import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
 
INPUT_NODE = 784  # 輸入層節(jié)點=圖片像素=28x28=784
OUTPUT_NODE = 10  # 輸出層節(jié)點數(shù)=圖片類別數(shù)目
 
LAYER1_NODE = 500  # 隱藏層節(jié)點數(shù)，只有一個隱藏層
BATCH_SIZE = 100  # 一個訓練包中的數(shù)據(jù)個數(shù)，數(shù)字越小
          # 越接近隨機梯度下降，越大越接近梯度下降
 
LEARNING_RATE_BASE = 0.8   # 基礎學習率
LEARNING_RATE_DECAY = 0.99  # 學習率衰減率
 
REGULARIZATION_RATE = 0.0001  # 正則化項系數(shù)
TRAINING_STEPS = 30000     # 訓練輪數(shù)
MOVING_AVG_DECAY = 0.99    # 滑動平均衰減率
 
# 定義一個輔助函數(shù)，給定神經(jīng)網(wǎng)絡的輸入和所有參數(shù)，計算神經(jīng)網(wǎng)絡的前向傳播結果
def inference(input_tensor, avg_class, weights1, biases1,
       weights2, biases2):
 
 # 當沒有提供滑動平均類時，直接使用參數(shù)當前取值
 if avg_class == None:
  # 計算隱藏層前向傳播結果
  layer1 = tf.nn.relu(tf.matmul(input_tensor, weights1) + biases1)
  # 計算輸出層前向傳播結果
  return tf.matmul(layer1, weights2) + biases2
 else:
  # 首先計算變量的滑動平均值，然后計算前向傳播結果
  layer1 = tf.nn.relu(
    tf.matmul(input_tensor, avg_class.average(weights1)) +
    avg_class.average(biases1))
  
  return tf.matmul(
    layer1, avg_class.average(weights2)) + avg_class.average(biases2)
 
# 訓練模型的過程
def train(mnist):
 x = tf.placeholder(tf.float32, [None, INPUT_NODE], name='x-input')
 y_ = tf.placeholder(tf.float32, [None, OUTPUT_NODE], name='y-input')
 
 # 生成隱藏層參數(shù)
 weights1 = tf.Variable(
   tf.truncated_normal([INPUT_NODE, LAYER1_NODE], stddev=0.1))
 biases1 = tf.Variable(tf.constant(0.1, shape=[LAYER1_NODE]))
 
 # 生成輸出層參數(shù)
 weights2 = tf.Variable(
   tf.truncated_normal([LAYER1_NODE, OUTPUT_NODE], stddev=0.1))
 biases2 = tf.Variable(tf.constant(0.1, shape=[OUTPUT_NODE]))
 
 # 計算前向傳播結果，不使用參數(shù)滑動平均值 avg_class=None
 y = inference(x, None, weights1, biases1, weights2, biases2)
 
 # 定義訓練輪數(shù)變量，指定為不可訓練
 global_step = tf.Variable(0, trainable=False)
 
 # 給定滑動平均衰減率和訓練輪數(shù)的變量，初始化滑動平均類
 variable_avgs = tf.train.ExponentialMovingAverage(
   MOVING_AVG_DECAY, global_step)
 
 # 在所有代表神經(jīng)網(wǎng)絡參數(shù)的可訓練變量上使用滑動平均
 variables_avgs_op = variable_avgs.apply(tf.trainable_variables())
 
 # 計算使用滑動平均值后的前向傳播結果
 avg_y = inference(x, variable_avgs, weights1, biases1, weights2, biases2)
 
 # 計算交叉熵作為損失函數(shù)
 cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
   logits=y, labels=tf.argmax(y_, 1))
 cross_entropy_mean = tf.reduce_mean(cross_entropy)
 
 # 計算L2正則化損失函數(shù)
 regularizer = tf.contrib.layers.l2_regularizer(REGULARIZATION_RATE)
 regularization = regularizer(weights1) + regularizer(weights2)
 
 loss = cross_entropy_mean + regularization
 
 # 設置指數(shù)衰減的學習率
 learning_rate = tf.train.exponential_decay(
   LEARNING_RATE_BASE,
   global_step,              # 當前迭代輪數(shù)
   mnist.train.num_examples / BATCH_SIZE, # 過完所有訓練數(shù)據(jù)的迭代次數(shù)
   LEARNING_RATE_DECAY)
 
 
 # 優(yōu)化損失函數(shù)
 train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(
   loss, global_step=global_step)
 
 # 反向傳播同時更新神經(jīng)網(wǎng)絡參數(shù)及其滑動平均值
 with tf.control_dependencies([train_step, variables_avgs_op]):
  train_op = tf.no_op(name='train')
 
 # 檢驗使用了滑動平均模型的神經(jīng)網(wǎng)絡前向傳播結果是否正確
 correct_prediction = tf.equal(tf.argmax(avg_y, 1), tf.argmax(y_, 1))
 accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
 
 
 # 初始化會話并開始訓練
 with tf.Session() as sess:
  tf.global_variables_initializer().run()
  
  # 準備驗證數(shù)據(jù)，用于判斷停止條件和訓練效果
  validate_feed = {x: mnist.validation.images,
          y_: mnist.validation.labels}
  
  # 準備測試數(shù)據(jù)，用于模型優(yōu)劣的最后評價標準
  test_feed = {x: mnist.test.images, y_: mnist.test.labels}
  
  # 迭代訓練神經(jīng)網(wǎng)絡
  for i in range(TRAINING_STEPS):
   if i%1000 == 0:
    validate_acc = sess.run(accuracy, feed_dict=validate_feed)
    print("After %d training step(s), validation accuracy using average " 
       "model is %g " % (i, validate_acc))
    
   xs, ys = mnist.train.next_batch(BATCH_SIZE)
   sess.run(train_op, feed_dict={x: xs, y_: ys})
  
  # 訓練結束后在測試集上檢測模型的最終正確率
  test_acc = sess.run(accuracy, feed_dict=test_feed)
  print("After %d training steps, test accuracy using average model "
     "is %g " % (TRAINING_STEPS, test_acc))
  
  
# 主程序入口
def main(argv=None):
 mnist = input_data.read_data_sets("/tmp/data", one_hot=True)
 train(mnist)
 
# Tensorflow主程序入口
if __name__ == '__main__':
 tf.app.run()

Extracting /tmp/data/train-images-idx3-ubyte.gz
Extracting /tmp/data/train-labels-idx1-ubyte.gz
Extracting /tmp/data/t10k-images-idx3-ubyte.gz
Extracting /tmp/data/t10k-labels-idx1-ubyte.gz
After 0 training step(s), validation accuracy using average model is 0.0462 
After 1000 training step(s), validation accuracy using average model is 0.9784 
After 2000 training step(s), validation accuracy using average model is 0.9806 
After 3000 training step(s), validation accuracy using average model is 0.9798 
After 4000 training step(s), validation accuracy using average model is 0.9814 
After 5000 training step(s), validation accuracy using average model is 0.9826 
After 6000 training step(s), validation accuracy using average model is 0.9828 
After 7000 training step(s), validation accuracy using average model is 0.9832 
After 8000 training step(s), validation accuracy using average model is 0.9838 
After 9000 training step(s), validation accuracy using average model is 0.983 
After 10000 training step(s), validation accuracy using average model is 0.9836 
After 11000 training step(s), validation accuracy using average model is 0.9822 
After 12000 training step(s), validation accuracy using average model is 0.983 
After 13000 training step(s), validation accuracy using average model is 0.983 
After 14000 training step(s), validation accuracy using average model is 0.9844 
After 15000 training step(s), validation accuracy using average model is 0.9832 
After 16000 training step(s), validation accuracy using average model is 0.9844 
After 17000 training step(s), validation accuracy using average model is 0.9842 
After 18000 training step(s), validation accuracy using average model is 0.9842 
After 19000 training step(s), validation accuracy using average model is 0.9838 
After 20000 training step(s), validation accuracy using average model is 0.9834 
After 21000 training step(s), validation accuracy using average model is 0.9828 
After 22000 training step(s), validation accuracy using average model is 0.9834 
After 23000 training step(s), validation accuracy using average model is 0.9844 
After 24000 training step(s), validation accuracy using average model is 0.9838 
After 25000 training step(s), validation accuracy using average model is 0.9834 
After 26000 training step(s), validation accuracy using average model is 0.984 
After 27000 training step(s), validation accuracy using average model is 0.984 
After 28000 training step(s), validation accuracy using average model is 0.9836 
After 29000 training step(s), validation accuracy using average model is 0.9842 
After 30000 training steps, test accuracy using average model is 0.9839