硬注意力与软注意力

软注意力机制就是虽然词语权重不同，但是在训练模型的时候雨露均沾，每个词语都用到，焦点词语的权重大。软性注意力（Soft Attention）机制是指在选择信息的时候，不是从N个信息中只选择1个，而是计算N个输入信息的加权平均，再输入到神经网络中计算。

硬注意力机制是从存储的多个信息中只挑出一条信息来，可能是概率最大的那个词向量

代码：

（一）软注意力

class SpatialAttn(nn.Module):
    """Spatial Attention (Sec. 3.1.I.1)"""

    def __init__(self):
        super(SpatialAttn, self).__init__()
        self.conv1 = ConvBlock(1, 1, 3, s=2, p=1)
        self.conv2 = ConvBlock(1, 1, 1)

    def forward(self, x):
        # global cross-channel averaging
        x = x.mean(1, keepdim=True) # 由hwc 变为 hw1
        # 3-by-3 conv
        h = x.size(2)
        x = self.conv1(x)
        # bilinear resizing
        x = F.upsample(x, (h,h), mode='bilinear', align_corners=True)
        # scaling conv
        x = self.conv2(x)
        return x
        ## 返回的是h*w*1 的 soft map

class ChannelAttn(nn.Module):

    """通道注意力机制"""

    def __init__(self, in_channels, reduction_rate=16):

        super(ChannelAttn, self).__init__()

        assert in_channels%reduction_rate == 0

        self.conv1 = ConvBlock(in_channels, in_channels // reduction_rate, 1)

        self.conv2 = ConvBlock(in_channels // reduction_rate, in_channels, 1)

    def forward(self, x):

        # 压缩操作（全局平均池化）
        x = F.avg_pool2d(x, x.size()[2:])

        # 激励操作（2个卷积层）
        x = self.conv1(x)
        x = self.conv2(x)

        return x

'''
空间和通道上的attention 融合
就是空间和通道上的attention做一个矩阵乘法
'''

class SoftAttn(nn.Module):
    """Soft Attention (Sec. 3.1.I)

    Aim: Spatial Attention + Channel Attention

    Output: attention maps with shape identical to input.
    """

    def __init__(self, in_channels):
        super(SoftAttn, self).__init__()
        self.spatial_attn = SpatialAttn()
        self.channel_attn = ChannelAttn(in_channels)
        self.conv = ConvBlock(in_channels, in_channels, 1)

    def forward(self, x):

        y_spatial = self.spatial_attn(x) # 空间注意力输出
        y_channel = self.channel_attn(x) # 通道注意力输出
        y = y_spatial * y_channel # 空间注意力和通道注意力相乘
        y = torch.sigmoid(self.conv(y)) # 卷积块输出
        return y

（二）硬注意力

class HardAttn(nn.Module):
    """Hard Attention (Sec. 3.1.II)"""

    def __init__(self, in_channels):
        super(HardAttn, self).__init__()
        self.fc = nn.Linear(in_channels, 4*2)
        self.init_params()

    def init_params(self):
        self.fc.weight.data.zero_()
        # 初始化参数
        # if x_t = 0 the performance is very low
        self.fc.bias.data.copy_(torch.tensor([0.3, -0.3, 0.3, 0.3, -0.3, 0.3, -0.3, -0.3], dtype=torch.float))

    def forward(self, x):
        # squeeze operation (global average pooling)
        x = F.avg_pool2d(x, x.size()[2:]).view(x.size(0), x.size(1))
        # predict transformation parameters
        theta = torch.tanh(self.fc(x))
        theta = theta.view(-1, 4, 2)
        return theta
         # 返回的是 2T T为区域数量。因为尺度会固定。所以只要学位移的值

# 并不太理解，只是记录一下

原文链接：https://blog.csdn.net/qq_40905284/article/details/130726310

posted @ 2023-09-20 10:03 逆袭成功的swift 阅读(396) 评论(0) 收藏举报

刷新页面返回顶部

硬注意力与软注意力

公告