mnist 초급

여러가지 케이스를 보고(학습) 그에 따라 어떤 형태인가를 확률로 계산한다.

계산할 때 softmax regression을 사용한다. 결과는 약 92%

import tensorflow as tf
 
# test data 다운로드
from tensorflow.examples.tutorials.mnist import input_data
 
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
 
# tf Graph Input
x = tf.placeholder(tf.float32, [None, 784])
 
# weight, bias
W = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
 
y = tf.nn.softmax(tf.matmul(x,W) + b)
 
y_ = tf.placeholder(tf.float32, [None, 10])
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_*tf.log(y), axis = 1))
 
# data를 변환하는 부분 : https://stackoverflow.com/questions/40088132/tensorflow-mnist-weight-and-bias-variables
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)
 
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
 
batch_size = 100
 
for i in range(1000):
    # batch_size 단위로 학습
    batch_xs, batch_ys = mnist.train.next_batch(batch_size)
    sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})
 
correct_prediction = tf.equal(tf.argmax(y,1), tf.argmax(y_,1))
 
accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
 
print(sess.run(accuracy, feed_dict={x: mnist.test.images, y_:mnist.test.labels}))