CBAM通道注意力的是tensorflow2.xs

matlab的旋律

CBAM通道注意力的是tensorflow2.x,  参考https://blog.csdn.net/weixin_39122088/article/details/10719197

from tensorflow.keras import layers, regularizers

import tensorflow as tf

1. #实现方式1
   
2. class ChannelAttention(layers.Layer):
   
3.         def __init__(self, in_planes, ratio=8):
   
4.              super(ChannelAttention, self).__init__()
   
5. 
   
6.             self.avg_out= layers.GlobalAveragePooling2D()
   
7.             self.max_out= layers.GlobalMaxPooling2D()
   
8. 
   
9.             self.fc1 = layers.Dense(in_planes//ratio, kernel_initializer='he_normal',
   
10.             kernel_regularizer=regularizers.l2(5e-4),
    
11.             activation=tf.nn.relu,
    
12.            use_bias=True, bias_initializer='zeros')
    
13.            self.fc2 = layers.Dense(in_planes, kernel_initializer='he_normal',
    
14.            kernel_regularizer=regularizers.l2(5e-4),
    
15.            use_bias=True, bias_initializer='zeros')
    
16. 
    
17. def call(self, inputs):
    
18.      avg_out = self.avg_out(inputs)
    
19.      max_out = self.max_out(inputs)
    
20.      out = tf.stack([avg_out, max_out], axis=1) # shape=(None, 2, fea_num)
    
21.      out = self.fc2(self.fc1(out))
    
22.      out = tf.reduce_sum(out, axis=1) # shape=(256, 512)
    
23.      out = tf.nn.sigmoid(out)
    
24.      out = layers.Reshape((1, 1, out.shape[1]))(out)
    
25. 
    
26.    return out

复制代码

1. #实现方式2
   
2. class ChannelAttention(layers.Layer):
   
3.      def __init__(self, in_planes):
   
4.          super(ChannelAttention, self).__init__()
   
5. 
   
6.          self.avg= layers.GlobalAveragePooling2D()
   
7.          self.max= layers.GlobalMaxPooling2D()
   
8. 
   
9.          self.fc1 = layers.Dense(in_planes//16, kernel_initializer='he_normal', activation='relu',
   
10.          use_bias=True, bias_initializer='zeros')
    
11.          self.fc2 = layers.Dense(in_planes, kernel_initializer='he_normal', use_bias=True,
    
12.          bias_initializer='zeros')
    
13. 
    
14.    def call(self, inputs):
    
15.         avg_out = self.fc2(self.fc1(self.avg(inputs)))
    
16.         max_out = self.fc2(self.fc1(self.max(inputs)))
    
17.         out = avg_out + max_out
    
18.         out = tf.nn.sigmoid(out)
    
19.         out = tf.reshape(out, [out.shape[0], 1, 1, out.shape[1]])
    
20.         out = tf.tile(out, [1, inputs.shape[1], inputs.shape[2], 1])
    
21. 
    
22.       return out

复制代码

空间注意力，参考https://blog.csdn.net/mqdlff_python/article/details/135119610

1. 
   
2. class SpatialAttention(layers.Layer):
   
3.       def __init__(self):
   
4.            super(SpatialAttention, self).__init__()
   
5.            self.conv =layers.Conv2D(filters=1, kernel_size=3, padding='same', activation='sigmoid')
   
6. 
   
7.      def call(self, inputs):
   
8.           max_pool = reduce_max(inputs, axis=-1, keepdims=True)
   
9.            spatial_attention = self.conv(max_pool)
   

复制代码