用TensorFlow搭建一個全連接神經網絡

用TensorFlow搭建一個全連接神經網絡

---

說明

• 本例子利用TensorFlow搭建一個全連接神經網絡，實現對MNIST手寫數字的識別。

---

先上代碼

from tensorflow.examples.tutorials.mnist import input_data
import tensorflow as tf

# prepare data
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)

xs = tf.placeholder(tf.float32, [None, 784])
ys = tf.placeholder(tf.float32, [None, 10])

# the model of the fully-connected network
weights = tf.Variable(tf.random_normal([784, 10]))
biases = tf.Variable(tf.zeros([1, 10]) + 0.1)
outputs = tf.matmul(xs, weights) + biases
predictions = tf.nn.softmax(outputs)
cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys * tf.log(predictions),
                                              reduction_indices=[1]))
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)

# compute the accuracy
correct_predictions = tf.equal(tf.argmax(predictions, 1), tf.argmax(ys, 1))
accuracy = tf.reduce_mean(tf.cast(correct_predictions, tf.float32))

with tf.Session() as sess:
    init = tf.global_variables_initializer()
    sess.run(init)
    for i in range(1000):
        batch_xs, batch_ys = mnist.train.next_batch(100)
        sess.run(train_step, feed_dict={
            xs: batch_xs,
            ys: batch_ys
        })
        if i % 50 == 0:
            print(sess.run(accuracy, feed_dict={
                xs: mnist.test.images,
                ys: mnist.test.labels
            }))

---

代碼解析

1. 讀取MNIST數據

mnist = input_data.read_data_sets('MNIST_data', one_hot=True)

2. 建立佔位符

xs = tf.placeholder(tf.float32, [None, 784])
ys = tf.placeholder(tf.float32, [None, 10])

• xs 代表圖片像素數據, 每張圖片(28×28)被展開成(1×784), 有多少圖片還未定, 所以shape爲None×784.
• ys 代表圖片標籤數據, 0-9十個數字被表示成One-hot形式, 即只有對應bit爲1, 其餘爲0.

3. 建立模型

weights = tf.Variable(tf.random_normal([784, 10]))
biases = tf.Variable(tf.zeros([1, 10]) + 0.1)
outputs = tf.matmul(xs, weights) + biases
predictions = tf.nn.softmax(outputs)
cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys * tf.log(predictions),
                                              reduction_indices=[1]))
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)

使用Softmax函數作爲激活函數：

ouput=Softmax(input×weight+bias)

4. 計算正確率

correct_predictions = tf.equal(tf.argmax(predictions, 1), tf.argmax(ys, 1))
accuracy = tf.reduce_mean(tf.cast(correct_predictions, tf.float32))

5. 使用模型

with tf.Session() as sess:
    init = tf.global_variables_initializer()
    sess.run(init)
    for i in range(1000):
        batch_xs, batch_ys = mnist.train.next_batch(100)
        sess.run(train_step, feed_dict={
            xs: batch_xs,
            ys: batch_ys
        })
        if i % 50 == 0:
            print(sess.run(accuracy, feed_dict={
                xs: mnist.test.images,
                ys: mnist.test.labels
            }))

---

運行結果

訓練1000個循環, 準確率在87%左右.

Extracting MNIST_data/train-images-idx3-ubyte.gz
Extracting MNIST_data/train-labels-idx1-ubyte.gz
Extracting MNIST_data/t10k-images-idx3-ubyte.gz
Extracting MNIST_data/t10k-labels-idx1-ubyte.gz
0.1041
0.632
0.7357
0.7837
0.7971
0.8147
0.8283
0.8376
0.8423
0.8501
0.8501
0.8533
0.8567
0.8597
0.8552
0.8647
0.8654
0.8701
0.8712
0.8712

---

參考

• https://morvanzhou.github.io/tutorials/machine-learning/tensorflow/5-01-classifier/

---

• 我的個人主頁：http://www.techping.cn/
• 我的CSDN博客：http://blog.csdn.net/techping
• 我的簡書：http://www.jianshu.com/users/b2a36e431d5e/
• 我的GitHub：https://github.com/techping