[up-board开发板]+tensorflow mnist深度神经网络

IC爬虫

我现在正在使用UP Board发帖！！！！！！！！！！！！！！！！

tensorflow的优势是非常简单就能构建对大型数据的的训练，但是这只是它众多亮点中的一个。在训练大规模数据的时候库可以使用多种模型，我分别测试了使用softmax回归模型和深度神经网络模型对数据训练，再对比这两种模型训练出来的精度。

使用softmax回归模型：

1. import tensorflow as tf
   
2. import numpy as np
   
3. import input_data
   
4. 
   
5. mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
   
6. sess = tf.InteractiveSession()
   
7. x = tf.placeholder("float", shape=[None, 784])
   
8. y_ = tf.placeholder("float", shape=[None, 10])
   
9. W = tf.Variable(tf.zeros([784,10]))
   
10. b = tf.Variable(tf.zeros([10]))
    
11. sess.run(tf.initialize_all_variables())
    
12. y = tf.nn.softmax(tf.matmul(x,W) + b)
    
13. cross_entropy = -tf.reduce_sum(y_*tf.log(y))
    
14. train_step = tf.train.GradientDescentOptimizer(0.01).minimize(cross_entropy)
    
15. for i in range(1000):
    
16.         batch = mnist.train.next_batch(50)
    
17.         train_step.run(feed_dict={x: batch[0], y_: batch[1]})
    
18. correct_prediction = tf.equal(tf.argmax(y,1), tf.argmax(y_,1))
    
19. accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
    
20. print accuracy.eval(feed_dict={x: mnist.test.images, y_: mnist.test.labels})
    

复制代码

计算出在测试数据上的准确率，大概是91%


深度卷积网络：

1. def weight_variable(shape):
   
2.         initial = tf.truncated_normal(shape, stddev=0.1)
   
3.         return tf.Variable(initial)
   
4. 
   
5. def bias_variable(shape):
   
6.         initial = tf.constant(0.1, shape=shape)
   
7.         return tf.Variable(initial)
   
8. def conv2d(x, W):
   
9.         return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
   
10. 
    
11. def max_pool_2x2(x):
    
12.         return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],strides=[1, 2, 2, 1], padding='SAME')
    
13. W_conv1 = weight_variable([5, 5, 1, 32])
    
14. b_conv1 = bias_variable([32])        
    
15. x_image = tf.reshape(x, [-1,28,28,1])
    
16. h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
    
17. h_pool1 = max_pool_2x2(h_conv1)
    
18. 
    
19. W_conv2 = weight_variable([5, 5, 32, 64])
    
20. b_conv2 = bias_variable([64])
    
21. 
    
22. h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
    
23. h_pool2 = max_pool_2x2(h_conv2)
    
24. 
    
25. W_fc1 = weight_variable([7 * 7 * 64, 1024])
    
26. b_fc1 = bias_variable([1024])
    
27. 
    
28. h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
    
29. h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)        
    
30. keep_prob = tf.placeholder("float")
    
31. h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
    
32. W_fc2 = weight_variable([1024, 10])
    
33. b_fc2 = bias_variable([10])
    
34. 
    
35. y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
    
36. cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv))
    
37. train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
    
38. correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
    
39. accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
    
40. sess.run(tf.initialize_all_variables())
    
41. for i in range(20000):
    
42.         batch = mnist.train.next_batch(50)
    
43.         if i%100 == 0:
    
44.                 train_accuracy = accuracy.eval(feed_dict={x:batch[0], y_: batch[1], keep_prob: 1.0})
    
45.             print "step %d, training accuracy %g"%(i, train_accuracy)
    
46.         train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})
    
47. 
    
48. print "test accuracy %g"%accuracy.eval(feed_dict={x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0})

复制代码

在最终测试集上的准确率大概是99.2%。

michael_llh

看不懂，想学习一些相关的内容，有什么好的推荐学习方式吗。
有在看Udacity的Tensorflow的一个视频

netlhx

吃瓜群众路过

samshishi

支持，优势是非常明显。

testqly

大佬，这板子能用来跑神经网络啊？CPU跑起来不会很慢吗？