基于Convolutional Block Attention Module (CBAM)的Multi-Attention模型设计与实现

客官°小女子只卖身不卖艺 2022-09-05 12:38 114阅读 0赞

本文主要介绍的 Multi-Attention 方法只是一个研究思路。只是尝试了在基本模型的顶部集成多注意力机制。

其Multi-Attention模型结构如下所示：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MDY1MTUxNQ_size_16_color_FFFFFF_t_70][]

模型本质上是并行添加了 CBAM 和 DeepMoji 注意力机制，并在最后将它们的特征进行合并。 此外，我们通过集成全局加权平均池 (GWAP) 方法，对 CBAM 机制及其空间模块的末尾部分进行了修改。

经过对该模型的输出，我们将Output特征放入全连接的神经网络模型中进行最终的模型训练。

其结构图如下：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MDY1MTUxNQ_size_16_color_FFFFFF_t_70 1][]

为了验证模型是否能够运行，我们在mnist手写汉字数据集上进行了实验。其训练主函数代码为：

image_size = 71
    def resize(mnist):
         train_data = []
         for img in mnist:
                resized_img = cv2.resize(img, (image_size, image_size))
                train_data.append(resized_img)
         return train_data
    (xtrain, train_target), (_, _) = tf.keras.datasets.mnist.load_data()
    
    # resize and prepare the input 
    xtrain = resize(xtrain)
    xtrain = np.expand_dims(xtrain, axis=-1)
    xtrain = np.repeat(xtrain, 3, axis=-1)
    xtrain = xtrain.astype('float32') / 255
    
    # train set / target 
    ytrain = tf.keras.utils.to_categorical(train_target, num_classes=10)
    class Classifier(tf.keras.Model):
        def __init__(self, dim):
            super(Classifier, self).__init__()
            # Defining All Layers in __init__
            # Layer of Block
            self.Base  = tf.keras.applications.Xception(
                input_shape=(image_size, image_size, 3),
                weights=None,
                include_top=False)
            # Keras Built-in
            self.GAP1 = tf.keras.layers.GlobalAveragePooling2D()
            self.GAP2 = tf.keras.layers.GlobalAveragePooling2D()
            self.BAT  = tf.keras.layers.BatchNormalization()
            self.ADD  = tf.keras.layers.Add()
            self.AVG  = tf.keras.layers.Average()
            self.DROP = tf.keras.layers.Dropout(rate=0.5)
            # Customs
            self.CAN  = ChannelAttentionModule()
            self.SPN1 = SpatialAttentionModule()
            self.SPN2 = SpatialAttentionModule()
            self.AWG  = AttentionWeightedAverage2D()
            # Tail
            self.DENS = tf.keras.layers.Dense(512, activation=tf.nn.relu)
            self.OUT  = tf.keras.layers.Dense(10,  activation='softmax', dtype=tf.float32)
        
        def call(self, input_tensor, training=False):
            # Base Inputs
            x  = self.Base(input_tensor)
            # Attention Modules 1
            # Channel Attention + Spatial Attention 
            canx   = self.CAN(x)*x
            spnx   = self.SPN1(canx)*canx
            # Global Weighted Average Poolin
            gapx   = self.GAP1(spnx)
            wvgx   = self.GAP2(self.SPN2(canx))
            gapavg = self.AVG([gapx, wvgx])
            # Attention Modules 2
            # Attention Weighted Average (AWG)
            awgavg = self.AWG(x)
            # Summation of Attentions
            x = self.ADD([gapavg, awgavg])
            # Tails
            x = self.BAT(x)
            x = self.DENS(x)
            x  = self.DROP(x, training=training)
            return self.OUT(x)
       
        def build_graph(self):
            x = tf.keras.layers.Input(shape=(image_size, image_size, 3))
            return tf.keras.Model(inputs=[x], outputs=self.call(x))
    
    model.compile( metrics=['accuracy'],loss = 'categorical_crossentropy', optimizer = 'adam')
    model.fit(xtrain, ytrain, epochs=5)

训练过程如下：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MDY1MTUxNQ_size_16_color_FFFFFF_t_70 2][]

完整代码可下载：[https://download.csdn.net/download/weixin\_40651515/21121787][https_download.csdn.net_download_weixin_40651515_21121787]

模型运行环境：python3.7.6、tensorflow==2.2.0、keras==2.3.1等。

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MDY1MTUxNQ_size_16_color_FFFFFF_t_70]: /images/20220829/dfec6a6948784974a484d13ca9326533.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MDY1MTUxNQ_size_16_color_FFFFFF_t_70 1]: /images/20220829/ab4c634f6d8a4e51a31ebe12a48b9467.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MDY1MTUxNQ_size_16_color_FFFFFF_t_70 2]: /images/20220829/8fc39ecb85bf46f49021c8f915259b8f.png
[https_download.csdn.net_download_weixin_40651515_21121787]: https://download.csdn.net/download/weixin_40651515/21121787