Module Setting



In [1]:

    
import tensorflow as tf
import pandas as pd
import numpy as np
import time

Read Data



In [2]:

    
dataset = pd.read_csv("dataset.txt")



In [3]:

    
dataset.head()









    Out[3]:







  
    
      
      Index
      Brain_Weight
      Body_Weight
    
  
  
    
      0
      1
      3.385
      44.5
    
    
      1
      2
      0.480
      15.5
    
    
      2
      3
      1.350
      8.1
    
    
      3
      4
      465.000
      423.0
    
    
      4
      5
      36.330
      119.5

위의 데이터셋을 보았을 때 어떤 전처리를 해야할까



In [4]:

    
def minmax(data):
    max_value = max(data)
    min_value = min(data)
    norm_value = (data-min_value)/ (max_value - min_value)
    return norm_value



In [5]:

    
x_data = dataset.Brain_Weight
y_data = dataset.Body_Weight
x_data = np.expand_dims(x_data, -1)
y_data = np.expand_dims(y_data, -1)



In [6]:

    
# data normalization
x_data = minmax(x_data)
y_data = minmax(y_data)

Parameter Setting

파라메터 세팅의 단계에선 Placeholder shape , learning rate, batch size, iteration 의 요소를 setting 한다.



In [7]:

    
batch_size = 20
print_every = 10
iteration = 100

x = tf.placeholder(tf.float32, shape=[None,1])
y = tf.placeholder(tf.float32, shape=[None,1])

Build model

모델을 구현한다.



In [8]:

    
w = tf.get_variable("weight", shape=[1,1], initializer=tf.random_normal_initializer(stddev=0.01))
b = tf.get_variable("bias", shape=[1,], initializer=tf.zeros_initializer())

hypothesis = tf.nn.bias_add(tf.matmul(x,w), b)

Train operator



In [9]:

    
cost = tf.reduce_mean(tf.square(hypothesis-y))
optimizer = tf.train.AdamOptimizer(0.01).minimize(cost)

Train

Tip : 항상 Learning time 을 측정하고 모델을 저장하는 습관을 들이자.



In [10]:

    
saver = tf.train.Saver()

start_time = time.time()
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    n_batch = int(len(x_data)/batch_size)
    for iteration in range(iteration):
        avg_loss = 0
        for step in range(n_batch):
            op = sess.run(optimizer, feed_dict = {x:x_data[step*n_batch:(step+1)*n_batch], y:y_data[step*n_batch:(step+1)*n_batch]})
            loss = sess.run(cost, feed_dict={x:x_data[step*n_batch:(step+1)*n_batch], y:y_data[step*n_batch:(step+1)*n_batch]})
            avg_loss += loss/n_batch
        if iteration % print_every == 0:
            print("cost at %d: %f"%(iteration+1,avg_loss))
    
    finish_time = time.time() - start_time
    save_path = saver.save(sess, "./model/model")     
    print("Label : ", x_data[:1])        
    print("predict : " , sess.run(hypothesis, feed_dict={x:x_data[:1]}))
    print("학습 소요 시간 : ", finish_time)









    



cost at 1: 0.000421
cost at 11: 0.000310
cost at 21: 0.000209
cost at 31: 0.000140
cost at 41: 0.000095
cost at 51: 0.000066
cost at 61: 0.000050
cost at 71: 0.000041
cost at 81: 0.000037
cost at 91: 0.000034
Label :  [[ 0.00050797]]
predict :  [[ 0.00939523]]
학습 소요 시간 :  0.6123685836791992

Restore and predict



In [11]:

    
sess = tf.Session()
saver = tf.train.Saver()
saver.restore(sess, 'model/model')
sess.run(hypothesis, feed_dict={x:x_data[:1]})









    



INFO:tensorflow:Restoring parameters from model/model






    Out[11]:





array([[ 0.00939523]], dtype=float32)