Spark.ML之PipeLine学习笔记

地址:

 
Spark PipeLine
是基于DataFrames的高层的API,可以方便用户构建和调试机器学习流水线
可以使得多个机器学习算法顺序执行,达到高效的数据处理的目的
 
DataFrame是来自Spark SQL的ML DataSet 可以存储一系列的数据类型,text,特征向量,Label和预测结果
 
Transformer:将DataFrame转化为另外一个DataFrame的算法,通过实现transform()方法
Estimator:将DataFrame转化为一个Transformer的算法,通过实现fit()方法
 
PipeLine:将多个Transformer和Estimator串成一个特定的ML Wolkflow
Parameter:Tansformer和Estimator共用同一个声明参数的API
 
Transformer和Estimator是PipeLine的Stage
Pipeline是一系列的Stage按照声明的顺序排列成的工作流
 
Transformer.transform()和Estimator.fit()都是无状态的
每一个Transformer和Estimator的实例都有唯一的ID在声明参数的时候非常有用
 
下面是一个线性的PipeLine的流程
 
上面创建的是线性的PipeLine,每一步都依赖上一步的结果
如果数据流可以组成有向不循环图(Directed Acyclic Graph DAG)
那么可以创建Non-Linear Pipeline
 
RuntimeCheching:因为PipeLine可以操作多种类型的DataFrame
所以不能使用编译时检测
那么PipeLine或者PipeLine Model使用运行时检测
这种检测使用了DataFrame Schema这个Schema是DataFrame列的数据类型的描述
 
Unique PipeLine Stage:PipeLine Stage应当都是唯一的实例,都拥有唯一的ID
 
Param是一个命名参数,带有自包含文档
ParamMap是一个参数与值的对(parameter,value)
 
将参数传递给算法主要有下面两种方式:
1. 为实例设置参数,若Ir是LogisticRegression的实例,调用Ir.SetMaxIter(10)意味着Ir.fit()做多调用10次
2. 传递一个ParamMap给fit()或者transform()那么位于map中的所有的parameter都会通过setter方法override以前的参数
 
很多时候将PipeLine保存到disk方便以后的使用是值得的
Spark 1.6时候,model Import/Export函数被添加到PipeLine API
大部分transformer和一些ML Model支持I/O
 
下面是基本组件的一些操作的例子:
 
 
  1. #导入向量和模型
  2. from pyspark.ml.linalg importVectors
  3. from pyspark.ml.classification importLogisticRegression
  4. #准备训练数据
  5. # Prepare training data from a list of (label, features) tuples.
  6. training = spark.createDataFrame([
  7. (1.0,Vectors.dense([0.0,1.1,0.1])),
  8. (0.0,Vectors.dense([2.0,1.0,-1.0])),
  9. (0.0,Vectors.dense([2.0,1.3,1.0])),
  10. (1.0,Vectors.dense([0.0,1.2,-0.5]))],["label","features"])
  11. #创建回归实例,这个实例是Estimator
  12. # Create a LogisticRegression instance. This instance is an Estimator.
  13. lr =LogisticRegression(maxIter=10, regParam=0.01)
  14. #打印出参数和文档
  15. # Print out the parameters, documentation, and any default values.
  16. print"LogisticRegression parameters:\n"+ lr.explainParams()+"\n"
  17. #使用Ir中的参数训练出Model1
  18. # Learn a LogisticRegression model. This uses the parameters stored in lr.
  19. model1 = lr.fit(training)
  20. # Since model1 is a Model (i.e., a transformer produced by an Estimator),
  21. # we can view the parameters it used during fit().
  22. # This prints the parameter (name: value) pairs, where names are unique IDs for this
  23. # LogisticRegression instance.
  24. #查看model1在fit()中使用的参数
  25. print"Model 1 was fit using parameters: "
  26. print model1.extractParamMap()
  27. #修改其中的一个参数
  28. # We may alternatively specify parameters using a Python dictionary as a paramMap
  29. paramMap ={lr.maxIter:20}
  30. #覆盖掉
  31. paramMap[lr.maxIter]=30# Specify 1 Param, overwriting the original maxIter.
  32. #更新参数对
  33. paramMap.update({lr.regParam:0.1, lr.threshold:0.55})# Specify multiple Params.
  34. # You can combine paramMaps, which are python dictionaries.
  35. #新的参数,合并为两组参数对
  36. paramMap2 ={lr.probabilityCol:"myProbability"}# Change output column name
  37. paramMapCombined = paramMap.copy()
  38. paramMapCombined.update(paramMap2)
  39. #重新得到model2并拿出来参数看看
  40. # Now learn a new model using the paramMapCombined parameters.
  41. # paramMapCombined overrides all parameters set earlier via lr.set* methods.
  42. model2 = lr.fit(training, paramMapCombined)
  43. print"Model 2 was fit using parameters: "
  44. print model2.extractParamMap()
  45. #准备测试的数据
  46. # Prepare test data
  47. test = spark.createDataFrame([
  48. (1.0,Vectors.dense([-1.0,1.5,1.3])),
  49. (0.0,Vectors.dense([3.0,2.0,-0.1])),
  50. (1.0,Vectors.dense([0.0,2.2,-1.5]))],["label","features"])
  51. # Make predictions on test data using the Transformer.transform() method.
  52. # LogisticRegression.transform will only use the 'features' column.
  53. # Note that model2.transform() outputs a "myProbability" column instead of the usual
  54. # 'probability' column since we renamed the lr.probabilityCol parameter previously.
  55. prediction = model2.transform(test)
  56. #得到预测的DataFrame打印出预测中的选中列
  57. selected = prediction.select("features","label","myProbability","prediction")
  58. for row in selected.collect():
  59. print row
 
 
下面是一个PipeLine的实例:
 
  1. from pyspark.ml importPipeline
  2. from pyspark.ml.classification importLogisticRegression
  3. from pyspark.ml.feature importHashingTF,Tokenizer
  4. #准备测试数据
  5. # Prepare training documents from a list of (id, text, label) tuples.
  6. training = spark.createDataFrame([
  7. (0L,"a b c d e spark",1.0),
  8. (1L,"b d",0.0),
  9. (2L,"spark f g h",1.0),
  10. (3L,"hadoop mapreduce",0.0)],["id","text","label"])
  11. #构建机器学习流水线
  12. # Configure an ML pipeline, which consists of three stages: tokenizer, hashingTF, and lr.
  13. tokenizer =Tokenizer(inputCol="text", outputCol="words")
  14. hashingTF =HashingTF(inputCol=tokenizer.getOutputCol(), outputCol="features")
  15. lr =LogisticRegression(maxIter=10, regParam=0.01)
  16. pipeline =Pipeline(stages=[tokenizer, hashingTF, lr])
  17. #训练出model
  18. # Fit the pipeline to training documents.
  19. model = pipeline.fit(training)
  20. #测试数据
  21. # Prepare test documents, which are unlabeled (id, text) tuples.
  22. test = spark.createDataFrame([
  23. (4L,"spark i j k"),
  24. (5L,"l m n"),
  25. (6L,"mapreduce spark"),
  26. (7L,"apache hadoop")],["id","text"])
  27. #预测,打印出想要的结果
  28. # Make predictions on test documents and print columns of interest.
  29. prediction = model.transform(test)
  30. selected = prediction.select("id","text","prediction")
  31. for row in selected.collect():
  32. print(row)





posted @ 2016-08-16 16:01  kongchung  阅读(12360)  评论(0编辑  收藏  举报