P和C

时间：2023-03-09 15:18:52
P和C 
import tensorflow as tf

import numpy as np

import math

import keras

from keras.layers import Conv2D,Reshape,Input

import numpy as np

import matplotlib.pyplot as plt

""" Channel attention module"""

if __name__ == '__main__':

    file = tf.read_file('img.jpg')

    x = tf.image.decode_jpeg(file)

    #print("Tensor:", x)

    sess = tf.Session()

    x1 = sess.run(x)

    print("x1:",x1)

    gamma = 0.05

    sess = tf.Session()

    x1 = sess.run(x)

    x1 = tf.expand_dims(x1, dim =0)

    print("x1.shape:", x1.shape)

    m_batchsize, height, width, C = x1.shape

    proj_query = Reshape((width * height, C))(x1)

    print("proj_query:", type(proj_query))

    print("proj_query:", proj_query.shape)

    proj_query = sess.run(proj_query)

    print(proj_query)

    proj_key = Reshape((width * height, C))(x1)

    proj_key = sess.run(proj_key).transpose(0, 2, 1)

    print(proj_key)

    print("proj_key:", type(proj_key))

    print("proj_key:", proj_key.shape)

    proj_query = proj_query.astype(np.float32)

    proj_key = proj_key.astype(np.float32)

    # N, C, C, bmm 批次矩阵乘法

    energy = tf.matmul(proj_key,proj_query)

    energy = sess.run(energy)

    print("energy:", energy)

    # 这里实现了softmax用最后一维的最大值减去了原始数据, 获得了一个不是太大的值

    # 沿着最后一维的C选择最大值, keepdim保证输出和输入形状一致, 除了指定的dim维度大小为1

    energy_new = tf.reduce_max(energy, -1, keep_dims=True)

    print("after_softmax_energy:",sess.run(energy_new))

    sess = tf.Session()

    e = energy_new

    print("b:", sess.run(energy_new))

    size = energy.shape[1]

    for i in range(size - 1):

        e = tf.concat([e, energy_new], axis=-1)

    energy_new = e

    print("energy_new2:", sess.run(energy_new))

    energy_new = energy_new - energy

    print("energy_new3:", sess.run(energy_new))

    attention = tf.nn.softmax(energy_new, axis=-1)

    print("attention:", sess.run(attention))

    proj_value = Reshape((width * height, C))(x1)

    proj_value = sess.run(proj_value)

    proj_value = proj_value.astype(np.float32)

    print("proj_value:", proj_value.shape)

    out = tf.matmul(proj_value, attention)

    out = sess.run(out)

    #plt.imshow(out)

    print("out1:", out)

    out = out.reshape(m_batchsize, width * height, C)

    #out1 = out.reshape(m_batchsize, C, height, width)

    print("out2:", out.shape)

    out = gamma * out + x

    #out = sess.run(out)

    #out = out.astype(np.int16)

    print("out3:", out)
import tensorflow as tf

import numpy as np

import math

import keras

from keras.layers import Conv2D,Reshape,Input

from keras.regularizers import l2

from keras.layers.advanced_activations import ELU, LeakyReLU

from keras import Model

import cv2

"""

Important：

1、A为CxHxW => Conv+BN+ReLU => B, C 都为CxHxW

2、Reshape B, C to CxN (N=HxW)

3、Transpose B to B’

4、Softmax(Matmul(B’, C)) => spatial attention map S为NxN(HWxHW)

5、如上式1, 其中sji测量了第i个位置在第j位置上的影响

6、也就是第i个位置和第j个位置之间的关联程度/相关性, 越大越相似.

7、A => Covn+BN+ReLU => D 为CxHxW => reshape to CxN

8、Matmul(D, S’) => CxHxW, 这里设置为DS

9、Element-wise sum(scale parameter alpha * DS, A) => the final output E 为 CxHxW (式2)

10、alpha is initialized as 0 and gradually learn to assign more weight.

"""

"""

        inputs :

            x : input feature maps( N X C X H X W)

        returns :

            out : attention value + input feature

            attention: N X (HxW) X (HxW)

"""

""" Position attention module"""

if __name__ == '__main__':

    #x = tf.random_uniform([2, 7, 7, 3],minval=0,maxval=255,dtype=tf.float32)

    file = tf.read_file('img.jpg')

    x = tf.image.decode_jpeg(file)

    #x = cv2.imread('ROIVIA3.jpg')

    print(x)

    gamma = 0.05

    sess = tf.Session()

    x1 =  sess.run(x)

    x1 = tf.expand_dims(x1, axis=0)

    print(x1.shape)

    in_dim = 3

    xlen = x1.shape[1]

    ylen = x1.shape[2]

    input = Input(shape=(xlen,ylen,3))

    query_conv = Conv2D(1, (1,1), activation='relu',kernel_initializer='he_normal')(input)

    key_conv = Conv2D(1, (1, 1), activation='relu', kernel_initializer='he_normal')(input)

    value_conv = Conv2D(3, (1, 1), activation='relu', kernel_initializer='he_normal')(input)

    print(query_conv)

    batchsize, height, width, C = x1.shape

    #print(C, height, width )

    # B => N, C, HW

    proj_query = Reshape(( width * height ,1))(query_conv)

    proj_key = Reshape(( width * height, 1))(key_conv)

    proj_value = Reshape((width * height, 3))(value_conv)

    print("proj_query:",proj_query)

    print("proj_key:", proj_key)

    print("proj_value:",proj_value.shape)

    model = Model(inputs=[input],outputs=[proj_query])

    model.compile(optimizer='adam',loss='binary_crossentropy')

    proj_query = model.predict(x1,steps=1)

    print("proj_query:",proj_query)

    # B' => N, HW, C

    proj_query = proj_query.transpose(0, 2, 1)

    print("proj_query2:", proj_query.shape)

    print("proj_query2:", type(proj_query))

    # C => N, C, HW

    model1 = Model(inputs=[input], outputs=[proj_key])

    model1.compile(optimizer='adam', loss='binary_crossentropy')

    proj_key = model1.predict(x1, steps=1)

    print("proj_key:", proj_key.shape)

    print(proj_key)

    # B'xC => N, HW, HW

    energy = tf.matmul(proj_key, proj_query)

    print("energy:",energy.shape)

    # S = softmax(B'xC) => N, HW, HW

    attention = tf.nn.softmax(energy, axis=-1)

    print("attention:", attention.shape)

    # D => N, C, HW

    model2 = Model(inputs=[input], outputs=[proj_value])

    model2.compile(optimizer='adam', loss='binary_crossentropy')

    proj_value = model2.predict(x1, steps=1)

    print("proj_value:",proj_value.shape)

    # DxS' => N, C, HW

    out = tf.matmul(proj_value, sess.run(attention).transpose(0, 2, 1))

    print("out:", out.shape)

    # N, C, H, W

    out = Reshape((height, width, 3))(out)

    print("out1:", out.shape)

    out = gamma * out + sess.run(x1)

    print("out2:", type(out))

随机推荐

  1. Android Studio中绘制simpleUML类图详细说明及使用

    一.Android Studio中安装simpleUML 1.下载simpleUML jar包 地址为:http://plugins.jetbrains.com/  搜索 simpleUMLCE 2. ...

  2. Linux下阅读源代码工具安装

    综合他们多篇博客,做一个自己的总结(从0开始,记录过程) 系统:ubuntu 16.04  vim:7.4.1689 内容来源: https://www.cnblogs.com/wangzhe1635 ...

  3. no module named win32api

    1 首先下载pywin32 https://sourceforge.net/projects/pywin32/files/pywin32/ 2进入虚拟环境 D:\env\jdscrapy\Lib\si ...

  4. Java文件File类学习总结

    java.io.File类 代表文件和目录,在开发中,读取文件.生成文件.删除文件.修改文件的属性都会用到该类. 常见构造方法: public File(String pathName){} 以pat ...

  5. Broadcom BCM94352z/DW1560驱动新姿势

    转自:https://blog.daliansky.net/Broadcom-BCM94352z-DW1560-drive-new-posture.html Broadcom WiFi/BlueToo ...

  6. Windows下虚拟机安装Mac OS X —– VMware Workstation12安装Mac OS X 10.11

    1下载  镜像:Instal OS X Yosemite 10.10.3(14D131).cdr        密码:qhhm 2 unlocker208文件(链接:https://pan.baidu ...

  7. Spark RDD 操作

    1. Spark RDD 创建操作 1.1 数据集合   parallelize 可以创建一个能够并行操作的RDD.其函数定义如下: ) scala> sc.defaultParallelism ...

  8. SSM框架指的是什么

    SSM(Spring+SpringMVC+MyBatis)框架集由Spring.SpringMVC.MyBatis三个开源框架整合而成, 常作为数据源较简单的web项目的框架. 其中spring是一个 ...

  9. JS 实现兼容浏览器报警提示声音

    <!DOCTYPE HTML> <head> <title>JS实现报警提示音</title> <meta http-equiv="co ...

  10. vue简单的导航栏

    <style> img{ display:block; margin:0 auto; width:500px; height:500px; } #app li{ list-style: n ...

相关文章