设为首页收藏本站
切换到窄版

鱼C论坛

 找回密码
 立即注册
鱼C论坛»论坛 › 技术交流 › Python交流 › Python实现PNN【tensorflow2.0】
返回列表 发新帖
查看: 880|回复: 1

[技术交流] Python实现PNN【tensorflow2.0】

[复制链接]
糖逗
发表于 2021-1-2 12:51:43 | 显示全部楼层 |阅读模式

马上注册，结交更多好友，享用更多功能^_^

您需要 登录 才可以下载或查看，没有账号？立即注册

x
本帖最后由 糖逗 于 2021-1-2 18:28 编辑

说明：待验证，有问题的地方欢迎批评指正

  1. #Inner Product-based Neural Network
  2. import tensorflow as tf
  3. from tensorflow.keras import layers, optimizers, Sequential
  4. from tensorflow import keras

  6. class PNN_Inner(keras.Model):
  7.     def __init__(self, D1):
  8.         super(PNN_Inner, self).__init__()
  9.         self.D1 = D1
  10.         
  11.     def build(self, input_shape):
  12.         self.N = input_shape[1]
  13.         self.M = input_shape[-1]
  14.         #[N, M, D1]
  15.         self.Wz = self.add_weight(shape = (self.N, self.M, self.D1),
  16.                                  trainable = True)
  17.         #[D1, N]
  18.         self.theta = self.add_weight(shape = (self.D1, self.N),
  19.                                  trainable = True)
  20.         #[batch, D1] -> [batch, 1]
  21.         self.l1 = layers.Dense(24, activation = tf.nn.leaky_relu)
  22.         self.bn1 = layers.BatchNormalization(axis = 1)
  23.         self.drop1 = layers.Dropout(0.5)
  24.         self.l2 = layers.Dense(12, activation = tf.nn.leaky_relu)
  25.         self.bn2 = layers.BatchNormalization(axis = 1)
  26.         self.drop2 = layers.Dropout(0.5)
  27.         self.l3 = layers.Dense(1, activation = tf.nn.sigmoid)
  28.         super(PNN_Inner, self).build(input_shape)
  29.    
  30.     def call(self, x, training = None):
  31.         #x:[batch, N, M]
  32.         #w:[N, M, D1]
  33.         linear = []
  34.         for i in range(self.D1):
  35.             #[N, M, 1] -> [N, M]
  36.             w = self.Wz[:, :, i]
  37.             #[batch, N. M] * [N, M] -> [batch, N, M]
  38.             temp = tf.multiply(x, w)
  39.             #[batch, N, M] -> [batch, 1]
  40.             temp = tf.expand_dims(tf.reduce_mean(temp, axis = [1, 2]), axis = 1)
  41.             linear.append(temp)
  42.         #linear:[batch, D1]
  43.         linear = tf.concat(linear, axis = 1)
  44.                            
  45.         product = []
  46.         #[batch, N, M] * [batch, N, M] -> [batch, N, M]
  47.         p = tf.multiply(x, x)
  48.         for i in range(self.D1):
  49.             #[N] -> [1, N]
  50.             theta = tf.expand_dims(self.theta[i,:], axis =0)
  51.             #[N, 1] @ [1, N] -> [N, N]
  52.             w = tf.matmul(tf.transpose(theta), theta)
  53.             #[batch, N, M] -> [batch, M, N]
  54.             f = tf.transpose(p, perm = [0, 2, 1])
  55.             #[batch, M, N] * [N, N] -> [batch, M, N]
  56.             temp = tf.matmul(f, w)
  57.             #[batch, M, N] -> [batch, 1]
  58.             temp = tf.expand_dims(tf.reduce_mean(temp, axis = [1, 2]), axis = 1)            
  59.             product.append(temp)
  60.         #product:[batch, D1]
  61.         product = tf.concat(product, axis = 1)
  62.         #[batch, D1] -> [batch, 2D1]
  63.         out = tf.concat([linear, product],axis = 1)
  64.         out = self.drop1(self.bn1(self.l1(out), training), training)
  65.         out = self.drop2(self.bn2(self.l2(out), training), training)
  66.         out = self.l3(out)
  67.         return out

  69.       
  70. model = PNN_Inner(D1 = 20)
  71. model.build((None,60, 30))
  72. model.summary()
复制代码



  1. #Outer Product-based Neural Network
  2. import tensorflow as tf
  3. from tensorflow.keras import layers, optimizers, Sequential
  4. from tensorflow import keras

  6. class PNN_Inner(keras.Model):
  7.     def __init__(self, D1):
  8.         super(PNN_Inner, self).__init__()
  9.         self.D1 = D1
  10.         
  11.     def build(self, input_shape):
  12.         self.N = input_shape[1]
  13.         self.M = input_shape[-1]
  14.         #[N, M, D1]
  15.         self.Wz = self.add_weight(shape = (self.N, self.M, self.D1),
  16.                                  trainable = True)
  17.         #[D1, M]
  18.         self.theta = self.add_weight(shape = (self.D1, self.M),
  19.                                  trainable = True)
  20.         #[batch, D1] -> [batch, 1]
  21.         self.l1 = layers.Dense(24, activation = tf.nn.leaky_relu)
  22.         self.bn1 = layers.BatchNormalization(axis = 1)
  23.         self.drop1 = layers.Dropout(0.5)
  24.         self.l2 = layers.Dense(12, activation = tf.nn.leaky_relu)
  25.         self.bn2 = layers.BatchNormalization(axis = 1)
  26.         self.drop2 = layers.Dropout(0.5)
  27.         self.l3 = layers.Dense(1, activation = tf.nn.sigmoid)
  28.         super(PNN_Inner, self).build(input_shape)
  29.    
  30.     def call(self, x, training = None):
  31.         #x:[batch, N, M]
  32.         #w:[N, M, D1]
  33.         linear = []
  34.         for i in range(self.D1):
  35.             #[N, M, 1] -> [N, M]
  36.             w = self.Wz[:, :, i]
  37.             #[batch, N. M] * [N, M] -> [batch, N, M]
  38.             temp = tf.multiply(x, w)
  39.             #[batch, N, M] -> [batch, 1]
  40.             temp = tf.expand_dims(tf.reduce_mean(temp, axis = [1, 2]), axis = 1)
  41.             linear.append(temp)
  42.         #linear:[batch, D1]
  43.         linear = tf.concat(linear, axis = 1)
  44.                            
  45.         product = []
  46.         #[batch, N, M] -> [batch, M]
  47.         fi = tf.reduce_mean(x, axis = 1)
  48.         #[batch, M] !* [batch, M] -> [batch, M, M]
  49.         p = tf.einsum('ai,aj->aij', fi, fi)
  50.         for i in range(self.D1):
  51.             #[M] -> [1, M]
  52.             theta = tf.expand_dims(self.theta[i,:], axis =0)
  53.             #[M, 1] @ [1, M] -> [M, M]
  54.             w = tf.matmul(tf.transpose(theta), theta)
  55.             #[batch, M, M] * [M, M] -> [batch, M, M]
  56.             temp = tf.matmul(p, w)
  57.             #[batch, M, M] -> [batch, 1]
  58.             temp = tf.expand_dims(tf.reduce_mean(temp, axis = [1, 2]), axis = 1)            
  59.             product.append(temp)
  60.         #product:[batch, D1]
  61.         product = tf.concat(product, axis = 1)
  62.         #[batch, D1] -> [batch, 2D1]
  63.         out = tf.concat([linear, product],axis = 1)
  64.         out = self.drop1(self.bn1(self.l1(out), training), training)
  65.         out = self.drop2(self.bn2(self.l2(out), training), training)
  66.         out = self.l3(out)
  67.         return out

  69.       
  70. model = PNN_Inner(D1 = 20)
  71. model.build((None,60, 30))
  72. model.summary()
复制代码

本帖被以下淘专辑推荐:

想知道小甲鱼最近在做啥？请访问 -> ilovefishc.com
回复

使用道具 举报

糖逗
 楼主| 发表于 2021-1-2 18:29:05 | 显示全部楼层
论文链接：https://arxiv.org/pdf/1611.00144v1.pdf
想知道小甲鱼最近在做啥？请访问 -> ilovefishc.com
回复 支持 反对

使用道具 举报

返回列表 发新帖
高级模式
B Color Image Link Quote Code Smilies
您需要登录后才可以回帖 登录 | 立即注册

本版积分规则

小黑屋|手机版|Archiver|鱼C工作室 ( 粤ICP备18085999号-1 | 粤公网安备 44051102000585号)

GMT+8, 2024-5-27 12:06

Powered by Discuz! X3.4

© 2001-2023 Discuz! Team.

快速回复 返回顶部 返回列表