min(X_Single(:)))/(max(X_Single(:))-min(X_Single(:)));
F1 = activations(net,X_Single_Normalize,2,'OutputAs','channels'); 
F2 = activations(net,X_Single_Normalize,3,'OutputAs','channels'); 
F3 = activations(net,X_Single_Normalize,4,'OutputAs','channels'); 
F4 = activations(net,X_Single_Normalize,5,'OutputAs','channels');

实际上，我收到了这个错误：

Error using nnet.internal.cnn.layer.FullyConnected/forwardPropagateSize  (line 101)
An input size for the layer must be defined in order to call forwardPropagateSize.

Error in SeriesNetwork>iDetermineLayerOutputSize (line 952)
inputSize = layers{i}.forwardPropagateSize(inputSize);

Error in SeriesNetwork/activations (line 775)
            outputSize = iDetermineLayerOutputSize(this.PrivateNetwork.Layers,
            layerID, inputSize );

Error in untitled1 (line 21)
F4 = activations(net,X_Single_Normalize,5,'OutputAs','channels');

任何人都可以帮我解决这个问题吗？

Answer 1

问题在于，从R2017b开始，NNT中的SeriesNetwork目前没有“最终确定”网络，以便为推理做好准备。

您可以将repro步骤简化为：

>> layers = [imageInputLayer([28 28 1],'Normalization','none');
convolution2dLayer(5,20);
reluLayer();
maxPooling2dLayer(2,'Stride',2);
fullyConnectedLayer(10);
softmaxLayer();];
>>
>> net = SeriesNetwork(layers);
>> 
>> A = rand(28,28);

>> activations(net,A,5);

作为一种解决方法，您需要使用 trainNetwork 来至少训练一次迭代。由此产生的网络应该准备好进行培训/推理。您可以使用可学习的参数在每个图层上设置学习率，以避免培训实际对您的体重初始化进行任何更新。

function example

layers = [imageInputLayer([28 28 1],'Normalization','none'),...
convolution2dLayer(5,20),...
reluLayer(),...
maxPooling2dLayer(2,'Stride',2),...
fullyConnectedLayer(10),...
softmaxLayer(),...
classificationLayer];

A = rand(28,28);

net = constructInitializedNetwork(layers);
activations(net,A,5);

function net = constructInitializedNetwork(layers)

X = rand(28,28);
Y = categorical(1,1:10);

layers = freezeLayers(layers);
options = 
trainingOptions('sgdm','MaxEpochs',1,'InitialLearnRate',eps);
net = trainNetwork(X,Y,layers,options);


function layers = freezeLayers(layers)

for idx = 1:length(layers)
    if isprop(layers(idx),'WeightLearnRateFactor')
       layers(idx).WeightLearnRateFactor = 0;
    end

    if isprop(layers(idx),'BiasLearnRateFactor')
       layers(idx).BiasLearnRateFactor = 0; 
    end
end

如何在Matlab中设置卷积神经网络的手动权重和偏差

1 个答案: