ckpt = tf.train.get_checkpoint_state('./model/')
print(ckpt.model_checkpoint_path)

ckpt = tf.train.get_checkpoint_state('./model/')

saver = tf.train.Saver({"v/ExponentialMovingAverage":v}) 

saver.restore(sess,'./model/model.ckpt-0')
saver.restore(sess,ckpt.model_checkpoint_path)

'''
使用原網(wǎng)絡(luò)保存的模型加載到自己重新定義的圖上
可以使用python變量名加載模型，也可以使用節(jié)點(diǎn)名
'''
import AlexNet as Net
import AlexNet_train as train
import random
import tensorflow as tf
 
IMAGE_PATH = './flower_photos/daisy/5673728_71b8cb57eb.jpg'
 
with tf.Graph().as_default() as g:
 
 x = tf.placeholder(tf.float32, [1, train.INPUT_SIZE[0], train.INPUT_SIZE[1], 3])
 y = Net.inference_1(x, N_CLASS=5, train=False)
 
 with tf.Session() as sess:
  # 程序前面得有 Variable 供 save or restore 才不報(bào)錯(cuò)
  # 否則會(huì)提示沒(méi)有可保存的變量
  saver = tf.train.Saver()
 
  ckpt = tf.train.get_checkpoint_state('./model/')
  img_raw = tf.gfile.FastGFile(IMAGE_PATH, 'rb').read()
  img = sess.run(tf.expand_dims(tf.image.resize_images(
   tf.image.decode_jpeg(img_raw),[224,224],method=random.randint(0,3)),0))
 
  if ckpt and ckpt.model_checkpoint_path:
   print(ckpt.model_checkpoint_path)
   saver.restore(sess,'./model/model.ckpt-0')
   global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
   res = sess.run(y, feed_dict={x: img})
   print(global_step,sess.run(tf.argmax(res,1)))

'''
直接使用使用保存好的圖
無(wú)需加載python定義的結(jié)構(gòu)，直接使用節(jié)點(diǎn)名稱加載模型
由于節(jié)點(diǎn)形狀已經(jīng)定下來(lái)了，所以有不便之處，placeholder定義batch后單張傳會(huì)報(bào)錯(cuò)
現(xiàn)階段不推薦使用，以后如果理解深入了可能會(huì)找到使用方法
'''
import AlexNet_train as train
import random
import tensorflow as tf
 
IMAGE_PATH = './flower_photos/daisy/5673728_71b8cb57eb.jpg'
 
 
ckpt = tf.train.get_checkpoint_state('./model/')       # 通過(guò)檢查點(diǎn)文件鎖定最新的模型
saver = tf.train.import_meta_graph(ckpt.model_checkpoint_path +'.meta') # 載入圖結(jié)構(gòu)，保存在.meta文件中
 
with tf.Session() as sess:
 saver.restore(sess,ckpt.model_checkpoint_path)      # 載入?yún)?shù)，參數(shù)保存在兩個(gè)文件中，不過(guò)restore會(huì)自己尋找
 
 img_raw = tf.gfile.FastGFile(IMAGE_PATH, 'rb').read()
 img = sess.run(tf.image.resize_images(
  tf.image.decode_jpeg(img_raw), train.INPUT_SIZE, method=random.randint(0, 3)))
 imgs = []
 for i in range(128):
  imgs.append(img)
 print(sess.run(tf.get_default_graph().get_tensor_by_name('fc3:0'),feed_dict={'Placeholder:0': imgs}))
 
 '''
 img = sess.run(tf.expand_dims(tf.image.resize_images(
  tf.image.decode_jpeg(img_raw), train.INPUT_SIZE, method=random.randint(0, 3)), 0))
 print(img)
 imgs = []
 for i in range(128):
  imgs.append(img)
 print(sess.run(tf.get_default_graph().get_tensor_by_name('conv1:0'),
     feed_dict={'Placeholder:0':img}))

# 連同圖結(jié)構(gòu)一同加載
ckpt = tf.train.get_checkpoint_state('./model/')
saver = tf.train.import_meta_graph(ckpt.model_checkpoint_path +'.meta')
with tf.Session() as sess:
 saver.restore(sess,ckpt.model_checkpoint_path)
    
# 只加載數(shù)據(jù)，不加載圖結(jié)構(gòu)，可以在新圖中改變batch_size等的值
# 不過(guò)需要注意，Saver對(duì)象實(shí)例化之前需要定義好新的圖結(jié)構(gòu)，否則會(huì)報(bào)錯(cuò)
saver = tf.train.Saver()
with tf.Session() as sess:
 ckpt = tf.train.get_checkpoint_state('./model/')
 saver.restore(sess,ckpt.model_checkpoint_path)

# 新建空白圖
self.graph = tf.Graph()
# 空白圖列為默認(rèn)圖
with self.graph.as_default():
 # 二進(jìn)制讀取模型文件
 with tf.gfile.FastGFile(os.path.join(model_dir,model_name),'rb') as f:
  # 新建GraphDef文件，用于臨時(shí)載入模型中的圖
  graph_def = tf.GraphDef()
  # GraphDef加載模型中的圖
  graph_def.ParseFromString(f.read())
  # 在空白圖中加載GraphDef中的圖
  tf.import_graph_def(graph_def,name='')
  # 在圖中獲取張量需要使用graph.get_tensor_by_name加張量名
  # 這里的張量可以直接用于session的run方法求值了
  # 補(bǔ)充一個(gè)基礎(chǔ)知識(shí)，形如'conv1'是節(jié)點(diǎn)名稱，而'conv1:0'是張量名稱，表示節(jié)點(diǎn)的第一個(gè)輸出張量
  self.input_tensor = self.graph.get_tensor_by_name(self.input_tensor_name)
  self.layer_tensors = [self.graph.get_tensor_by_name(name + ':0') for name in self.layer_operation_names]