H2O

官网：http://www.h2o.ai/

H2o开源的机器学习框架，支持glm，rf，gbm，深度学习等算法，借助hadoop spark计算平台，实现large scale 机器学习

H2o 机器学习包http://s6/mw690/001N0mEhzy72ZqEp4Vvb5&690

R版本 https://cran.r-project.org/web/packages/h2o/index.html

Python版本 https://pypi.python.org/pypi/h2o/

文档 http://h2o-release.s3.amazonaws.com/h2o/rel-turchin/9/docs-website/h2o-docs/index.html

 

基于h2o的gbm参数调整实验

 

https://github.com/h2oai/h2o-3/blob/master/h2o-docs/src/product/tutorials/gbm/gbmTuning.Rmd

1 下载安装 R h2o

install.packages("h2o")

2 启动h2o

> library(h2o)

> h2o.init(nthreads = 3)

 Connection successful!

 

R is connected to the H2O cluster:

    H2O cluster uptime:         5 hours 9 minutes

    H2O cluster version:        3.8.2.6

    H2O cluster name:           H2O_started_from_R_xxx_phg216

    H2O cluster total nodes:    1

    H2O cluster total memory:   1.42 GB

    H2O cluster total cores:    4

    H2O cluster allowed cores:  3

    H2O cluster healthy:        TRUE

    H2O Connection ip:          localhost

    H2O Connection port:        54321

    H2O Connection proxy:       NA

    R Version:                  R version 3.2.3 (2015-12-10)

 

3 导入数据

## 'path' can point to a local file, hdfs, s3, nfs, Hive, directories, etc.

df <- h2o.importFile(path = "http://s3.amazonaws.com/h2o-public-test-data/smalldata/gbm_test/titanic.csv")

dim(df)

head(df)

tail(df)

summary(df,exact_quantiles=TRUE)

 

## pick a response for the supervised problem

response <- "survived"

 

## the response variable is an integer, we will turn it into a categorical/factor for binary classification

df[[response]] <- as.factor(df[[response]])          

 

## use all other columns (except for the name) as predictors

predictors <- setdiff(names(df), c(response, "name"))

4 数据切分

splits <- h2o.splitFrame(

  data = df,

  ratios = c(0.6,0.2),   ## only need to specify 2 fractions, the 3rd is implied

  destination_frames = c("train.hex", "valid.hex", "test.hex"), seed = 1234

)

train <- splits[[1]]

valid <- splits[[2]]

test  <- splits[[3]]

 

5 基础模型

## We only provide the required parameters, everything else is default

gbm <- h2o.gbm(x = predictors, y = response, training_frame = train)

 

## Show a detailed model summary

gbm

 

## Get the AUC on the validation set

h2o.auc(h2o.performance(gbm, newdata = valid))

6 模型调参优化

## Depth 10 is usually plenty of depth for most datasets, but you never know

hyper_params = list( max_depth = seq(1,29,2) )

#hyper_params = list( max_depth = c(4,6,8,12,16,20) ) ##faster for larger datasets

grid <- h2o.grid(

  ## hyper parameters

  hyper_params = hyper_params,

  ## full Cartesian hyper-parameter search

  search_criteria = list(strategy = "Cartesian"),

  ## which algorithm to run

  algorithm="gbm",

  ## identifier for the grid, to later retrieve it

  grid_id="depth_grid",

  ## standard model parameters

  x = predictors, 

  y = response, 

  training_frame = train, 

  validation_frame = valid,

  ## more trees is better if the learning rate is small enough 

  ## here, use "more than enough" trees - we have early stopping

  ntrees = 10000,                                                            

  ## smaller learning rate is better

  ## since we have learning_rate_annealing, we can afford to start with a bigger learning rate

  learn_rate = 0.05,                                                         

  ## learning rate annealing: learning_rate shrinks by 1% after every tree 

  ## (use 1.00 to disable, but then lower the learning_rate)

  learn_rate_annealing = 0.99,                                               

  ## sample 80% of rows per tree

  sample_rate = 0.8,                                                       

  ## sample 80% of columns per split

  col_sample_rate = 0.8, 

  ## fix a random number generator seed for reproducibility

  seed = 1234,                                                             

  ## early stopping once the validation AUC doesn't improve by at least 0.01% for 5 consecutive scoring events

  stopping_rounds = 5,

  stopping_tolerance = 1e-4,

  stopping_metric = "AUC", 

  ## score every 10 trees to make early stopping reproducible (it depends on the scoring interval)

  score_tree_interval = 10                                                

)

## by default, display the grid search results sorted by increasing logloss (since this is a classification task)

grid                                                                       

## sort the grid models by decreasing AUC

sortedGrid <- h2o.getGrid("depth_grid", sort_by="auc", decreasing = TRUE)    

sortedGrid

## find the range of max_depth for the top 5 models

topDepths = sortedGrid@summary_table$max_depth[1:5]                       

minDepth = min(as.numeric(topDepths))

maxDepth = max(as.numeric(topDepths))

 

hyper_params = list( 

  ## restrict the search to the range of max_depth established above

  max_depth = seq(minDepth,maxDepth,1),                                      

  ## search a large space of row sampling rates per tree

  sample_rate = seq(0.2,1,0.01),                                             

  ## search a large space of column sampling rates per split

  col_sample_rate = seq(0.2,1,0.01),                                         

  ## search a large space of column sampling rates per tree

  col_sample_rate_per_tree = seq(0.2,1,0.01),                                

  ## search a large space of how column sampling per split should change as a function of the depth of the split

  col_sample_rate_change_per_level = seq(0.9,1.1,0.01),                      

  ## search a large space of the number of min rows in a terminal node

  min_rows = 2^seq(0,log2(nrow(train))-1,1),                                 

  ## search a large space of the number of bins for split-finding for continuous and integer columns

  nbins = 2^seq(4,10,1),                                                     

  ## search a large space of the number of bins for split-finding for categorical columns

  nbins_cats = 2^seq(4,12,1),                                                

  ## search a few minimum required relative error improvement thresholds for a split to happen

  min_split_improvement = c(0,1e-8,1e-6,1e-4),                               

  ## try all histogram types (QuantilesGlobal and RoundRobin are good for numeric columns with outliers)

  histogram_type = c("UniformAdaptive","QuantilesGlobal","RoundRobin")       

)

search_criteria = list(

  ## Random grid search

  strategy = "RandomDiscrete",      

  ## limit the runtime to 60 minutes

  max_runtime_secs = 3600,         

  ## build no more than 100 models

  max_models = 100,                  

  ## random number generator seed to make sampling of parameter combinations reproducible

  seed = 1234,                        

  ## early stopping once the leaderboard of the top 5 models is converged to 0.1% relative difference

  stopping_rounds = 5,                

  stopping_metric = "AUC",

  stopping_tolerance = 1e-3

)

grid <- h2o.grid(

  ## hyper parameters

  hyper_params = hyper_params,

  ## hyper-parameter search configuration (see above)

  search_criteria = search_criteria,

  ## which algorithm to run

  algorithm = "gbm",

  ## identifier for the grid, to later retrieve it

  grid_id = "final_grid", 

  ## standard model parameters

  x = predictors, 

  y = response, 

  training_frame = train, 

  validation_frame = valid,

  ## more trees is better if the learning rate is small enough

  ## use "more than enough" trees - we have early stopping

  ntrees = 10000,                                                            

  ## smaller learning rate is better

  ## since we have learning_rate_annealing, we can afford to start with a bigger learning rate

  learn_rate = 0.05,                                                         

  ## learning rate annealing: learning_rate shrinks by 1% after every tree 

  ## (use 1.00 to disable, but then lower the learning_rate)

  learn_rate_annealing = 0.99,                                               

  ## early stopping based on timeout (no model should take more than 1 hour - modify as needed)

  max_runtime_secs = 3600,                                                 

  ## early stopping once the validation AUC doesn't improve by at least 0.01% for 5 consecutive scoring events

  stopping_rounds = 5, stopping_tolerance = 1e-4, stopping_metric = "AUC", 

  ## score every 10 trees to make early stopping reproducible (it depends on the scoring interval)

  score_tree_interval = 10,                                                

  ## base random number generator seed for each model (automatically gets incremented internally for each model)

  seed = 1234                                                             

)

## Sort the grid models by AUC

sortedGrid <- h2o.getGrid("final_grid", sort_by = "auc", decreasing = TRUE)    

sortedGrid

 

7 模型验证和测试

gbm <- h2o.getModel(sortedGrid@model_ids[[1]])

print(h2o.auc(h2o.performance(gbm, newdata = test)))

 

交叉验证

for (i in 1:5) {

  gbm <- h2o.getModel(sortedGrid@model_ids[[i]])

  cvgbm <- do.call(h2o.gbm,

        ## update parameters in place

        {

          p <- gbm@parameters

          p$model_id = NULL          ## do not overwrite the original grid model

          p$training_frame = df      ## use the full dataset

          p$validation_frame = NULL  ## no validation frame

          p$nfolds = 5               ## cross-validation

          p

        }

  )

  print(gbm@model_id)

  print(cvgbm@model$cross_validation_metrics_summary[5,]) ## Pick out the "AUC" row

}

 

gbm <- h2o.getModel(sortedGrid@model_ids[[1]])

preds <- h2o.predict(gbm, test)

head(preds)

gbm@model$validation_metrics@metrics$max_criteria_and_metric_scores

 

web ui查看各种结果，模型，评估等等

http://localhost:54321/flow/index.html

 http://s2/mw690/001N0mEhzy72Zr9lyQF11&690

结果保存

h2o.saveModel(gbm, "/tmp/bestModel.csv", force=TRUE)

h2o.exportFile(preds, "/tmp/bestPreds.csv", force=TRUE)

 

8 模型部署

h2o.download_pojo(gbm)

得到pojo代码

通过java 打分服务，可以将模型部署到实际工业应用场景

（1）下载pojo代码  javabean

$ mkdir experiment

$ cd experiment

$ mv ~/Downloads/gbm_pojo_test.java .

$ curl http://localhost:54321/3/h2o-genmodel.jar > h2o-genmodel.jar

（2）编写 打分程序

import java.io.*;

import hex.genmodel.easy.RowData;

import hex.genmodel.easy.EasyPredictModelWrapper;

import hex.genmodel.easy.prediction.*;

 

public class main {

  private static String modelClassName = "gbm_pojo_test";

 

  public static void main(String[] args) throws Exception {

    hex.genmodel.GenModel rawModel;

    rawModel = (hex.genmodel.GenModel) Class.forName(modelClassName).newInstance();

    EasyPredictModelWrapper model = new EasyPredictModelWrapper(rawModel);

 

    RowData row = new RowData();

    row.put("Year", "1987");

    row.put("Month", "10");

    row.put("DayofMonth", "14");

    row.put("DayOfWeek", "3");

    row.put("CRSDepTime", "730");

    row.put("UniqueCarrier", "PS");

    row.put("Origin", "SAN");

    row.put("Dest", "SFO");

 

    BinomialModelPrediction p = model.predictBinomial(row);

    System.out.println("Label (aka prediction) is flight departure delayed: " + p.label);

    System.out.print("Class probabilities: ");

    for (int i = 0; i < p.classProbabilities.length; i++) {

      if (i > 0) {

        System.out.print(",");

      }

      System.out.print(p.classProbabilities[i]);

    }

    System.out.println("");

  }

}

（3）编译，输出打分结果

$ javac -cp h2o-genmodel.jar -J-Xmx2g -J-XX:MaxPermSize=128m gbm_pojo_test.java main.java

$ java -cp .:h2o-genmodel.jar main

The following output displays:

        Label (aka prediction) is flight departure delayed: YES

        Class probabilities: 0.4790490513429604,0.5209509486570396

 

9 关闭h2o 集群

> h2o.shutdown()

Are you sure you want to shutdown the H2O instance running at http://localhost:54321/ (Y/N)? n

[1] TRUE

 

参考：

[1] http://blog.h2o.ai/

[2] https://github.com/h2oai

总结，相对于其他开源的机器学习算法包，h2o是一个机器学习产品，更加好用适用，从实际问题出发，结合茶品的思维，开发实现的机器学习框架，适合工业应用。