import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data
from tensorflow.contrib.tensorboard.plugins import projector
import time

IMAGE_PIXELS = 28
hidden_unit = 100
output_nums = 10
learning_rate = 0.001
train_steps = 50000
batch_size = 500
test_data_size = 10000
#日志目錄(這里根據(jù)自己的目錄修改)
logdir = 'D:/Develop_Software/Anaconda3/WorkDirectory/summary/mnist'
#導(dǎo)入mnist數(shù)據(jù)
mnist = input_data.read_data_sets('MNIST_data', one_hot = True)

 #全局訓(xùn)練步數(shù)
global_step = tf.Variable(0, name = 'global_step', trainable = False)
with tf.name_scope('input'):
 #輸入數(shù)據(jù)
 with tf.name_scope('x'):
 x = tf.placeholder(
  dtype = tf.float32, shape = (None, IMAGE_PIXELS * IMAGE_PIXELS))
 #收集x圖像的會總數(shù)據(jù)
 with tf.name_scope('x_summary'):
 shaped_image_batch = tf.reshape(
  tensor = x,
  shape = (-1, IMAGE_PIXELS, IMAGE_PIXELS, 1),
  name = 'shaped_image_batch')
 tf.summary.image(name = 'image_summary',
      tensor = shaped_image_batch,
      max_outputs = 10)
 with tf.name_scope('y_'):
 y_ = tf.placeholder(dtype = tf.float32, shape = (None, 10))

with tf.name_scope('hidden_layer'):
 with tf.name_scope('hidden_arg'):
 #隱層模型參數(shù)
 with tf.name_scope('hid_w'):
  
  hid_w = tf.Variable(
   tf.truncated_normal(shape = (IMAGE_PIXELS * IMAGE_PIXELS, hidden_unit)),
   name = 'hidden_w')
  #添加獲取隱層權(quán)重統(tǒng)計值匯總數(shù)據(jù)的匯總操作
  tf.summary.histogram(name = 'weights', values = hid_w)
  with tf.name_scope('hid_b'):
  hid_b = tf.Variable(tf.zeros(shape = (1, hidden_unit), dtype = tf.float32),
       name = 'hidden_b')
 #隱層輸出
 with tf.name_scope('relu'):
 hid_out = tf.nn.relu(tf.matmul(x, hid_w) + hid_b)
with tf.name_scope('softmax_layer'):
 with tf.name_scope('softmax_arg'):
 #softmax層參數(shù)
 with tf.name_scope('sm_w'):
  
  sm_w = tf.Variable(
   tf.truncated_normal(shape = (hidden_unit, output_nums)),
   name = 'softmax_w')
  #添加獲取softmax層權(quán)重統(tǒng)計值匯總數(shù)據(jù)的匯總操作
  tf.summary.histogram(name = 'weights', values = sm_w)
  with tf.name_scope('sm_b'):
  sm_b = tf.Variable(tf.zeros(shape = (1, output_nums), dtype = tf.float32), 
       name = 'softmax_b')
 #softmax層的輸出
 with tf.name_scope('softmax'):
 y = tf.nn.softmax(tf.matmul(hid_out, sm_w) + sm_b)
 #梯度裁剪，因為概率取值為[0, 1]為避免出現(xiàn)無意義的log(0)，故將y值裁剪到[1e-10, 1]
 y_clip = tf.clip_by_value(y, 1.0e-10, 1 - 1.0e-5)
with tf.name_scope('cross_entropy'):
 #使用交叉熵代價函數(shù)
 cross_entropy = -tf.reduce_sum(y_ * tf.log(y_clip) + (1 - y_) * tf.log(1 - y_clip))
 #添加獲取交叉熵的匯總操作
 tf.summary.scalar(name = 'cross_entropy', tensor = cross_entropy)
 
with tf.name_scope('train'):
 #若不使用同步訓(xùn)練機制，使用Adam優(yōu)化器
 optimizer = tf.train.AdamOptimizer(learning_rate = learning_rate)
 #單步訓(xùn)練操作，
 train_op = optimizer.minimize(cross_entropy, global_step = global_step)
#加載測試數(shù)據(jù)
test_image = mnist.test.images
test_label = mnist.test.labels
test_feed = {x:test_image, y_:test_label}

with tf.name_scope('accuracy'):
 prediction = tf.equal(tf.argmax(input = y, axis = 1),
      tf.argmax(input = y_, axis = 1))
 accuracy = tf.reduce_mean(
  input_tensor = tf.cast(x = prediction, dtype = tf.float32))
#創(chuàng)建嵌入變量
embedding_var = tf.Variable(test_image, trainable = False, name = 'embedding')
saver = tf.train.Saver({'embedding':embedding_var})
#創(chuàng)建元數(shù)據(jù)文件，將MNIST圖像測試集對應(yīng)的標簽寫入文件
def CreateMedaDataFile():
 with open(logdir + '/metadata.tsv', 'w') as f:
 label = np.nonzero(test_label)[1]
 for i in range(test_data_size):
  f.write('%d\n' % label[i])
#創(chuàng)建投影配置參數(shù)
def CreateProjectorConfig():
 config = projector.ProjectorConfig()
 embeddings = config.embeddings.add()
 embeddings.tensor_name = 'embedding:0'
 embeddings.metadata_path = logdir + '/metadata.tsv'
 
 projector.visualize_embeddings(writer, config)
 #聚集匯總操作
merged = tf.summary.merge_all()
#創(chuàng)建會話的配置參數(shù)
sess_config = tf.ConfigProto(
 allow_soft_placement = True,
 log_device_placement = False)
#創(chuàng)建會話
with tf.Session(config = sess_config) as sess:
 #創(chuàng)建FileWriter實例
 writer = tf.summary.FileWriter(logdir = logdir, graph = sess.graph)
 #初始化全局變量
 sess.run(tf.global_variables_initializer())
 time_begin = time.time()
 print('Training begin time: %f' % time_begin)
 while True:
 #加載訓(xùn)練批數(shù)據(jù)
 batch_x, batch_y = mnist.train.next_batch(batch_size)
 train_feed = {x:batch_x, y_:batch_y}
 loss, _, summary= sess.run([cross_entropy, train_op, merged], feed_dict = train_feed)
 step = global_step.eval()
 #如果step為100的整數(shù)倍
 if step % 100 == 0:
  now = time.time()
  print('%f: global_step = %d, loss = %f' % (
   now, step, loss))
  #向事件文件中添加匯總數(shù)據(jù)
  writer.add_summary(summary = summary, global_step = step)
 #若大于等于訓(xùn)練總步數(shù)，退出訓(xùn)練
 if step >= train_steps:
  break
 time_end = time.time()
 print('Training end time: %f' % time_end)
 print('Training time: %f' % (time_end - time_begin))
 #測試模型精度
 test_accuracy = sess.run(accuracy, feed_dict = test_feed)
 print('accuracy: %f' % test_accuracy)
 
 saver.save(sess = sess, save_path = logdir + '/embedding_var.ckpt')
 CreateMedaDataFile()
 CreateProjectorConfig()
 #關(guān)閉FileWriter
 writer.close()