LeNet分类单变量时间序列报错：数据格式化问题排查求助

问题描述 #

尝试用LeNet对含300时间步的单变量时间序列做分类，遇到预测与标签维度不匹配的错误，先后试过矩阵格式和单元格数组格式的输入，均报错。对channel和feature的概念混淆，不确定序列输入层的参数设置是否正确。

原代码 #

num_channels = 1;
num_classes = 3;
filterSize = 5;
numFilters = 32;
num_of_features = size(X_train(1, :), 2);
assert(num_of_features==300);

net = dlnetwork;

LeNeT = [
    % featureInputLayer(num_of_features, "Name", "input")
    sequenceInputLayer(num_of_features, "Name", "input", "MinLength", num_of_features)
    convolution1dLayer( 5, 6,"Name","conv1")
    tanhLayer("Name","tanh1")
    averagePooling1dLayer( 2, "Name","pool1","Stride", 2)
    tanhLayer("Name","tanh2")
    convolution1dLayer( 5, 16,"Name","conv2")
    tanhLayer("Name","tanh3")
    averagePooling1dLayer( 2,"Name","pool2","Stride", 2)
    tanhLayer("Name","tanh4")
    convolution1dLayer( 5, 120,"Name","conv3")
    tanhLayer("Name","tanh5")
    fullyConnectedLayer( 84, "Name","fc1")
    tanhLayer("Name","tanh6")
    fullyConnectedLayer( num_classes, "Name","new_fc")
    softmaxLayer("Name","prob")];

net = addLayers(net,LeNeT);
net = initialize(net);

train_opts = trainingOptions("adam", ...
    MaxEpochs=20, ...
    MiniBatchSize=128, ... 
    InitialLearnRate=0.01, ...
    Shuffle="once", ...
    ValidationData={X_val,Y_val}, ...
    Plots="none", ...
    Metrics="accuracy", ...
    Verbose=true);

%% Train the network.
net = trainnet(X_train, Y_train, net, "crossentropy", train_opts);

%% Test the network.
accuracy = testnet(net, X_test, Y_test, "accuracy");
scores = minibatchpredict(net, X_test);
predicted_labels = scores2label(scores, categories(Y));
figure
confusionchart(Y_test, predicted_labels)

数据维度 #

• X_train = 11788 x 300 double（行=样本，列=时间步）
• Y_train = 11788 x 1 categorical（行=样本，3类标签）
• X_val = 5894 x 300 double
• Y_val = 5894 x 1 categorical
• X_test = 5895 x 300 double
• Y_test = 5895 x 1 categorical

报错信息 #

Training stopped: Error occurred
Error using trainnet (line 54)
Error evaluating loss function.

Error in main (line 158)
net = trainnet(X_train, Y_train, net, "crossentropy", train_opts);
      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Caused by:
    Error using validateTrueValues (line 54)
    Size of predictions and targets must match.
    Size of predictions:
        3(C) × 1(B) × 2940(T)
    Size of targets:
        3(C) × 1(B) × 11788(T)

核心问题分析 #

1. 特征与时间步混淆：单变量时间序列中，每个时间步仅含1个特征值，因此特征数=1，时间步=300。原代码将时间步300设为sequenceInputLayer的inputSize，完全颠倒了两者的关系。
2. 网络输出维度不匹配：序列分类任务需要将卷积后的序列维度压缩为单值（每个样本对应一个分类结果），原网络未添加池化层压缩序列维度，导致输出保留了冗余的序列长度，与标签维度（每个样本对应一个标签）不匹配。
3. 数据格式错误：MATLAB的1D卷积和序列输入层要求输入维度为特征数×时间步×样本数，原数据格式为样本数×时间步，未满足要求。

修复方案 #

1. 修正数据格式 #

将矩阵格式的输入转换为特征数×时间步×样本数的维度：

% 训练数据转换：1（特征数）×300（时间步）×11788（样本数）
X_train = permute(X_train, [2, 1]); % 转为300×11788
X_train = reshape(X_train, 1, 300, []); % 调整为1×300×11788

% 验证数据同理
X_val = permute(X_val, [2, 1]);
X_val = reshape(X_val, 1, 300, []);

% 测试数据同理
X_test = permute(X_test, [2, 1]);
X_test = reshape(X_test, 1, 300, []);

若使用单元格数组格式，每个元素需为时间步×特征数的向量：

X_train_cell = cell(size(X_train, 1), 1);
for i = 1:size(X_train, 1)
    X_train_cell{i} = X_train(i,:)'; % 300×1，时间步×特征数
end

2. 修正网络结构 #

调整sequenceInputLayer的inputSize为1，并添加globalAveragePooling1dLayer压缩序列维度：

LeNeT = [
    sequenceInputLayer(1, "Name", "input", "MinLength", 300) % inputSize=1（单特征）
    convolution1dLayer(5, 6,"Name","conv1")
    tanhLayer("Name","tanh1")
    averagePooling1dLayer(2, "Name","pool1","Stride", 2)
    tanhLayer("Name","tanh2")
    convolution1dLayer(5, 16,"Name","conv2")
    tanhLayer("Name","tanh3")
    averagePooling1dLayer(2,"Name","pool2","Stride", 2)
    tanhLayer("Name","tanh4")
    convolution1dLayer(5, 120,"Name","conv3")
    tanhLayer("Name","tanh5")
    globalAveragePooling1dLayer("Name", "global_pool") % 压缩序列维度为1
    fullyConnectedLayer(84, "Name","fc1")
    tanhLayer("Name","tanh6")
    fullyConnectedLayer(num_classes, "Name","new_fc")
    softmaxLayer("Name","prob")];

3. 概念澄清 #

• Feature（特征）：时间序列中每个时间步的观测值数量，单变量序列特征数为1，多变量序列特征数等于变量数。
• Channel（通道）：在1D卷积中，通道数等价于特征数，单特征对应单通道。

内容的提问来源于stack exchange，提问作者Jenő Fekete