坐标注意力机制-Tensorflow2

最新推荐文章于 2025-01-17 14:41:39 发布

早起学习晚上搬砖

最新推荐文章于 2025-01-17 14:41:39 发布

阅读量1.2k

点赞数

文章标签： tensorflow keras 深度学习

本文链接：https://blog.csdn.net/zzl18681269883/article/details/123561547

版权

本文介绍了一个名为Coordinate Attention的模型，它利用坐标敏感的注意力机制在卷积神经网络中，通过池化操作和sigmoid激活函数，实现对输入特征的空间依赖捕捉。核心代码展示了如何构造并调用该机制，适用于具有特定宽度和高度的图像特征处理。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

"""
    坐标注意力机制代码
"""
import numpy as np
from tensorflow.keras.layers import AvgPool2D, Conv2D, BatchNormalization, Activation
from tensorflow.keras import Model
from tensorflow.keras.activations import sigmoid
from tensorflow import split,concat,transpose
import tensorflow as tf



def ac_swish(x):
    temp = tf.nn.relu6(x+3) / 6
    return x * temp

class Coordinate_Attention(Model):
    def __init__(self, w, h, inp, oup, groups=32):
        """
        :param w:   width
        :param h:   height
        :param inp: input channels
        :param oup: output channels
        :param groups:
        """
        super(Coordinate_Attention, self).__init__()
        self.w = w
        self.h = h
        self.inp = inp
        self.oup = oup
        self.groups = groups

        self.pool_h = AvgPool2D(pool_size=(1, self.w), strides=1, padding='same')
        self.pool_w = AvgPool2D(pool_size=(self.h, 1), strides=1, padding='same')

        self.mip = max(8, self.inp // self.groups)

        self.conv1 = Conv2D(filters=self.mip, kernel_size=(1,1), strides=1, padding='same')
        self.bn1 = BatchNormalization()
        self.conv2 = Conv2D(filters=self.oup, kernel_size=(1,1), strides=1, padding='same')
        self.conv3 = Conv2D(filters=self.oup, kernel_size=(1,1), strides=1, padding='same')

        self.ac = Activation(ac_swish)

    def call(self, inputs):
        residual = inputs
        x = residual
        # n, c, h, w = x.shape

        x_h = self.pool_h(x)
        x_w = self.pool_w(x)

        x_w = transpose(x_w, [0, 2, 1, 3])

        y = concat([x_h, x_w], axis=1)
        y = self.conv1(y)
        y = self.bn1(y)
        y = self.ac(y)

        x_h, x_w = split(y, 2, axis=1)
        x_w = transpose(x_w, [0, 2, 1, 3])

        x_h = sigmoid(self.conv2(x_h))
        x_w = sigmoid(self.conv3(x_w))

        y = residual * x_w * x_h
        print(np.shape(y))
        return y


if __name__ == '__main__':
    model = Coordinate_Attention(w=14, h=14, inp=512, oup=512)
    # inputs = tf.keras.Input(shape=(14,14,512))
    # model.call(inputs=inputs)
    model.build((1, 14, 14, 512))
    model.summary()