举两个例子分别关于全局注意力机制和局部注意力机制，并将代码注释进行标注

以下是两个例子分别展示了全局注意力机制和局部注意力机制的示例代码：

1. 全局注意力机制的示例代码：

import torch
import torch.nn as nn

class GlobalAttention(nn.Module):
    def __init__(self, input_size, hidden_size):
        super(GlobalAttention, self).__init__()
        self.linear_in = nn.Linear(input_size, hidden_size)
        self.linear_out = nn.Linear(hidden_size, 1)

    def forward(self, inputs):
        # 输入inputs的形状：(batch_size, sequence_length, input_size)
        batch_size, sequence_length, input_size = inputs.size()

        # 对输入进行线性变换
        transformed_inputs = self.linear_in(inputs)  # (batch_size, sequence_length, hidden_size)

        # 计算权重
        weights = self.linear_out(torch.tanh(transformed_inputs))  # (batch_size, sequence_length, 1)
        weights = torch.softmax(weights, dim=1)  # 注意力权重归一化

        # 加权求和
        context = torch.bmm(weights.transpose(1, 2), inputs)  # (batch_size, 1, input_size)
        context = context.squeeze(1)  # (batch_size, input_size)

        return context

# 使用全局注意力机制的示例
input_size = 100
hidden_size = 50
sequence_length = 10
batch_size = 32

inputs = torch.randn(batch_size, sequence_length, input_size)
attention = GlobalAttention(input_size, hidden_size)
output = attention(inputs)
print(output.shape)  # 输出形状：(32, 100)

2. 局部注意力机制的示例代码：

import torch
import torch.nn as nn

class LocalAttention(nn.Module):
    def __init__(self, input_size, hidden_size, window_size):
        super(LocalAttention, self).__init__()
        self.linear_in = nn.Linear(input_size, hidden_size)
        self.linear_out = nn.Linear(hidden_size, 1)
        self.window_size = window_size

    def forward(self, inputs):
        # 输入inputs的形状：(batch_size, sequence_length, input_size)
        batch_size, sequence_length, input_size = inputs.size()

        # 对输入进行线性变换
        transformed_inputs = self.linear_in(inputs)  # (batch_size, sequence_length, hidden_size)

        # 计算局部注意力权重
        weights = self.linear_out(torch.tanh(transformed_inputs))  # (batch_size, sequence_length, 1)
        weights = torch.softmax(weights, dim=1)  # 注意力权重归一化

        # 根据窗口大小调整注意力权重
        center = sequence_length // 2
        start = max(0, center - self.window_size // 2)
        end = min(sequence_length, center + self.window_size // 2 + 1)
        weights[:, :start] = 0
        weights[:, end:] = 0
        weights = weights / weights.sum(dim=1, keepdim=True)  # 重新归一化

        # 加权求和
        context = torch.bmm(weights.transpose(1, 2), inputs)  # (batch_size, 1, input_size)
        context = context.squeeze(1)  # (batch_size, input_size)

        return context

# 使用局部注意力机制的示例
input_size = 100
hidden_size = 50
sequence_length = 10
window_size = 3
batch_size = 32

inputs = torch.randn(batch_size, sequence_length, input_size)
attention = LocalAttention(input_size, hidden_size, window_size)
output = attention(inputs)
print(output.shape)  # 输出形状：(32, 100)

以上代码仅为示例，实际应用中可能需要根据具体任务和需求进行调整和修改。注意力权重的计算和调整方式也可以根据具体的局部或全局关注策略进行调整。