Productionizing H2O

About POJOs and MOJOs

h2o allows you to convert the models you have built to either a (pojo) or a model object, optimized (mojo).

H2O-generated MOJO and POJO models are intended to be easily embeddable in any Java environment. The only compilation and runtime dependency for a generated model is the h2o-genmodel.jar file produced as the build output of these packages. This file is a library that supports scoring. For POJOs, it contains the base classes from which the POJO is derived from. (You can see “extends GenModel” in a POJO class. The GenModel class is part of this library.) For MOJOs, it also contains the required readers and interpreters. The h2o-genmodel.jar file is required when POJO/MOJO models are deployed to production.

users can refer to the quick start topics that follow for more information about generating pojos and mojos.

Developers can refer to the the POJO and MOJO Model Javadoc.

MOJO Quick Start

This section describes how to build and implement a MOJO (Model Object, Optimized) to use predictive scoring. Java developers should refer to the Javadoc for more information, including packages.

What is a MOJO?

a mojo (model object, optimized) is an alternative to h2o’s pojo. as with pojos, h2o allows you to convert models that you build to mojos, which can then be deployed for scoring in real time.

Notes:

  • MOJOs are supported for AutoML, Deep Learning, DRF, GBM, GLM, GLRM, K-Means, PCA, Stacked Ensembles, SVM, Word2vec, and XGBoost models.
  • MOJOs are only supported for encodings that are either default or enum.
  • MOJO predict cannot parse columns enclosed in double quotes (for example, “”2””).

Benefits of MOJOs over POJOs

宁夏微乐滑水麻将下载while pojos continue to be supported, some customers encountered issues with large pojos not compiling. (note that pojos are not supported for source files larger than 1g.) mojos do not have a size restriction and address the size issue by taking the tree out of the pojo and using generic tree-walker code to navigate the model. the resulting executable is much smaller and faster than a pojo.

宁夏微乐滑水麻将下载at large scale, new models are roughly 20-25 times smaller in disk space, 2-3 times faster during “hot” scoring (after jvm is able to optimize the typical execution paths), and 10-40 times faster in “cold” scoring (when jvm doesn’t know yet know the execution paths) compared to pojos. the efficiency gains are larger the bigger the size of the model.

h2o conducted in-house testing using models with 5000 trees of depth 25. at very small scale (50 trees / 5 depth), pojos were found to perform ≈10% faster than mojos for binomial and regression models, but 50% slower than mojos for multinomial models.

Building a MOJO

mojos are built in much the same way as pojos. the example code below shows how to start h2o, build a model using either r or python, and then compile and run the mojo. this example uses gbm, but any supported algorithm can be used to build a model and run the mojo.

The examples below describe how to start H2O and create a model using R and Python. The download_mojo() function saves the model as a zip file. You can unzip the file to view the options used to build the file along with each tree built in the model. Note that each tree file is saved as a binary file type.

Step 1: Build and Extract a Model

# 1. Open a terminal window and start R.

# 2. Run the following commands to build a simple GBM model.
library(h2o)
h2o.init(nthreads=-1)
path <- system.file("extdata", "prostate.csv", package="h2o")
h2o_df <- h2o.importFile(path)
h2o_df$CAPSULE <- as.factor(h2o_df$CAPSULE)
model <- h2o.gbm(y="CAPSULE",
                 x=c("AGE", "RACE", "PSA", "GLEASON"),
                 training_frame=h2o_df,
                 distribution="bernoulli",
                 ntrees=100,
                 max_depth=4,
                 learn_rate=0.1)

# Download the MOJO and the resulting h2o-genmodel.jar file
# to a new **experiment** folder. Note that the ``h2o-genmodel.jar`` file
# is a library that supports scoring and contains the required readers
# and interpreters. This file is required when MOJO models are deployed
# to production. Be sure to specify the entire path, not just the relative path.
modelfile <- h2o.download_mojo(model, path="~/experiments/", get_genmodel_jar=TRUE)
print("Model saved to " + modelfile)
Model saved to /Users/user/GBM_model_R_1475248925871_74.zip"
# 1. Open a terminal window and start python.

# 2. Run the following commands to build a simple GBM model.
# The model, along with the **h2o-genmodel.jar** file will
# then be downloaded to an **experiment** folder.
import h2o
from h2o.estimators.gbm import H2OGradientBoostingEstimator
h2o.init()
h2o_df = h2o.load_dataset("prostate.csv")
h2o_df["CAPSULE"] = h2o_df["CAPSULE"].asfactor()
model=H2OGradientBoostingEstimator(distribution="bernoulli",
                                   ntrees=100,
                                   max_depth=4,
                                   learn_rate=0.1)
model.train(y="CAPSULE",
            x=["AGE","RACE","PSA","GLEASON"],
            training_frame=h2o_df)

# Download the MOJO and the resulting ``h2o-genmodel.jar`` file
# to a new **experiment** folder. Note that the ``h2o-genmodel.jar`` file
# is a library that supports scoring and contains the required readers
# and interpreters. This file is required when MOJO models are deployed
# to production. Be sure to specify the entire path, not just the relative path.
modelfile = model.download_mojo(path="~/experiment/", get_genmodel_jar=True)
print("Model saved to " + modelfile)
Model saved to /Users/user/GBM_model_python_1475248925871_888.zip
// Compile the source:
javac -classpath ~/h2o/h2o-3.20.0.1/h2o.jar src/h2oDirect/h2oDirect.java

// Execute as a classfile. This also downloads the LoanStats4 demo,
// which trains a GBM model.
Erics-MBP-2:h2oDirect ericgudgion$ java -cp /Users/ericgudgion/NetBeansProjects/h2oDirect/src/:/Users/ericgudgion/h2o/h2o-3.20.0.1/h2o.jar h2oDirect.h2oDirect /Demos/Lending-Club/LoanStats4.csv
...
06-14 20:40:29.420 192.168.1.160:54321   55005  main      INFO: Found XGBoost backend with library: xgboost4j_minimal
06-14 20:40:29.428 192.168.1.160:54321   55005  main      INFO: Your system supports only minimal version of XGBoost (no GPUs, no multithreading)!
06-14 20:40:29.428 192.168.1.160:54321   55005  main      INFO: ----- H2O started  -----
06-14 20:40:29.428 192.168.1.160:54321   55005  main      INFO: Build git branch: rel-wright
...
...
Starting H2O with IP 192.168.1.160:54321
Loading data from file
...
Loaded file /Demos/Lending-Club/LoanStats4.csv size 3986423 Cols:19 Rows:39029
...
Creating GBM Model
Training Model
...
Training Results
Model Metrics Type: Binomial
 Description: N/A
 model id: GBM_model_1529023227180_1
 frame id: dataset-key
 MSE: 0.11255783
 RMSE: 0.3354964
 AUC: 0.82892376
 logloss: 0.36827797
 mean_per_class_error: 0.26371866
 default threshold: 0.261136531829834
...
Model AUC 0.8289237508508612
Model written out as a mojo to file /Demos/Lending-Club/LoanStats4.csv.zip

// Save as h2oDirect.java
package h2oDirect;

import hex.tree.gbm.GBM;
import hex.tree.gbm.GBMModel;
import hex.tree.gbm.GBMModel.GBMParameters;
import java.io.FileOutputStream;
import java.io.IOException;
import java.net.InetAddress;
import water.Key;
import water.fvec.Frame;
import water.fvec.NFSFileVec;
import water.parser.ParseDataset;
import water.*;


public class h2oDirect {


    /**
     * @param args the command line arguments
     */
    public static void main(String[] args) throws IOException {

      String h2oargs = "-nthreads -1 ";
      H2OApp.main(h2oargs.split(" "));
      System.out.println("Starting H2O with IP "+H2O.getIpPortString());

      H2O.waitForCloudSize(1, 3000);

      System.out.println("Loading data from file ");
      String inputfile = args[0];
      NFSFileVec datafile = NFSFileVec.make(inputfile);
      Frame dataframe = ParseDataset.parse(Key.make("dataset-key") , datafile._key);
      System.out.println("Loaded file "+inputfile+" size "+datafile.byteSize()+" Cols:"+dataframe.numCols()+" Rows:"+dataframe.numRows());


      for (int v=0; v<dataframe.numCols(); v++) {
      System.out.println(dataframe.name(v)+" "+dataframe.vec(v).get_type_str());
      }

      int c = dataframe.find("bad_loan");

      dataframe.replace(c, dataframe.vec(c).toCategoricalVec());


      // drop the id and member_id columns from model
      dataframe.remove(dataframe.find("id"));
      dataframe.remove(dataframe.find("member_id"));

      System.out.println("Creating GBM Model");

      GBMParameters modelparms = new GBMParameters();
      modelparms._train = dataframe._key;
      modelparms._response_column = "bad_loan";

      System.out.println("Training Model");
      GBM model = new GBM(modelparms);
      GBMModel gbm = model.trainModel().get();

      System.out.println("Training Results");
      System.out.println(gbm._output);
      System.out.println("Model AUC "+gbm.auc());


      String outputfile = inputfile+".zip";
      FileOutputStream modeloutput = new FileOutputStream(outputfile);
      gbm.getMojo().writeTo(modeloutput);
      modeloutput.close();
      System.out.println("Model written out as a mojo to file "+outputfile);

      System.out.println("H2O shutdown....");
      H2O.shutdown(0);

    }

}
import water.rapids.ast.prims.advmath.AstCorrelation

object RandomForestFileInput {

  import water.H2O
  import water.H2OApp
  import water.fvec.Vec
  import water.fvec.NFSFileVec
  import water.fvec._

  import hex.tree.drf.DRF
  import hex.tree.drf.DRFModel
  import hex.tree.drf.DRFModel.DRFParameters
  import water.parser.ParseDataset
  import water.Key
  import water.Futures
  import water._

  import scala.io.Source
  import scala.reflect._

  import java.io.FileOutputStream
  import java.io.FileWriter

     def main(args: Array[String]): Unit = {
      println("H2O Random Forest FileInput example\n")

      if (args.length==0) {
        println("Input file missing, please pass datafile as the first parameter")
        return
      }

      // Start H2O instance and wait for 3 seconds for instance to complete startup
      println("Starting H2O")
      val h2oargs = "-nthreads -1 -quiet"

      H2OApp.main(h2oargs.split(" "))
      H2O.waitForCloudSize(1, 3000)

      println("H2O available")

      // Load datafile passed as first parameter and print the size of the file as confirmation
      println("Loading data from file ")
      val inputfile = args(0)
      val parmsfile = args(1)
      def ignore: Boolean = System.getProperty("ignore","false").toBoolean

      val datafile = NFSFileVec.make(inputfile)
      val dataframe = ParseDataset.parse(Key.make("dataset-key") , datafile._key)
      println("Loaded file "+inputfile+" size "+datafile.byteSize()+" Cols:"+dataframe.numCols()+" Rows:"+dataframe.numRows())

      println(dataframe.anyVec().get_type_str)

      for (v <- 0 to dataframe.numCols()-1) {
        println(dataframe.name(v))
      }

      val c = dataframe.find("bad_loan")
      dataframe.replace(c, dataframe.vecs()(c).toCategoricalVec())

      // drop the id and member_id columns from model
      dataframe.remove(dataframe.find("id"))
      dataframe.remove(dataframe.find("member_id"))


      // set Random Forest parameters
      println("creating model parameters")
      var modelparams = new DRFParameters()
      var fields = modelparams.getClass.getFields

      for (line <- Source.fromFile(parmsfile).getLines) {
          println("Reading parameter from file: "+line)
          var linedata = line.split(" ")


         for(v <- fields){
           if ( v.getName.matches(linedata(0))) {
             val method1 = v.getDeclaringClass.getDeclaredField(linedata(0) )
             method1.setAccessible(true)
             println("Found "+linedata(0)+" Var "+v+" Accessable "+method1.isAccessible()+" Type "+method1.getType )
             v.setAccessible(true)
             v.setInt(modelparams, linedata(1).toInt)
           }
         }
      }


      // hard coded values
      modelparams._train = dataframe._key
      modelparams._response_column = "bad_loan"

       if (ignore) {
         println("Adding fields to ignore from file "+parmsfile+"FieldtoIgnore")
         var ignoreNames = new Array[String](dataframe.numCols())
         var in=0
         for (line <- Source.fromFile(parmsfile+"FieldtoIgnore").getLines) {
           ignoreNames(in) = line
           in+=1
         }
         modelparams._ignored_columns=ignoreNames
       }


      println("Parameters set ")

      // train model
      println("Starting training")
      var job: DRF = new DRF(modelparams)
      var model: DRFModel = job.trainModel().get()

      println("Training completed")

      // training metrics
      println(model._output.toString())
      println("Model AUC: "+model.auc())
      println(model._output._variable_importances)

      // If you want to look at variables that are important and then model on them
      // the following will write them out, then use only those in other model training
      // handy when you have a thousand columns but want to train on only the important ones.
      // Then before calling the model... call modelparams._ignored_columns= Array("inq_last_6mths")
      // FileWriter

       if (ignore) {
         val file = new FileOutputStream(parmsfile + "FieldtoIgnore")

         var n = 0
         var in = 0
         var ignoreNames = new Array[String](dataframe.numCols())
         val fieldnames = model._output._varimp._names
         println("Fields to add to _ignored_columns field")
         for (i <- model._output._varimp.scaled_values()) {
           if (i < 0.3) {
             println(n + " = " + fieldnames(n) + " = " + i)
             Console.withOut(file) {
               println(fieldnames(n))
             }
             ignoreNames(in) = fieldnames(n)
             in += 1
           }
           n += 1
         }
         println("Drop these:")
         for (i <- 0 to in) {
           println(fieldnames(i))
         }
         file.close()
         println()
       }

      // save model
      var outputfile = inputfile+"_model_pojo.txt"
      var modeloutput: FileOutputStream = new FileOutputStream(outputfile)
      println("Saving model to "+outputfile)
      model.toJava(modeloutput, false, true)
      modeloutput.close()

      outputfile = inputfile+"_model_jason.txt"
      modeloutput = new FileOutputStream(outputfile)
      println("Saving Jason to "+outputfile)
      Console.withOut(modeloutput) {  println(model.toJsonString()) }
      modeloutput.close()

      outputfile = inputfile+"_model_mojo.zip"
      modeloutput = new FileOutputStream(outputfile)
      println("Saving mojo to "+outputfile)
      model.getMojo.writeTo(modeloutput)
      modeloutput.close()

       println(models: hex.ensemble.StackedEnsemble )

      println("Completed")
      H2O.shutdown(0)

  }
}

Step 2: Compile and Run the MOJO

  1. Open a new terminal window and change directories to the experiment folder:

    $ cd experiment
    
  2. Create your main program in the experiment宁夏微乐滑水麻将下载 folder by creating a new file called main.java (for example, using “vim main.java”). Include the following contents. Note that this file references the GBM model created above using R.

    import java.io.*;
    import hex.genmodel.easy.RowData;
    import hex.genmodel.easy.EasyPredictModelWrapper;
    import hex.genmodel.easy.prediction.*;
    import hex.genmodel.MojoModel;
    
    public class main {
      public static void main(String[] args) throws Exception {
        EasyPredictModelWrapper model = new EasyPredictModelWrapper(MojoModel.load("GBM_model_R_1475248925871_74.zip"));
    
        RowData row = new RowData();
        row.put("AGE", "68");
        row.put("RACE", "2");
        row.put("DCAPS", "2");
        row.put("VOL", "0");
        row.put("GLEASON", "6");
    
        BinomialModelPrediction p = model.predictBinomial(row);
        System.out.println("Has penetrated the prostatic capsule (1=yes; 0=no): " + 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("");
      }
    }
    

GBM and DRF return classProbabilities, but not all MOJOs will return a classProbabilities field. Refer to the ModelPrediction definition for each algorithm to find the correct field(s) to access. This is available in the H2O-3 GitHub repo at: . You can also view the hex.genmodel.easy.prediction classes in the Javadoc.

In addition to classProbabilities, in GBM and DRF you can choose to generate additional leafNodeAssignments and contributions fields. The leafNodeAssignments field will show the decision path through each tree. The contributions宁夏微乐滑水麻将下载 field will provide Shapley contributions. Note that these fields may slow down the MOJO as they add computation. Below is the Java code showing how return both the leaf node assignment and the contributions:

import java.io.*;
import hex.genmodel.easy.RowData;
import hex.genmodel.easy.EasyPredictModelWrapper;
import hex.genmodel.easy.prediction.*;
import hex.genmodel.MojoModel;

public class main {
  public static void main(String[] args) throws Exception {
    EasyPredictModelWrapper.Config config = new EasyPredictModelWrapper.Config()
       .setModel(MojoModel.load("GBM_model_R_1475248925871_74.zip"))
       .setEnableLeafAssignment(true)
       .setEnableContributions(true);
    EasyPredictModelWrapper model = new EasyPredictModelWrapper(config);

    RowData row = new RowData();
    row.put("AGE", "68");
    row.put("RACE", "2");
    row.put("DCAPS", "2");
    row.put("VOL", "0");
    row.put("GLEASON", "6");

    BinomialModelPrediction p = model.predictBinomial(row);
    System.out.println("Has penetrated the prostatic capsule (1=yes; 0=no): " + 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("Leaf node assignments: ");
    for (int i=0; i < p.leafNodeAssignments; i++) {
      if (i > 0) {
        System.out.print(p.leafNodeAssignments[i]);
      }
    }
    System.out.println("");

    System.out.println("Shapley contributions: ");
    for (int i=0; i < p.contributions; i++) {
      if (i > 0) {
        System.out.print(",");
      }
      System.out.print(p.contributions[i]);
    }
    System.out.println("");
  }
}

宁夏微乐滑水麻将下载for glrm, the returned field is the x coefficients for the archetypes by default. in addition to that, you can choose to generate the reconstructed data row as well. again, this may slow down the mojo due to added computation. below is the java code showing how to obtain both the x factors and the reconstructed data after you have generated the glrm mojo:

import java.io.*;
import hex.genmodel.easy.RowData;
import hex.genmodel.easy.EasyPredictModelWrapper;
import hex.genmodel.easy.prediction.*;
import hex.genmodel.MojoModel;

public class main {
  public static void main(String[] args) throws Exception {
  EasyPredictModelWrapper.Config config = new EasyPredictModelWrapper.Config().setModel(MojoModel.load("GLRM_model_python_1530295749484_1.zip")).setEnableGLRMReconstrut(true);
  EasyPredictModelWrapper model = new EasyPredictModelWrapper(config);

  RowData row = new RowData();
  row.put("CAPSULE", "0");
  row.put("AGE", "68");
  row.put("RACE", "2");
  row.put("DPROS", "4");
  row.put("DCAPS", "2");
  row.put("PSA", "31.9");
  row.put("VOL", "0");
  row.put("GLEASON", "6");

  DimReductionModelPrediction p = model.predictDimReduction(row);
  String[] colnames = model.m.getNames();
  System.out.println("X coefficients for input row: ");
  for (int i = 0; i < p.dimensions.length; i++) {
    if (i > 0) {
      System.out.println(",");
    }
    System.out.print("Arch "+i+" coefficient: "+p.dimensions[i]);
  }
  System.out.println("");
  System.out.println("Reconstructed input row: ");
  for (int i = 0; i < p.reconstructed.length; i++) {
    if (i > 0) {
      System.out.println(",");
    }
    System.out.print(colnames[i]+": "+p.reconstructed[i]);
  }
  System.out.println("");
}
  1. compile in terminal window 2.

    $ javac -cp h2o-genmodel.jar -J-Xms2g -J-XX:MaxPermSize=128m main.java
    
  2. run in terminal window 2.

    # Linux and OS X users
    $ java -cp .:h2o-genmodel.jar main
    
    # Windows users
    $ java -cp .;h2o-genmodel.jar main
    

宁夏微乐滑水麻将下载the following output displays:

Has penetrated the prostatic capsule (1 yes; 0 no): 0
Class probabilities: 0.8059929056296662,0.19400709437033375

if you have chosen to enable leaf node assignments, you will also see 100 leaf node assignments for your data row:

Has penetrated the prostatic capsule (1 yes; 0 no): 0
Class probabilities: 0.8059929056296662,0.19400709437033375
Leaf node assignments:   RRRR,RRR,RRRR,RRR,RRL,RRRR,RLRR,RRR,RRR,RRR,RLRR,...

for the glrm mojo, after running the java code, you will see the following:

X coefficients for input row:
Arch 0 coefficient: -0.5930494611027051,
Arch 1 coefficient: 1.0459847877909487,
Arch 2 coefficient: 0.5849220609025815
Reconstructed input row:
CAPSULE: 0.5204822003860688,
AGE: 10.520294102886806,
RACE: 4.1422863477607645,
DPROS: 2.970424071063664,
DCAPS: 6.361196172145799,
PSA: 1.905415090602722,
VOL: 0.7123169431687857,
GLEASON: 6.625024806196047

Viewing a MOJO Model

宁夏微乐滑水麻将下载a java tool for converting binary mojo files into human viewable graphs is packaged with h2o. this tool produces output that “dot” (which is part of graphviz) can turn into an image. (see the for more information.)

宁夏微乐滑水麻将下载here is an example output for a gbm model:

GBM MOJO model

宁夏微乐滑水麻将下载the following code snippet shows how to download a mojo from r and run the printmojo tool on the command line to make a .png file. to better control the look and feel of your tree, we provide two options for printmojo:

  • --decimalplaces (or -d) allows you to control the number of decimal points shown for numbers.
  • --fontsize (or -f) controls the font size. The default font size is 14. When using this option, be careful not to choose a font size that is so large that you cannot see your whole tree. We recommend using a font size no larger than 20.
library(h2o)
h2o.init()
df <- h2o.importFile("http://s3.amazonaws.com/h2o-public-test-data/smalldata/airlines/allyears2k_headers.zip")
model <- h2o.gbm(model_id = "model",
                training_frame = df,
                x = c("Year", "Month", "DayofMonth", "DayOfWeek", "UniqueCarrier"),
                y = "IsDepDelayed",
                max_depth = 3,
                ntrees = 5)
h2o.download_mojo(model, getwd(), FALSE)

# Now download the latest stable h2o release from http://www.sunuq.icu/download/
# and run the PrintMojo tool from the command line.
#
# (For MacOS: brew install graphviz)
java -cp h2o.jar hex.genmodel.tools.PrintMojo --tree 0 -i model.zip -o model.gv -f 20 -d 3
dot -Tpng model.gv -o model.png
open model.png

FAQ

  • Are MOJOs thread safe?
Yes, all of H2O-3 MOJOs are thread safe.
  • How can I use an XGBoost MOJO with Maven?

宁夏微乐滑水麻将下载if you declare a dependency on h2o-genmodel, then you also have to include the h2o-genmodel-ext-xgboost dependency if you are planning to use xgboost models. for example:

<groupId>ai.h2o</groupId>
<artifactId>xgboost-mojo-example</artifactId>
<version>1.0-SNAPSHOT</version>

dependency>
    <groupId>ai.h2o</groupId>
    <artifactId>h2o-genmodel-ext-xgboost</artifactId>
    <version>3.18.0.8</version>
</dependency>
<dependency>
    <groupId>ai.h2o</groupId>
    <artifactId>h2o-genmodel</artifactId>
    <version>3.18.0.8</version>
</dependency>

POJO Quick Start

This section describes how to build and implement a POJO to use predictive scoring. Java developers should refer to the Javadoc宁夏微乐滑水麻将下载 for more information, including packages.

Notes:

  • POJOs are not supported for source files larger than 1G. For more information, refer to the POJO FAQ section below. POJOs are also not supported for GLRM or Stacked Ensembles models.
  • POJO predict cannot parse columns enclosed in double quotes (for example, “”2””).

What is a POJO?

h2o allows you to convert the models you have built to a plain old java object (pojo), which can then be easily deployed within your java app and scheduled to run on a specified dataset.

宁夏微乐滑水麻将下载pojos allow users to build a model using h2o and then deploy the model to score in real-time, using the pojo model or a rest api call to a scoring server.

The only compilation and runtime dependency for a generated model is the h2o-genmodel.jar file produced as the build output of these packages. This file is a library that supports scoring, and it contains the base classes from which the POJO is derived from. (You can see “extends GenModel” in a pojo class. The GenModel class is part of this library.) The h2o-genmodel.jar宁夏微乐滑水麻将下载 file is required when POJO models are deployed to production.

Building a POJO

宁夏微乐滑水麻将下载the example code below shows how to start h2o flow, compile and run an example flow, and then compile and run the pojo. this example uses gbm, but any supported algorithm can be used to build a model and run the pojo.

  1. 宁夏微乐滑水麻将下载start h2o in terminal window #1:

    $ java -jar h2o.jar

  2. 宁夏微乐滑水麻将下载build a model using your web browser:

    1. Go to
    2. Click View Example Flows near the right edge of the screen. Here is a screenshot of what to look for:
    View Example Flows
    1. Click GBM_Airlines_Classification.flow
    2. If a confirmation prompt appears asking you to “Load Notebook”, click it.
    3. From the “Flow” menu choose the “Run all cells” option
    4. Scroll down and find the “Model” cell in the notebook. Click on the Download POJO button as shown in the following screenshot:
    Download POJO

    Note宁夏微乐滑水麻将下载: The instructions below assume that the POJO model was downloaded to the “Downloads” folder.

  3. Download model pieces in a new terminal window. Note that H2O must still be running in terminal window #1:

    $ mkdir experiment
    $ cd experiment
    $ mv ~/Downloads/gbm_pojo_test.java .
    $ curl http://localhost:54321/3/h2o-genmodel.jar > h2o-genmodel.jar
    
  4. Create your main program in terminal window #2 by creating a new file called main.java (vim main.java宁夏微乐滑水麻将下载) with the following contents:

    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("");
      }
    }
    
  5. compile the pojo in terminal window 2:

    $ javac -cp h2o-genmodel.jar -J-Xmx2g -J-XX:MaxPermSize=128m gbm_pojo_test.java main.java
    
  6. run the pojo in terminal window 2.

For Linux and OS X users: $ java -cp .:h2o-genmodel.jar main

For Windows users: $ java -cp .;h2o-genmodel.jar main

the following output displays:

Label (aka prediction) is flight departure delayed: YES
Class probabilities: 0.4319916897116479,0.5680083102883521

Extracting Models from H2O

宁夏微乐滑水麻将下载generated models can be extracted from h2o in the following ways:

From the H2O Flow Web UI

When viewing a model, click the Download POJO button at the top of the model cell, as shown in the example in the Quick start section. You can also preview the POJO inside Flow, but it will only show the first thousand lines or so in the web browser, truncating large models.

From R or Python

the following code snippets show an example of h2o building a model and downloading its corresponding pojo from an r script and a python script.

library(h2o)
h2o.init()
path <- system.file("extdata", "prostate.csv", package = "h2o")
h2o_df <- h2o.importFile(path)
h2o_df$CAPSULE <- as.factor(h2o_df$CAPSULE)
model <- h2o.glm(y = "CAPSULE",
                x = c("AGE", "RACE", "PSA", "GLEASON"),
                training_frame = h2o_df,
                family = "binomial")
h2o.download_pojo(model)
import h2o
h2o.init()
from h2o.estimators.glm import H2OGeneralizedLinearEstimator
path = "http://s3.amazonaws.com/h2o-public-test-data/smalldata/prostate/prostate.csv.zip"
h2o_df = h2o.import_file(path)
h2o_df['CAPSULE'] = h2o_df['CAPSULE'].asfactor()
model = H2OGeneralizedLinearEstimator(family = "binomial")
model.train(y = "CAPSULE",
            x = ["AGE", "RACE", "PSA", "GLEASON"],
            training_frame = h2o_df)
h2o.download_pojo(model)

Use Cases

宁夏微乐滑水麻将下载the following use cases are demonstrated with code examples:

  • Reading new data from a CSV file and predicting on it: The PredictCsv class is used by the H2O test harness to make predictions on new data points.
  • Getting a new observation from a JSON request and returning a prediction
  • Calling a user-defined function directly from hive: See the .

FAQ

  • Are POJOs thread safe?
Yes, all of H2O-3 POJOs are thread safe.
  • How do I score new cases in real-time in a production environment?
If you’re using the UI, click the Preview POJO button for your model. This produces a Java class with methods that you can reference and use in your production app.
  • What kind of technology would I need to use?
Anything that runs in a JVM. The POJO is a standalone Java class with no dependencies on H2O.
  • How should I format my data before calling the POJO?

here are our requirements (assuming you are using the “easy” prediction api for the pojo as described in the ).

  • Input columns must only contain categorical levels that were seen during training
  • Any additional input columns not used for training are ignored
  • If no input column is specified, it will be treated as an NA
  • Some models do not handle NAs well (e.g., GLM)
  • Any transformations applied to data before model training must also be applied before calling the POJO predict method
  • How do I run a POJO on a Spark Cluster?
The POJO provides just the math logic to do predictions, so you won’t find any Spark (or even H2O) specific code there. If you want to use the POJO to make predictions on a dataset in Spark, create a map to call the POJO for each row and save the result to a new column, row-by-row.
  • How do I communicate with a remote cluster using the REST API?
You can dl the POJO using the REST API but when calling the POJO predict function, it’s in the same JVM, not across a REST API.
  • Is it possible to make predictions using my H2O cluster with the REST API?
Yes, but this way of making predictions is separate from the POJO. For more information about in-H2O predictions (as opposed to POJO predictions), see the documentation for the H2O REST API endpoint /3/Predictions.
  • Why did I receive the following error when trying to compile the POJO?

宁夏微乐滑水麻将下载the following error is generated when the source file is larger than 1g.

Michals-MBP:b michal$ javac -cp h2o-genmodel.jar -J-Xmx2g -J-XX:MaxPermSize=128m drf_b9b9d3be_cf5a_464a_b518_90701549c12a.java
An exception has occurred in the compiler (1.7.0_60). Please file a bug at the Java Developer Connection (http://java.sun.com/webapps/bugreport)  after checking the Bug Parade for duplicates. Include your program and the following diagnostic in your report.  Thank you.
java.lang.IllegalArgumentException
    at java.nio.ByteBuffer.allocate(ByteBuffer.java:330)
    at com.sun.tools.javac.util.BaseFileManager$ByteBufferCache.get(BaseFileManager.java:308)
    at com.sun.tools.javac.util.BaseFileManager.makeByteBuffer(BaseFileManager.java:280)
    at com.sun.tools.javac.file.RegularFileObject.getCharContent(RegularFileObject.java:112)
    at com.sun.tools.javac.file.RegularFileObject.getCharContent(RegularFileObject.java:52)
    at com.sun.tools.javac.main.JavaCompiler.readSource(JavaCompiler.java:571)
    at com.sun.tools.javac.main.JavaCompiler.parse(JavaCompiler.java:632)
    at com.sun.tools.javac.main.JavaCompiler.parseFiles(JavaCompiler.java:909)
    at com.sun.tools.javac.main.JavaCompiler.compile(JavaCompiler.java:824)
    at com.sun.tools.javac.main.Main.compile(Main.java:439)
    at com.sun.tools.javac.main.Main.compile(Main.java:353)
    at com.sun.tools.javac.main.Main.compile(Main.java:342)
    at com.sun.tools.javac.main.Main.compile(Main.java:333)
    at com.sun.tools.javac.Main.compile(Main.java:76)
    at com.sun.tools.javac.Main.main(Main.java:61)

Example Design Patterns

宁夏微乐滑水麻将下载here is a collection of example design patterns for how to productionize h2o.

Consumer loan application

Characteristic Value
Pattern name Jetty servlet
Example training language R
Example training data source CSV file
Example scoring data source User input to Javascript application running in browser
Scoring environment REST API service provided by Jetty servlet
Scoring engine H2O POJO
Scoring latency SLA Real-time
Resource Location
Git repos
Slides http://sunuq.icu/h2o-tutorials/latest-stable/tutorials/building-a-smarter-application/index.html
Videos

Craigslist application

Characteristic Value
Pattern name Sparkling water streaming
Example training language Scala
Example training data source CSV file
Example scoring data source User input to Javascript application running in browser
Scoring engine H2O cluster
Scoring latency SLA Real-time
Resource Location
Git repos
Blogs

Slides

Malicious domain application

Characteristic Value
Pattern name AWS Lambda
Example training language Python
Example training data source CSV file
Example scoring data source User input to Javascript application running in browser
Scoring environment AWS Lambda REST API endpoint
Scoring engine H2O POJO
Scoring latency SLA Real-time
Resource Location
Git repos
Slides
Videos

Storm bolt

Characteristic Value
Pattern name Storm bolt
Example training language R
Example training data source CSV file
Example scoring data source Storm spout
Scoring environment POJO embedded in a Storm bolt
Scoring engine H2O POJO
Scoring latency SLA Real-time
Resource Location
Git repos
Tutorials http://sunuq.icu/h2o-tutorials/latest-stable/tutorials/streaming/storm/index.html

Invoking POJO directly in R

Characteristic Value
Pattern name POJO in R
Example training language R
Example training data source (Need example)
Example scoring data source (Need example)
Scoring environment R
Scoring engine H2O POJO
Scoring latency SLA Batch

Hive UDF

Characteristic Value
Pattern name Hive UDF
Example training language R
Example training data source HDFS directory with hive part files output by a SELECT
Example scoring data source Hive
Scoring environment Hive SELECT query (parallel MapReduce) running UDF
Scoring engine H2O POJO
Scoring latency SLA Batch
Resource Location
Git repos
POJO Tutorial http://sunuq.icu/h2o-tutorials/latest-stable/tutorials/hive_udf_template/hive_udf_pojo_template/index.html
MOJO Tutorial http://sunuq.icu/h2o-tutorials/latest-stable/tutorials/hive_udf_template/hive_udf_mojo_template/index.html

MOJO as a JAR Resource

Characteristic Value
Pattern name MOJO JAR
Example training language R
Example training data source Iris
Example scoring data source Single Row
Scoring environment Portable
Scoring engine H2O MOJO
Scoring latency SLA Real-time example, but can be adapted (use in Hive UDF etc.)
Resource Location
Git repos

Steam Scoring Server from sunuq.icu

Characteristic Value
Pattern name Steam
Scoring data source REST API client
Scoring environment Steam scoring server
Scoring engine H2O POJO
Scoring latency SLA Real-time
Resource Location
Web sites

Scoring Server on AWS

you can deploy a restful server on aws using the marketplace ami (h2o inference server - hourly). notice that this is a paid ami.

transfer the mojo file into the /tmp folder of this instance before launching. if your mojo is in s3, assign a role that provides s3 access to the instance.

Run following bash script as “userdata” to transfer your MOJO into the instance before you launch the instance. Be sure you change the mojofile path below.

#cloud-boothook
#!/bin/bash
export mojofile="s3://yourbucket/yourmojo.zip"
aws s3 cp $mojofile /tmp/pipeline.mojo

After this instance has launched, you can make real time inference using the following command. Remember to change the IP address. Input data is provided through the row宁夏微乐滑水麻将下载 parameter in the URL.

curl "http://<yourIP>:8080/model?type=1&row=2000,2000"

Additional Resources