引言

TensorFlow 是 Google 基于 DistBelief 进行研发的第二代人工智能学习系统,被广泛用于语音识别或图像识别等多项机器深度学习领域。其命名来源于本身的运行原理。Tensor(张量)意味着 N 维数组,Flow(流)意味着基于数据流图的计算,TensorFlow 代表着张量从图象的一端流动到另一端计算过程,是将复杂的数据结构传输至人工智能神经网中进行分析和处理的过程。

TensorFlow 完全开源,任何人都可以使用。可在小到一部智能手机、大到数千台数据中心服务器的各种设备上运行。

『机器学习进阶笔记』系列是将深入解析 TensorFlow 系统的技术实践,从零开始,由浅入深,与大家一起走上机器学习的进阶之路。

CUDA 与 TensorFlow 安装

按以往经验,TensorFlow 安装一条 pip 命令就可以解决,前提是有 fq 工具,没有的话去找找墙内别人分享的地址。而坑多在安装支持 gpu,需预先安装英伟达的 cuda,这里坑比较多,推荐使用 ubuntu deb 的安装方式来安装 cuda,run.sh 的方式总感觉有很多问题,cuda 的安装具体可以参考。 注意链接里面的 tensorflow 版本是以前的,tensorflow 现在官方上的要求是 cuda7.5+cudnnV4,请在安装的时候注意下。

Hello World

  1. import tensorflow as tf
  2. hello = tf.constant('Hello, TensorFlow!')
  3. sess = tf.Session()
  4. print sess.run(hello)

首先,通过 tf.constant 创建一个常量,然后启动 Tensorflow 的 Session,调用 sess 的 run 方法来启动整个 graph。
接下来我们做下简单的数学的方法:

  1. import tensorflow as tf
  2. a = tf.constant(2)
  3. b = tf.constant(3)
  4. with tf.Session() as sess:
  5. print "a=2, b=3"
  6. print "Addition with constants: %i" % sess.run(a+b)
  7. print "Multiplication with constants: %i" % sess.run(a*b)
  8. # output
  9. a=2, b=3
  10. Addition with constants: 5
  11. Multiplication with constants: 6

接下来用 tensorflow 的 placeholder 来定义变量做类似计算:
placeholder 的使用见https://www.tensorflow.org/versions/r0.8/api_docs/python/io_ops.html#placeholder

  1. import tensorflow as tf
  2. a = tf.placeholder(tf.int16)
  3. b = tf.placeholder(tf.int16)
  4. add = tf.add(a, b)
  5. mul = tf.mul(a, b)
  6. with tf.Session() as sess:
  7. # Run every operation with variable input
  8. print "Addition with variables: %i" % sess.run(add, feed_dict={a: 2, b: 3})
  9. print "Multiplication with variables: %i" % sess.run(mul, feed_dict={a: 2, b: 3})
  10. # output:
  11. Addition with variables: 5
  12. Multiplication with variables: 6
  13. matrix1 = tf.constant([[3., 3.]])
  14. matrix2 = tf.constant([[2.],[2.]])
  15. with tf.Session() as sess:
  16. result = sess.run(product)
  17. print result

线性回归

以下代码来自GitHub - aymericdamien/TensorFlow-Examples: TensorFlow Tutorial and Examples for beginners, 仅作学习用

  1. import tensorflow as tf
  2. import numpy
  3. import matplotlib.pyplot as plt
  4. rng = numpy.random
  5. # Parameters
  6. learning_rate = 0.01
  7. training_epochs = 2000
  8. display_step = 50
  9. # Training Data
  10. train_X = numpy.asarray([3.3,4.4,5.5,6.71,6.93,4.168,9.779,6.182,7.59,2.167,7.042,10.791,5.313,7.997,5.654,9.27,3.1])
  11. train_Y = numpy.asarray([1.7,2.76,2.09,3.19,1.694,1.573,3.366,2.596,2.53,1.221,2.827,3.465,1.65,2.904,2.42,2.94,1.3])
  12. n_samples = train_X.shape[0]
  13. # tf Graph Input
  14. X = tf.placeholder("float")
  15. Y = tf.placeholder("float")
  16. # Create Model
  17. # Set model weights
  18. W = tf.Variable(rng.randn(), name="weight")
  19. b = tf.Variable(rng.randn(), name="bias")
  20. # Construct a linear model
  21. activation = tf.add(tf.mul(X, W), b)
  22. # Minimize the squared errors
  23. cost = tf.reduce_sum(tf.pow(activation-Y, 2))/(2*n_samples) #L2 loss
  24. optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost) #Gradient descent
  25. # Initializing the variables
  26. init = tf.initialize_all_variables()
  27. # Launch the graph
  28. with tf.Session() as sess:
  29. sess.run(init)
  30. # Fit all training data
  31. for epoch in range(training_epochs):
  32. for (x, y) in zip(train_X, train_Y):
  33. sess.run(optimizer, feed_dict={X: x, Y: y})
  34. #Display logs per epoch step
  35. if epoch % display_step == 0:
  36. print "Epoch:", '%04d' % (epoch+1), "cost=", \
  37. "{:.9f}".format(sess.run(cost, feed_dict={X: train_X, Y:train_Y})), \
  38. "W=", sess.run(W), "b=", sess.run(b)
  39. print "Optimization Finished!"
  40. print "cost=", sess.run(cost, feed_dict={X: train_X, Y: train_Y}), \
  41. "W=", sess.run(W), "b=", sess.run(b)
  42. #Graphic display
  43. plt.plot(train_X, train_Y, 'ro', label='Original data')
  44. plt.plot(train_X, sess.run(W) * train_X + sess.run(b), label='Fitted line')
  45. plt.legend()
  46. plt.show()

逻辑回归

  1. import tensorflow as tf
  2. # Import MINST data
  3. from tensorflow.examples.tutorials.mnist import input_data
  4. mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)
  5. # Parameters
  6. learning_rate = 0.01
  7. training_epochs = 25
  8. batch_size = 100
  9. display_step = 1
  10. # tf Graph Input
  11. x = tf.placeholder(tf.float32, [None, 784]) # mnist data image of shape 28*28=784
  12. y = tf.placeholder(tf.float32, [None, 10]) # 0-9 digits recognition => 10 classes
  13. # Set model weights
  14. W = tf.Variable(tf.zeros([784, 10]))
  15. b = tf.Variable(tf.zeros([10]))
  16. # Construct model
  17. pred = tf.nn.softmax(tf.matmul(x, W) + b) # Softmax
  18. # Minimize error using cross entropy
  19. cost = tf.reduce_mean(-tf.reduce_sum(y*tf.log(pred), reduction_indices=1))
  20. # Gradient Descent
  21. optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
  22. # Initializing the variables
  23. init = tf.initialize_all_variables()
  24. # Launch the graph
  25. with tf.Session() as sess:
  26. sess.run(init)
  27. # Training cycle
  28. for epoch in range(training_epochs):
  29. avg_cost = 0.
  30. total_batch = int(mnist.train.num_examples/batch_size)
  31. # Loop over all batches
  32. for i in range(total_batch):
  33. batch_xs, batch_ys = mnist.train.next_batch(batch_size)
  34. # Run optimization op (backprop) and cost op (to get loss value)
  35. _, c = sess.run([optimizer, cost], feed_dict={x: batch_xs,
  36. y: batch_ys})
  37. # Compute average loss
  38. avg_cost += c / total_batch
  39. # Display logs per epoch step
  40. if (epoch+1) % display_step == 0:
  41. print "Epoch:", '%04d' % (epoch+1), "cost=", "{:.9f}".format(avg_cost)
  42. print "Optimization Finished!"
  43. # Test model
  44. correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
  45. # Calculate accuracy
  46. accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
  47. print "Accuracy:", accuracy.eval({x: mnist.test.images, y: mnist.test.labels})
  48. # result :
  49. Epoch: 0001 cost= 29.860467369
  50. Epoch: 0002 cost= 22.001451784
  51. Epoch: 0003 cost= 21.019925554
  52. Epoch: 0004 cost= 20.561320320
  53. Epoch: 0005 cost= 20.109135756
  54. Epoch: 0006 cost= 19.927862290
  55. Epoch: 0007 cost= 19.548687116
  56. Epoch: 0008 cost= 19.429119071
  57. Epoch: 0009 cost= 19.397068211
  58. Epoch: 0010 cost= 19.180813479
  59. Epoch: 0011 cost= 19.026808132
  60. Epoch: 0012 cost= 19.057875510
  61. Epoch: 0013 cost= 19.009575057
  62. Epoch: 0014 cost= 18.873240641
  63. Epoch: 0015 cost= 18.718575359
  64. Epoch: 0016 cost= 18.718761925
  65. Epoch: 0017 cost= 18.673640560
  66. Epoch: 0018 cost= 18.562128253
  67. Epoch: 0019 cost= 18.458205289
  68. Epoch: 0020 cost= 18.538211225
  69. Epoch: 0021 cost= 18.443384213
  70. Epoch: 0022 cost= 18.428727668
  71. Epoch: 0023 cost= 18.304270616
  72. Epoch: 0024 cost= 18.323529782
  73. Epoch: 0025 cost= 18.247192113
  74. Optimization Finished!
  75. (10000, 784)
  76. Accuracy 0.9206

这里有个小插曲,ipython notebook 在一个 notebook 打开时,一直在占用 GPU 资源,可能是之前有一个 notebook 一直打开着,然后占用着 GPU 资源,然后在计算 Accuracy 的”InternalError: Dst tensor is not initialized.” 然后找了 github 上面也有这个问题InternalError: Dst tensor is not initialized., 可以肯定是 GPU 的 memory 相关的问题,所以就尝试加上 tf.device(‘/cpu:0’),将 Accuracy 这步拉到 cpu 上计算,但是又出现 OOM 的问题,最后 nvidia-smi 时,发现有一个 python 脚本一直占用 3g 多的显存,把它 kill 之后恢复了,之前还比较吐槽怎么可能 10000*784 个 float 就把显存撑爆呢,原来是自己的问题。

这里逻辑回归,model 是一个 softmax 函数用来做多元分类,大概意思是选择 10 当中最后预测概率最高作为最终的分类。

其实基本的 tensorflow 没有特别好讲的,语法的课程什么可以去看看基本的文档,之后我会找一点经典有趣的 tensorflow 的代码应用来看看,毕竟『show me the code 』才是程序猿应有的态度。

本文由『UCloud 内核与虚拟化研发团队』提供。

关于作者:

Burness( ), UCloud 平台研发中心深度学习研发工程师,tflearn Contributor,做过电商推荐、精准化营销相关算法工作,专注于分布式深度学习框架、计算机视觉算法研究,平时喜欢玩玩算法,研究研究开源的项目,偶尔也会去一些数据比赛打打酱油,生活中是个极客,对新技术、新技能痴迷。

「UCloud 机构号」将独家分享云计算领域的技术洞见、行业资讯以及一切你想知道的相关讯息。

欢迎提问 & 求关注 o(////▽////)q~

以上。
https://zhuanlan.zhihu.com/p/22410917