直接附上代码：
1 import numpy as np 2 import sklearn.preprocessing as prep 3 import tensorflow as tf 4 from tensorflow.examples.tutorials.mnist import input_data 5 6 def xavier_init(fan_in,fan_out,constant=1): 7 low=-constant*np.sqrt(6.0/(fan_in+fan_out)) 8 high=constant*np.sqrt(6.0/(fan_in+fan_out)) 9 return tf.random_uniform((fan_in,fan_out),minval=low,maxval=high,dtype=tf.float32) 10 11 class AdditiveGaussianNoiseAutoencoder(object): 12 def __init__(self,n_input,n_hidden,transfer_function=tf.nn.softplus,optimizer=tf.train.AdamOptimizer(),scale=0.1): 13 self.n_input=n_input 14 self.n_hidden=n_hidden 15 self.transfer=transfer_function 16 self.scale=tf.placeholder(tf.float32) 17 self.training_scale=scale 18 network_weights=self._initialize_weights() 19 self.weights=network_weights 20 21 self.x=tf.placeholder(tf.float32,[None,self.n_input]) 22 self.hidden=self.transfer(tf.add(tf.matmul(self.x+scale*tf.random_normal((n_input,)),self.weights['w1']),self.weights['b1'])) 23 self.reconstruction=tf.add(tf.matmul(self.hidden,self.weights['w2']),self.weights['b2']) 24 self.cost=0.5*tf.reduce_sum(tf.pow(tf.sub(self.reconstruction,self.x),2.0)) 25 self.optimizer=optimizer.minimize(self.cost) 26 27 init=tf.initialize_all_variables() 28 self.sess=tf.Session() 29 self.sess.run(init) 30 31 def _initialize_weights(self): 32 all_weights=dict() 33 all_weights['w1']=tf.Variable(xavier_init(self.n_input,self.n_hidden)) 34 all_weights['b1']=tf.Variable(tf.zeros([self.n_hidden],dtype=tf.float32)) 35 all_weights['w2']=tf.Variable(tf.zeros([self.n_hidden,self.n_input],dtype=tf.float32)) 36 all_weights['b2']=tf.Variable(tf.zeros([self.n_input],dtype=tf.float32)) 37 38 return all_weights 39 40 def partial_fit(self,X): 41 42 cost,opt=self.sess.run((self.cost,self.optimizer),feed_dict={self.x:X,self.scale:self.training_scale}) 43 44 return cost 45 46 def calc_total_cost(self,X): 47 return self.sess.run(self.cost,feed_dict={self.x:X,self.scale:self.training_scale}) 48 49 def transform(self,X): 50 return self.sess.run(self.hidden,feed_dict={self.x:X,self.scale:self.training_scale}) 51 52 def generate(self,hidden=None): 53 if hidden is None: 54 hidden=np.random.normal(size=self.weights["b1"]) 55 return self.sess.run(self.reconstruction,feed_dict={self.hidden:hidden}) 56 57 def reconstruct(self,X): 58 return self.sess.run(self.reconstruction,feed_dict={self.x:X,self.scale:self.training_scale}) 59 60 def getWeights(self): 61 return self.sess.run(self.weights['w1']) 62 63 def getBiases(self): 64 return self.sess.run(self.weights['b1']) 65 66 mnist=input_data.read_data_sets('MNIST_data',one_hot=True) 67 68 def standard_scale(X_train,X_test): 69 preprocessor=prep.StandardScaler().fit(X_train) 70 X_train=preprocessor.transform(X_train) 71 X_test=preprocessor.transform(X_test) 72 return X_train,X_test 73 74 def get_random_block_from_data(data,batch_size): 75 start_index=np.random.randint(0,len(data)-batch_size) 76 return data[start_index:(start_index+batch_size)] 77 78 X_train,X_test=standard_scale(mnist.train.images,mnist.test.images) 79 n_samples=int(mnist.train.num_examples) 80 training_epochs=20 81 batch_size=128 82 diaplay_step=1 83 autoencoder=AdditiveGaussianNoiseAutoencoder(n_input=784,n_hidden=200,transfer_function=tf.nn.softplus,optimizer=tf.train.AdamOptimizer(learning_rate=0.001),scale=0.01) 84 for epoch in range(training_epochs): 85 avg_cost=0 86 total_batch=int(n_samples/batch_size) 87 for i in range(total_batch): 88 batch_xs=get_random_block_from_data(X_train,batch_size) 89 90 cost=autoencoder.partial_fit(batch_xs) 91 avg_cost+=cost/n_samples*batch_size 92 93 if epoch%diaplay_step==0: 94 print("Epoch:",'%04d'%(epoch+1),"cost=","{:.9f}".format(avg_cost)) 95 96 print("Total cost: "+str(autoencoder.calc_total_cost(X_test)))