用pytorch实现基于注意力机制的特征融合

以下是基于注意力机制的特征融合的PyTorch实现示例：

import torch
import torch.nn as nn

class Attention(nn.Module):
    def __init__(self, input_size):
        super(Attention, self).__init__()
        self.input_size = input_size
        self.linear = nn.Linear(input_size, input_size)
        self.softmax = nn.Softmax(dim=1)

    def forward(self, input):
        # input shape: (batch_size, seq_len, input_size)
        energy = self.linear(input)
        # energy shape: (batch_size, seq_len, input_size)
        energy = torch.tanh(energy)
        # energy shape: (batch_size, seq_len, input_size)
        attention = self.softmax(energy.sum(dim=2, keepdim=True))
        # attention shape: (batch_size, seq_len, 1)
        weighted_input = torch.mul(input, attention)
        # weighted_input shape: (batch_size, seq_len, input_size)
        weighted_sum = weighted_input.sum(dim=1)
        # weighted_sum shape: (batch_size, input_size)
        return weighted_sum

class FeatureFusion(nn.Module):
    def __init__(self, input_size1, input_size2, hidden_size):
        super(FeatureFusion, self).__init__()
        self.linear1 = nn.Linear(input_size1, hidden_size)
        self.linear2 = nn.Linear(input_size2, hidden_size)
        self.attention = Attention(hidden_size)

    def forward(self, input1, input2):
        # input1 shape: (batch_size, seq_len1, input_size1)
        # input2 shape: (batch_size, seq_len2, input_size2)
        hidden1 = self.linear1(input1)
        # hidden1 shape: (batch_size, seq_len1, hidden_size)
        hidden2 = self.linear2(input2)
        # hidden2 shape: (batch_size, seq_len2, hidden_size)
        fused_hidden = torch.cat((hidden1, hidden2), dim=1)
        # fused_hidden shape: (batch_size, seq_len1 + seq_len2, hidden_size)
        fused_hidden = self.attention(fused_hidden)
        # fused_hidden shape: (batch_size, hidden_size)
        return fused_hidden

上述代码中，Attention类实现了注意力机制，FeatureFusion类实现了两个特征的融合。在Attention类中，首先通过一个线性层将输入的特征映射到一个新的空间，然后通过tanh函数将其压缩到[-1, 1]的范围内，再通过softmax函数计算每个特征的权重，最后将输入的特征和权重相乘得到加权和。在FeatureFusion类中，首先将两个特征通过线性层映射到相同的维度，然后将它们拼接在一起，再通过Attention类得到融合后的特征。