下面用一個線下回歸模型來記載儲存模型、載入模型做預測

參考文章:

訓練一個線下回歸模型並儲存看程式碼：

1. import tensorflow as tf
2. import numpy as np
3. import matplotlib.pyplot as plt
4. money=np.array([[109],[82],[99],[72],[87],[78],[86],[84],[94],[57]]).astype(np.float32)
5. click=np.array([[11],[8],[8],[6],[7],[7],[7],[8],[9],[5]]).astype(np.float32)
6. x_test
   =money[0:5].reshape(-1,1)
7. y_test=click[0:5]
8. x_train=money[5:].reshape(-1,1)
9. y_train=click[5:]
10. x=tf.placeholder(tf.float32,[None,1],name='x')#儲存要輸入的格式
11. w=tf.Variable(tf.zeros([1,1]))
12. b=tf.Variable(tf.zeros([1]))
13. y=tf.matmul(x,w)+b
14. tf.add_to_collection('pred_network', y)#用於載入模型獲取要預測的網路結構
15. y_=tf.placeholder(tf.float32
    ,[None,1])
16. cost=tf.reduce_sum(tf.pow((y-y_),2))
17. train_step=tf.train.GradientDescentOptimizer(0.000001).minimize(cost)
18. init=tf.global_variables_initializer()
19. sess=tf.Session()
20. sess.run(init)
21. cost_history=[]
22. saver = tf.train.Saver()
23. for i in range(100):
24. feed={x:x_train,y_:y_train}
25. sess.run(train_step,feed_dict=
    feed)
26. cost_history.append(sess.run(cost,feed_dict=feed))
27. # 輸出最終的W,b和cost值
28. print("109的預測值是:",sess.run(y, feed_dict={x:[[109]]}))
29. print("W_Value: %f"% sess.run(w),"b_Value: %f"% sess.run(b),"cost_Value: %f"% sess.run(cost, feed_dict=feed))
30. saver_path = saver.save(sess,"/Users/shuubiasahi/Desktop/tensorflow/modelsave/model.ckpt",global_step=100)
31. print("model saved in file: ", saver_path)

模型訓練結果：

1. 2017-10-1123:05:12.606557: W tensorflow/core/platform/cpu_feature_guard.cc:45]TheTensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
2. 109的預測值是:[[9.84855175]]

儲存模型檔案分析：

一、 saver.restore()時填的檔名，因為在saver.save的時候，每個checkpoint會儲存三個檔案，如 
model.ckpt-100.meta, model.ckpt-100.index, model.ckpt-100.data-00000-of-00001 
在import_meta_graph時填的就是meta檔名，我們知道權值都儲存在model.ckpt-100.data-00000-of-00001這個檔案中，但是如果在restore方法中填這個檔名，就會報錯，應該填的是字首，這個字首可以使用tf.train.latest_checkpoint(checkpoint_dir)這個方法獲取。
二、模型的y中有用到placeholder，在sess.run()的時候肯定要feed對應的資料，因此還要根據具體placeholder的名字，從graph中使用get_operation_by_name方法獲取。

載入訓練好的模型：

1. import tensorflow as tf
2. with tf.Session()as sess:
3. new_saver=tf.train.import_meta_graph('/Users/shuubiasahi/Desktop/tensorflow/modelsave/model.ckpt-100.meta')
4. new_saver.restore(sess,"/Users/shuubiasahi/Desktop/tensorflow/modelsave/model.ckpt-100")
5. graph = tf.get_default_graph()
6. x=graph.get_operation_by_name('x').outputs[0]
7. y=tf.get_collection("pred_network")[0]
8. print("109的預測值是:",sess.run(y, feed_dict={x:[[109]]}))

載入模型後的預測結果:

1. 2017-10-1123:07:33.176523: W tensorflow/core/platform/cpu_feature_guard.cc:45]TheTensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
2. 109的預測值是:[[9.84855175]]
3. Process finished with exit code 0