将cnn与lstm连接需要注意什么 lstm和cnn结合文本

转载

mob64ca14101b2f 2024-05-30 09:52:14

文章标签 将cnn与lstm连接需要注意什么 lstm 分类 cnn Max 文章分类 机器学习人工智能

⛄ 内容介绍

一种基于长短时记忆网络和卷积神经网络的文本分类方法,首先,利用词向量将输入文本进行向量表示,通过三层CNN提取文本的局部特征,进而整合出全文语义,同时,使用LSTM存储文本序列中历史信息的特征,以获取文本的上下文依赖关系,其次,将输入向量分别与各层CNN的输出相融合,缓解深层神经网络中层与层之间特征传递时出现的特征丢失问题.本发明适用于文本分类,具有运行效率高,运行时间短的优点,解决了传统分类技术在处理复杂抽象和上下文的强相关性文本时的缺陷.

将cnn与lstm连接需要注意什么 lstm和cnn结合文本_分类

将cnn与lstm连接需要注意什么 lstm和cnn结合文本_lstm_02

将cnn与lstm连接需要注意什么 lstm和cnn结合文本_将cnn与lstm连接需要注意什么_03

将cnn与lstm连接需要注意什么 lstm和cnn结合文本_lstm_04

将cnn与lstm连接需要注意什么 lstm和cnn结合文本_分类_05

将cnn与lstm连接需要注意什么 lstm和cnn结合文本_将cnn与lstm连接需要注意什么_06

⛄ 部分代码

% Max_iter: Maximum iterations, N: Populatoin size, Convergence_curve: Convergence curve
rand('seed',sum(100 * clock)); % Random number seed
z = 0.03; % Adjustable parameter
% Initialize the position of slime mould
lb = ones(1,dim).*lb; % Lower boundary 
ub = ones(1,dim).*ub; % Upper boundary
X = initialization(N,dim,ub,lb); % It can be downloaded from https://github.com/Shihong-Yin
bestPositions = zeros(1,dim); % Optimal food source location
Destination_fitness = inf; % Change this to -inf for maximization problems
weight = ones(N,dim); % Fitness weight of each slime mould
Convergence_curve = zeros(1,Max_iter);
% Main loop
for it = 1:Max_iter
    % Check the boundary and calculate the fitness
    FU = X>ub;  FL = X<lb;  X = (X.*(~(FU+FL)))+ub.*FU+lb.*FL;
    PopFitness = cec21_bias_shift_rot_func(X',func_num)'; % https://github.com/Shihong-Yin
    % Sort the fitness thus update the bF and wF
    [SmellOrder,SmellIndex] = sort(PopFitness); % Eq.(2.6)
    bestFitness = SmellOrder(1);
    worstFitness = SmellOrder(N);
    S = bestFitness-worstFitness+eps; % Plus eps to avoid denominator zero
    % Calculate the fitness weight of each slime mould
    for i = 1:N
        if i <= N/2 % Eq.(2.5)
            weight(SmellIndex(i),:) = 1+rand(1,dim)*log10((bestFitness-SmellOrder(i))/S+1);
        else
            weight(SmellIndex(i),:) = 1-rand(1,dim)*log10((bestFitness-SmellOrder(i))/S+1);
        end
    end
    % Update the best position and destination fitness
    if bestFitness < Destination_fitness
        bestPositions = X(SmellIndex(1),:);
        Destination_fitness = bestFitness;
    end
    a = atanh(-(it/Max_iter)+1); % Eq.(2.4)
    vb = unifrnd(-a,a,N,dim); % Eq.(2.3)
    b = 1-it/Max_iter;
    vc = unifrnd(-b,b,N,dim);
    p = tanh(abs(PopFitness-Destination_fitness)); % Eq.(2.2)
    r = rand(N,dim);
    A = randi([1,N],N,dim); % Two positions randomly selected from population
    B = randi([1,N],N,dim);
    % Update the Position of search agents
    for i = 1:N
        if rand < z % Eq.(2.7)
            X(i,:) = (ub-lb).*rand(1,dim)+lb; % The original code is (ub-lb)*rand+lb;
        else
            for j = 1:dim
                if r(i,j) < p(i) % Eq.(2.1)
                    X(i,j) = bestPositions(j)+vb(i,j)*(weight(i,j)*X(A(i,j),j)-X(B(i,j),j));
                else
                    X(i,j) = vc(i,j)*X(i,j);
                end
            end
        end
    end
    Convergence_curve(it) = Destination_fitness;
end
end