在densenet中添加se模块

在DenseNet中添加SE（Squeeze-and-Excitation）模块可以通过以下步骤实现：

1. 定义SE模块函数：

import tensorflow as tf

def se_block(input_tensor, compress_rate=16):
    # 获取输入张量的形状
    input_shape = input_tensor.get_shape().as_list()

    # 计算通道数
    num_channels = input_shape[-1]

    # 压缩通道数
    reduced_channels = num_channels // compress_rate

    # Squeeze操作：全局平均池化
    squeeze = tf.reduce_mean(input_tensor, [1, 2])

    # Excitation操作：全连接层
    excitation = tf.layers.dense(squeeze, reduced_channels, activation=tf.nn.relu)
    excitation = tf.layers.dense(excitation, num_channels, activation=tf.nn.sigmoid)

    # Reshape张量
    excitation = tf.reshape(excitation, [-1, 1, 1, num_channels])

    # Scale操作：原始张量与Excitation张量相乘
    scale = input_tensor * excitation

    return scale

2. 在DenseBlock中应用SE模块：

def dense_block(input_tensor, num_layers, growth_rate, bottleneck_width=4):
    # 存储所有密集块中的特征映射
    features = [input_tensor]

    # 创建num_layers个密集层
    for i in range(num_layers):
        # 计算Bottleneck层的输出通道数
        bottleneck_channels = growth_rate * bottleneck_width

        # BN-ReLU-Conv1x1-BN-ReLU-Conv3x3
        x = batch_norm_relu(features[-1])
        x = conv2d(x, bottleneck_channels, kernel_size=1)
        x = batch_norm_relu(x)
        x = conv2d(x, growth_rate, kernel_size=3)

        # 应用SE模块
        x = se_block(x)

        # 将当前层的输出添加到特征列表中
        features.append(x)

    # 将所有特征映射串联在一起
    output_tensor = tf.concat(features, axis=-1)

    return output_tensor

注意：在应用SE模块时，应该添加在密集块的最后一个卷积层之后。