变分自编码器python代码

import tensorflow as tf import numpy as np import os np.random.seed(0) tf.set_random_seed(0) #xavier参数初始化 def xavier_init(shape_in,shape_out, constant=1): low = -constant*np.sqrt(6.0/(shape_in + shape_out)) high = constant*np.sqrt(6.0/(shape_in + shape_out)) return tf.Variable(tf.random_uniform((shape_in, shape_out),minval = low, maxval=high,dtype=tf.float32)) def bias(shape): return tf.constant(0.1, shape=shape) class VariationalAutoencoder(object): def __init__(self, transfer_fct=tf.nn.relu, # softplus sigmoid, softplus , n_hidden_recog_1 = 500, n_hidden_recog_2 = 500, n_hidden_gener_1 = 500, n_hidden_gener_2 = 500, n_input = 784, n_z = 2, learning_rate=0.001, batch_size=100): self.transfer_fct = transfer_fct self.learning_rate = learning_rate self.batch_size = batch_size self.n_hidden_recog_1 = n_hidden_recog_1 self.n_hidden_recog_2 = n_hidden_recog_2 self.n_hidden_gener_1 = n_hidden_gener_1 self.n_hidden_gener_2 = n_hidden_gener_2 self.n_input = n_input self.n_z = n_z # 输入占位符 self.x = tf.placeholder(tf.float32, [self.batch_size, 784]) # # 创建自编码器网络 # self._create_network() # # 损失函数和优化器 # self._create_loss_optimizer() def _create_network(self): self.z_mean, self.z_log_sigma_sq = self._recognition_network() #重新参数化生成z eps = tf.random_normal((self.batch_size, self.n_z), 0, 1, dtype=tf.float32) with tf.variable_scope('z'): # z = mu + sigma*epsilon self.z = tf.add(self.z_mean, tf.multiply(tf.sqrt(tf.exp(self.z_log_sigma_sq)), eps)) self._summary_helper(self.z) self.x_reconstr_mean = self._generator_network() #编码网络encode network def _recognition_network(self): with tf.variable_scope('recognition_network_layer_1'): rec_weights_h1 = xavier_init(self.n_input, self.n_hidden_recog_1) rec_biases_h1 = bias([self.n_hidden_recog_1]) tf.summary.histogram('rec_weights_h1', rec_weights_h1) tf.summary.histogram('rec_biases_h1', rec_biases_h1) rec_layer_1 = self.transfer_fct(tf.matmul(self.x, rec_weights_h1) + rec_biases_h1,name = 'recognition_network_layer_1') self._summary_helper(rec_layer_1) with tf.variable_scope('recognition_network_layer_2'): rec_weights_h2 = xavier_init(self.n_hidden_recog_1, self.n_hidden_recog_2) rec_biases_h2 = bias([self.n_hidden_recog_2]) tf.summary.histogram('rec_weights_h2', rec_weights_h2) tf.summary.histogram('rec_biases_h2', rec_biases_h2) rec_layer_2 = self.transfer_fct(tf.matmul(rec_layer_1, rec_weights_h2)+rec_biases_h2,name = 'recognition_network_layer_2') self._summary_helper(rec_layer_2) with tf.variable_scope('recognition_network_mean'): weights_out_mean = xavier_init(self.n_hidden_recog_2, self.n_z) biases_out_mean = bias([self.n_z]) tf.summary.histogram('weights_out_mean', weights_out_mean) tf.summary.histogram('biases_out_mean', biases_out_mean) z_mean = tf.add(tf.matmul(rec_layer_2, weights_out_mean),biases_out_mean, name = 'recognition_network_mean') self._summary_helper(z_mean) with tf.variable_scope('recognition_network_sigma'): weights_out_log_sigma = xavier_init(self.n_hidden_recog_2, self.n_z) biases_out_log_sigma = bias([self.n_z]) tf.summary.histogram('weights_out_log_sigma', weights_out_log_sigma) tf.summary.histogram('biases_out_log_sigma', biases_out_log_sigma) z_log_sigma_sq = tf.add(tf.matmul(rec_layer_2, weights_out_log_sigma),biases_out_log_sigma, name = 'recognition_network_sigma') self._summary_helper(z_log_sigma_sq) return (z_mean, z_log_sigma_sq) #解码器/生成网络(decoder network) def _generator_network(self): with tf.variable_scope('generator_network_layer_1'): gen_weights_h1 = xavier_init(self.n_z, self.n_hidden_gener_1) gen_biases_h1 = bias([self.n_hidden_gener_1]) tf.summary.histogram('gen_weights_h1', gen_weights_h1) tf.summary.histogram('gen_biases_h1', gen_biases_h1) gen_layer_1 = self.transfer_fct(tf.matmul(self.z, gen_weights_h1) + gen_biases_h1, name = 'generator_network_layer_1') self._summary_helper(gen_layer_1) with tf.variable_scope('generator_network_layer_2'): gen_weights_h2 = xavier_init(self.n_hidden_gener_1, self.n_hidden_gener_2) gen_biases_h2 = bias([self.n_hidden_gener_2]) tf.summary.histogram('gen_weights_h2', gen_weights_h2) tf.summary.histogram('gen_biases_h2', gen_biases_h2) gen_layer_2 = self.transfer_fct(tf.matmul(gen_layer_1, gen_weights_h2) + gen_biases_h2, name = 'generator_network_layer_1') self._summary_helper(gen_layer_2) with tf.variable_scope('generator_network_x'): gen_weights_out = xavier_init(self.n_hidden_gener_2, self.n_input) gen_biases_out = bias([self.n_input]) tf.summary.histogram('gen_weights_out', gen_weights_out) tf.summary.histogram('gen_biases_out', gen_biases_out) x_reconstr_mean = tf.nn.sigmoid(tf.matmul(gen_layer_2, gen_weights_out) + gen_biases_out, name = 'generator_network_x') self._summary_helper(x_reconstr_mean) return x_reconstr_mean def _create_loss_optimizer(self): # 损失函数由两部分组成: with tf.variable_scope('loss'): reconstr_loss = -tf.reduce_sum(self.x * tf.log(1e-10 + self.x_reconstr_mean) + (1-self.x) * tf.log(1e-10 + 1 - self.x_reconstr_mean), 1) latent_loss = - 0.5 * tf.reduce_sum(1 + self.z_log_sigma_sq - tf.square(self.z_mean) - tf.exp(self.z_log_sigma_sq), 1) self.cost = tf.reduce_mean(reconstr_loss + latent_loss) # average over batch tf.summary.scalar('cost',self.cost) # Use ADAM optimizer self.optimizer = \ tf.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(self.cost) def _summary_helper(self, variable): tf.summary.histogram(variable.op.name + '/activation', variable) tf.summary.scalar(variable.op.name + '/mean', tf.reduce_mean(variable))