TensorFlow自编码器（AutoEncoder）之MNIST实践

可以使用自编码器进行降维，也可以通过加入噪声学习来实现降噪。

以下使用单个隐藏层来训练 mnist 数据集并共享对称权重参数。

模型本身并不难，调试过程中有几点需要注意：

1、建立模型

import numpy as np
import tensorflow as tf
class AutoEncoder(object):
    '''
    使用对称结构，解码器重用编码器的权重参数
    '''
    def __init__(self, input_shape, h1_size, lr):
        tf.reset_default_graph()# 重置默认计算图，有时出错后内存还一团糟
        with tf.variable_scope('auto_encoder', reuse=tf.AUTO_REUSE):
            self.W1 = self.weights(shape=(input_shape, h1_size), name='h1')
            self.b1 = self.bias(h1_size)
            self.W2 = tf.transpose(tf.get_variable('h1'))  # 共享参数，使用其转置
            self.b2 = self.bias(input_shape)

        self.lr = lr
        self.input = tf.placeholder(shape=(None, input_shape),
                                    dtype=tf.float32)
        self.h1_out = tf.nn.softplus(tf.matmul(self.input, self.W1) + self.b1)# softplus，类relu
        self.out = tf.matmul(self.h1_out, self.W2) + self.b2
        self.optimizer = tf.train.AdamOptimizer(learning_rate=self.lr)
        self.loss = 0.1 * tf.reduce_sum(
            tf.pow(tf.subtract(self.input, self.out), 2))
        self.train_op = self.optimizer.minimize(self.loss)
        self.sess = tf.Session()
        self.sess.run(tf.global_variables_initializer())
    def fit(self, X, epoches=100, batch_size=128, epoches_to_display=10):
        batchs_per_epoch = X.shape[0] // batch_size
        for i in range(epoches):
            epoch_loss = []
            for j in range(batchs_per_epoch):
                X_train = X[j * batch_size:(j + 1) * batch_size]

                loss, _ = self.sess.run([self.loss, self.train_op],
                                        feed_dict={self.input: X_train})
                epoch_loss.append(loss)
            if i % epoches_to_display == 0:
                print('avg_loss at epoch %d :%f' % (i, np.mean(epoch_loss)))
#         return self.sess.run(W1)
#     权重初始化参考别人的，这个居然很重要！用自己设定的截断正态分布随机没有效果
    def weights(self, shape, name, constant=1):
        fan_in = shape[0]
        fan_out = shape[1]
        low = -constant * np.sqrt(6.0 / (fan_in + fan_out))
        high = constant * np.sqrt(6.0 / (fan_in + fan_out))
        init = tf.random_uniform_initializer(minval=low, maxval=high)
        return tf.get_variable(name=name,
                               shape=shape,
                               initializer=init,

                               dtype=tf.float32)
    def bias(self, size):
        return tf.Variable(tf.constant(0, dtype=tf.float32, shape=[size]))
    def encode(self, X):
        return self.sess.run(self.h1_out, feed_dict={self.input: X})
    def decode(self, h):
        return self.sess.run(self.out, feed_dict={self.h1_out: h})
    def reconstruct(self, X):
        return self.sess.run(self.out, feed_dict={self.input: X})

2、加载数据和预处理

from keras.datasets import mnist
(X_train, y_train), (X_test, y_test) = mnist.load_data()
import random

X_train = X_train.reshape(-1, 784)
# 测试集里随机10个图片用做测试
test_idxs = random.sample(range(X_test.shape[0]), 10)
data_test = X_test[test_idxs].reshape(-1, 784)
# 标准化
import sklearn.preprocessing as prep
processer = prep.StandardScaler().fit(X_train) # 这里还是用全部数据好，这个也很关键！
X_train = processer.transform(X_train)
X_test = processer.transform(data_test)
# 随机5000张图片用做训练
idxs = random.sample(range(X_train.shape[0]), 5000)
data_train = X_train[idxs]

3、训练

model = AutoEncoder(784, 200, 0.001)  # 学习率对loss影响也有点大
model.fit(data_train, batch_size=128, epoches=200)  # 200轮即可

4、测试、视觉对比图

decoded_test = model.reconstruct(X_test)

import matplotlib.pyplot as plt
%matplotlib inline
shape = (28, 28)
fig, axes = plt.subplots(2,10,
                         figsize=(10, 2),
                         subplot_kw={
                             'xticks': [],
                             'yticks': []
                         },
                         gridspec_kw=dict(hspace=0.1, wspace=0.1))
for i in range(10):
    axes[0][i].imshow(np.reshape(X_test[i], shape))
    axes[1][i].imshow(np.reshape(decoded_test[i], shape))
plt.show()

结果如下：

以上，点高斯噪声可以添加到输入以增加鲁棒性。