layout: global title: Spark MLContext Programming Guide description: Spark MLContext Programming Guide
- This will become a table of contents (this text will be scraped). {:toc}
Overview
The Spark MLContext API offers a programmatic interface for interacting with SystemML from Spark using languages such as Scala and Java. When interacting with MLContext from Spark, DataFrames and RDDs can be passed to SystemML. These data representations are converted to a binary-block data format, allowing for SystemML's optimizations to be performed.
Spark Shell (Scala) Example
Start Spark Shell with SystemML
To use SystemML with the Spark Shell, the SystemML jar can be referenced using the Spark Shell's --jars option. Instructions to build the SystemML jar can be found in the SystemML GitHub README.
{% highlight bash %} ./bin/spark-shell --executor-memory 4G --driver-memory 4G --jars SystemML.jar {% endhighlight %}
Here is an example of Spark Shell with SystemML and YARN.
{% highlight bash %} ./bin/spark-shell --master yarn-client --num-executors 3 --driver-memory 5G --executor-memory 5G --executor-cores 4 --jars SystemML.jar {% endhighlight %}
Create MLContext
An MLContext object can be created by passing its constructor a reference to the SparkContext.
scala> val ml = new MLContext(sc) ml: org.apache.sysml.api.MLContext = org.apache.sysml.api.MLContext@33e38c6b {% endhighlight %}
Create DataFrame
For demonstration purposes, we'll create a DataFrame consisting of 100,000 rows and 1,000 columns of random doubles.
scala> import org.apache.spark.sql.types.{StructType,StructField,DoubleType} import org.apache.spark.sql.types.{StructType, StructField, DoubleType}
scala> import scala.util.Random import scala.util.Random
scala> val numRows = 100000 numRows: Int = 100000
scala> val numCols = 1000 numCols: Int = 1000
scala> val data = sc.parallelize(0 to numRows-1).map { _ => Row.fromSeq(Seq.fill(numCols)(Random.nextDouble)) } data: org.apache.spark.rdd.RDD[org.apache.spark.sql.Row] = MapPartitionsRDD[1] at map at :33
scala> val schema = StructType((0 to numCols-1).map { i => StructField(“C” + i, DoubleType, true) } ) schema: org.apache.spark.sql.types.StructType = StructType(StructField(C0,DoubleType,true), StructField(C1,DoubleType,true), StructField(C2,DoubleType,true), StructField(C3,DoubleType,true), StructField(C4,DoubleType,true), StructField(C5,DoubleType,true), StructField(C6,DoubleType,true), StructField(C7,DoubleType,true), StructField(C8,DoubleType,true), StructField(C9,DoubleType,true), StructField(C10,DoubleType,true), StructField(C11,DoubleType,true), StructField(C12,DoubleType,true), StructField(C13,DoubleType,true), StructField(C14,DoubleType,true), StructField(C15,DoubleType,true), StructField(C16,DoubleType,true), StructField(C17,DoubleType,true), StructField(C18,DoubleType,true), StructField(C19,DoubleType,true), StructField(C20,DoubleType,true), StructField(C21,DoubleType,true), ...
scala> val df = sqlContext.createDataFrame(data, schema) df: org.apache.spark.sql.DataFrame = [C0: double, C1: double, C2: double, C3: double, C4: double, C5: double, C6: double, C7: double, C8: double, C9: double, C10: double, C11: double, C12: double, C13: double, C14: double, C15: double, C16: double, C17: double, C18: double, C19: double, C20: double, C21: double, C22: double, C23: double, C24: double, C25: double, C26: double, C27: double, C28: double, C29: double, C30: double, C31: double, C32: double, C33: double, C34: double, C35: double, C36: double, C37: double, C38: double, C39: double, C40: double, C41: double, C42: double, C43: double, C44: double, C45: double, C46: double, C47: double, C48: double, C49: double, C50: double, C51: double, C52: double, C53: double, C54: double, C55: double, C56: double, C57: double, C58: double, C5...
{% endhighlight %}
Helper Methods
For convenience, we'll create some helper methods. The SystemML output data is encapsulated in an MLOutput object. The getScalar() method extracts a scalar value from a DataFrame returned by MLOutput. The getScalarDouble() method returns such a value as a Double, and the getScalarInt() method returns such a value as an Int.
scala> def getScalar(outputs: MLOutput, symbol: String): Any = | outputs.getDF(sqlContext, symbol).first()(1) getScalar: (outputs: org.apache.sysml.api.MLOutput, symbol: String)Any
scala> def getScalarDouble(outputs: MLOutput, symbol: String): Double = | getScalar(outputs, symbol).asInstanceOf[Double] getScalarDouble: (outputs: org.apache.sysml.api.MLOutput, symbol: String)Double
scala> def getScalarInt(outputs: MLOutput, symbol: String): Int = | getScalarDouble(outputs, symbol).toInt getScalarInt: (outputs: org.apache.sysml.api.MLOutput, symbol: String)Int
{% endhighlight %}
{% endhighlight %}
Convert DataFrame to Binary-Block Matrix
SystemML is optimized to operate on a binary-block format for matrix representation. For large datasets, conversion from DataFrame to binary-block can require a significant quantity of time. Explicit DataFrame to binary-block conversion allows algorithm performance to be measured separately from data conversion time.
The SystemML binary-block matrix representation can be thought of as a two-dimensional array of blocks, where each block consists of a number of rows and columns. In this example, we specify a matrix consisting of blocks of size 1000x1000. The experimental dataFrameToBinaryBlock() method of RDDConverterUtilsExt is used to convert the DataFrame df to a SystemML binary-block matrix, which is represented by the datatype JavaPairRDD[MatrixIndexes, MatrixBlock].
scala> import org.apache.sysml.runtime.matrix.MatrixCharacteristics; import org.apache.sysml.runtime.matrix.MatrixCharacteristics
scala> val numRowsPerBlock = 1000 numRowsPerBlock: Int = 1000
scala> val numColsPerBlock = 1000 numColsPerBlock: Int = 1000
scala> val mc = new MatrixCharacteristics(numRows, numCols, numRowsPerBlock, numColsPerBlock) mc: org.apache.sysml.runtime.matrix.MatrixCharacteristics = [100000 x 1000, nnz=-1, blocks (1000 x 1000)]
scala> val sysMlMatrix = RDDConverterUtils.dataFrameToBinaryBlock(sc, df, mc, false) sysMlMatrix: org.apache.spark.api.java.JavaPairRDD[org.apache.sysml.runtime.matrix.data.MatrixIndexes,org.apache.sysml.runtime.matrix.data.MatrixBlock] = org.apache.spark.api.java.JavaPairRDD@2bce3248
{% endhighlight %}
{% endhighlight %}
DML Script
For this example, we will utilize the following DML Script called shape.dml that reads in a matrix and outputs the number of rows and the number of columns, each represented as a matrix.
{% highlight r %} X = read($Xin) m = matrix(nrow(X), rows=1, cols=1) n = matrix(ncol(X), rows=1, cols=1) write(m, $Mout) write(n, $Nout) {% endhighlight %}
Execute Script
Let's execute our DML script, as shown in the example below. The call to reset() of MLContext is not necessary here, but this method should be called if you need to reset inputs and outputs or if you would like to call execute() with a different script.
An example of registering the DataFrame df as an input to the X variable is shown but commented out. If a DataFrame is registered directly, it will implicitly be converted to SystemML‘s binary-block format. However, since we’ve already explicitly converted the DataFrame to the binary-block fixed variable systemMlMatrix, we will register this input to the X variable. We register the m and n variables as outputs.
When SystemML is executed via DMLScript (such as in Standalone Mode), inputs are supplied as either command-line named arguments or positional argument. These inputs are specified in DML scripts by prepending them with a $. Values are read from or written to files using read/write (DML) and load/save (PyDML) statements. When utilizing the MLContext API, inputs and outputs can be other data representations, such as DataFrames. The input and output data are bound to DML variables. The named arguments in the shape.dml script do not have default values set for them, so we create a Map to map the required named arguments to blank Strings so that the script can pass validation.
The shape.dml script is executed by the call to execute(), where we supply the Map of required named arguments. The execution results are returned as the MLOutput fixed variable outputs. The number of rows is obtained by calling the getStaticInt() helper method with the outputs object and "m". The number of columns is retrieved by calling getStaticInt() with outputs and "n".
scala> //ml.registerInput(“X”, df) // implicit conversion of DataFrame to binary-block
scala> ml.registerInput(“X”, sysMlMatrix, numRows, numCols)
scala> ml.registerOutput(“m”)
scala> ml.registerOutput(“n”)
scala> val nargs = Map(“Xin” -> " ", “Mout” -> " ", “Nout” -> " ") nargs: scala.collection.immutable.Map[String,String] = Map(Xin -> " ", Mout -> " ", Nout -> " ")
scala> val outputs = ml.execute(“shape.dml”, nargs) 15/10/12 16:29:15 WARN : Your hostname, derons-mbp.usca.ibm.com resolves to a loopback/non-reachable address: 127.0.0.1, but we couldn't find any external IP address! 15/10/12 16:29:15 WARN OptimizerUtils: Auto-disable multi-threaded text read for ‘text’ and ‘csv’ due to thread contention on JRE < 1.8 (java.version=1.7.0_80). outputs: org.apache.sysml.api.MLOutput = org.apache.sysml.api.MLOutput@4d424743
scala> val m = getScalarInt(outputs, “m”) m: Int = 100000
scala> val n = getScalarInt(outputs, “n”) n: Int = 1000
{% endhighlight %}
{% endhighlight %}
DML Script as String
The MLContext API allows a DML script to be specified as a String. Here, we specify a DML script as a fixed String variable called minMaxMeanScript. This DML will find the minimum, maximum, and mean value of a matrix.
{% endhighlight %}
{% endhighlight %}
Scala Wrapper for DML
We can create a Scala wrapper for our invocation of the minMaxMeanScript DML String. The minMaxMean() method takes a JavaPairRDD[MatrixIndexes, MatrixBlock] parameter, which is a SystemML binary-block matrix representation. It also takes a rows parameter indicating the number of rows in the matrix, a cols parameter indicating the number of columns in the matrix, and an MLContext parameter. The minMaxMean() method returns a tuple consisting of the minimum value in the matrix, the maximum value in the matrix, and the computed mean value of the matrix.
scala> import org.apache.sysml.runtime.matrix.data.MatrixBlock import org.apache.sysml.runtime.matrix.data.MatrixBlock
scala> import org.apache.spark.api.java.JavaPairRDD import org.apache.spark.api.java.JavaPairRDD
scala> def minMaxMean(mat: JavaPairRDD[MatrixIndexes, MatrixBlock], rows: Int, cols: Int, ml: MLContext): (Double, Double, Double) = { | ml.reset() | ml.registerInput(“Xin”, mat, rows, cols) | ml.registerOutput(“minOut”) | ml.registerOutput(“maxOut”) | ml.registerOutput(“meanOut”) | val outputs = ml.executeScript(minMaxMeanScript) | val minOut = getScalarDouble(outputs, “minOut”) | val maxOut = getScalarDouble(outputs, “maxOut”) | val meanOut = getScalarDouble(outputs, “meanOut”) | (minOut, maxOut, meanOut) | } minMaxMean: (mat: org.apache.spark.api.java.JavaPairRDD[org.apache.sysml.runtime.matrix.data.MatrixIndexes,org.apache.sysml.runtime.matrix.data.MatrixBlock], rows: Int, cols: Int, ml: org.apache.sysml.api.MLContext)(Double, Double, Double)
{% endhighlight %}
{% endhighlight %}
Invoking DML via Scala Wrapper
Here, we invoke minMaxMeanScript using our minMaxMean() Scala wrapper method. It returns a tuple consisting of the minimum value in the matrix, the maximum value in the matrix, and the mean value of the matrix.
{% endhighlight %}
{% endhighlight %}
Java Example
Next, let's consider a Java example. The MLContextExample class creates an MLContext object from a JavaSparkContext. Next, it reads in a matrix CSV file as a JavaRDD<String> object. It registers this as input X. It registers two outputs, m and n. A HashMap maps the expected command-line arguments of the shape.dml script to spaces so that it passes validation. The shape.dml script is executed, and the number of rows and columns in the matrix are output to standard output.
{% highlight java %} package org.apache.sysml;
import java.util.HashMap;
import org.apache.spark.SparkConf; import org.apache.spark.api.java.JavaRDD; import org.apache.spark.api.java.JavaSparkContext; import org.apache.spark.sql.DataFrame; import org.apache.spark.sql.SQLContext;
import org.apache.sysml.api.MLContext; import org.apache.sysml.api.MLOutput;
public class MLContextExample {
public static void main(String[] args) throws Exception {
SparkConf conf = new SparkConf().setAppName("MLContextExample").setMaster("local");
JavaSparkContext sc = new JavaSparkContext(conf);
SQLContext sqlContext = new SQLContext(sc);
MLContext ml = new MLContext(sc);
JavaRDD<String> csv = sc.textFile("A.csv");
ml.registerInput("X", csv, "csv");
ml.registerOutput("m");
ml.registerOutput("n");
HashMap<String, String> cmdLineArgs = new HashMap<String, String>();
cmdLineArgs.put("X", " ");
cmdLineArgs.put("m", " ");
cmdLineArgs.put("n", " ");
MLOutput output = ml.execute("shape.dml", cmdLineArgs);
DataFrame mDf = output.getDF(sqlContext, "m");
DataFrame nDf = output.getDF(sqlContext, "n");
System.out.println("rows:" + mDf.first().getDouble(1));
System.out.println("cols:" + nDf.first().getDouble(1));
}
}
{% endhighlight %}
Zeppelin Notebook Example - Linear Regression Algorithm
Next, we'll consider an example of a SystemML linear regression algorithm run from Spark through an Apache Zeppelin notebook. Instructions to clone and build Zeppelin can be found at the GitHub Apache Zeppelin site. This example also will look at the Spark ML linear regression algorithm.
This Zeppelin notebook example can be imported by choosing Import note -> Add from URL from the Zeppelin main page, then insert the following URL:
https://raw.githubusercontent.com/apache/incubator-systemml/master/samples/zeppelin-notebooks/2AZ2AQ12B/note.json
Alternatively download note.json, then import it by choosing Import note -> Choose a JSON here from the Zeppelin main page.
A conf/zeppelin-env.sh file is created based on conf/zeppelin-env.sh.template. For this demonstration, it features SPARK_HOME, SPARK_SUBMIT_OPTIONS, and ZEPPELIN_SPARK_USEHIVECONTEXT environment variables:
export SPARK_HOME=/Users/example/spark-1.5.1-bin-hadoop2.6 export SPARK_SUBMIT_OPTIONS="--jars /Users/example/systemml/system-ml/target/SystemML.jar" export ZEPPELIN_SPARK_USEHIVECONTEXT=false
Start Zeppelin using the zeppelin.sh script:
bin/zeppelin.sh
After opening Zeppelin in a brower, we see the “SystemML - Linear Regression” note in the list of available Zeppelin notes.
If we go to the “SystemML - Linear Regression” note, we see that the note consists of several cells of code.
Let's briefly consider these cells.
Trigger Spark Startup
This cell triggers Spark to initialize by calling the SparkContext sc object. Information regarding these startup operations can be viewed in the console window in which zeppelin.sh is running.
Cell: {% highlight scala %} // Trigger Spark Startup sc {% endhighlight %}
Output: {% highlight scala %} res8: org.apache.spark.SparkContext = org.apache.spark.SparkContext@6ce70bf3 {% endhighlight %}
Generate Linear Regression Test Data
The Spark LinearDataGenerator is used to generate test data for the Spark ML and SystemML linear regression algorithms.
Cell: {% highlight scala %} // Generate data import org.apache.spark.mllib.util.LinearDataGenerator
val numRows = 10000 val numCols = 1000 val rawData = LinearDataGenerator.generateLinearRDD(sc, numRows, numCols, 1).toDF()
// Repartition into a more parallelism-friendly number of partitions val data = rawData.repartition(64).cache() {% endhighlight %}
Output: {% highlight scala %} import org.apache.spark.mllib.util.LinearDataGenerator numRows: Int = 10000 numCols: Int = 1000 rawData: org.apache.spark.sql.DataFrame = [label: double, features: vector] data: org.apache.spark.sql.DataFrame = [label: double, features: vector] {% endhighlight %}
Train using Spark ML Linear Regression Algorithm for Comparison
For purpose of comparison, we can train a model using the Spark ML linear regression algorithm.
Cell: {% highlight scala %} // Spark ML import org.apache.spark.ml.regression.LinearRegression
// Model Settings val maxIters = 100 val reg = 0 val elasticNetParam = 0 // L2 reg
// Fit the model val lr = new LinearRegression() .setMaxIter(maxIters) .setRegParam(reg) .setElasticNetParam(elasticNetParam) val start = System.currentTimeMillis() val model = lr.fit(data) val trainingTime = (System.currentTimeMillis() - start).toDouble / 1000.0
// Summarize the model over the training set and gather some metrics val trainingSummary = model.summary val r2 = trainingSummary.r2 val iters = trainingSummary.totalIterations val trainingTimePerIter = trainingTime / iters {% endhighlight %}
Output: {% highlight scala %} import org.apache.spark.ml.regression.LinearRegression maxIters: Int = 100 reg: Int = 0 elasticNetParam: Int = 0 lr: org.apache.spark.ml.regression.LinearRegression = linReg_a7f51d676562 start: Long = 1444672044647 model: org.apache.spark.ml.regression.LinearRegressionModel = linReg_a7f51d676562 trainingTime: Double = 12.985 trainingSummary: org.apache.spark.ml.regression.LinearRegressionTrainingSummary = org.apache.spark.ml.regression.LinearRegressionTrainingSummary@227ba28b r2: Double = 0.9677118209276552 iters: Int = 17 trainingTimePerIter: Double = 0.7638235294117647 {% endhighlight %}
Spark ML Linear Regression Summary Statistics
Summary statistics for the Spark ML linear regression algorithm are displayed by this cell.
Cell: {% highlight scala %} // Print statistics println(s“R2: ${r2}”) println(s“Iterations: ${iters}”) println(s“Training time per iter: ${trainingTimePerIter} seconds”) {% endhighlight %}
Output: {% highlight scala %} R2: 0.9677118209276552 Iterations: 17 Training time per iter: 0.7638235294117647 seconds {% endhighlight %}
SystemML Linear Regression Algorithm
The linearReg fixed String variable is set to a linear regression algorithm written in DML, SystemML's Declarative Machine Learning language.
Cell: {% highlight scala %} // SystemML kernels val linearReg = """
THIS SCRIPT SOLVES LINEAR REGRESSION USING THE CONJUGATE GRADIENT ALGORITHM
INPUT PARAMETERS:
--------------------------------------------------------------------------------------------
NAME TYPE DEFAULT MEANING
--------------------------------------------------------------------------------------------
X String --- Matrix X of feature vectors
Y String --- 1-column Matrix Y of response values
icpt Int 0 Intercept presence, shifting and rescaling the columns of X:
0 = no intercept, no shifting, no rescaling;
1 = add intercept, but neither shift nor rescale X;
2 = add intercept, shift & rescale X columns to mean = 0, variance = 1
reg Double 0.000001 Regularization constant (lambda) for L2-regularization; set to nonzero
for highly dependend/sparse/numerous features
tol Double 0.000001 Tolerance (epsilon); conjugate graduent procedure terminates early if
L2 norm of the beta-residual is less than tolerance * its initial norm
maxi Int 0 Maximum number of conjugate gradient iterations, 0 = no maximum
--------------------------------------------------------------------------------------------
OUTPUT:
B Estimated regression parameters (the betas) to store
Note: Matrix of regression parameters (the betas) and its size depend on icpt input value:
OUTPUT SIZE: OUTPUT CONTENTS: HOW TO PREDICT Y FROM X AND B:
icpt=0: ncol(X) x 1 Betas for X only Y ~ X %% B[1:ncol(X), 1], or just X %% B
icpt=1: ncol(X)+1 x 1 Betas for X and intercept Y ~ X %*% B[1:ncol(X), 1] + B[ncol(X)+1, 1]
icpt=2: ncol(X)+1 x 2 Col.1: betas for X & intercept Y ~ X %*% B[1:ncol(X), 1] + B[ncol(X)+1, 1]
Col.2: betas for shifted/rescaled X and intercept
fileX = ""; fileY = ""; fileB = "";
intercept_status = ifdef ($icpt, 0); # $icpt=0; tolerance = ifdef ($tol, 0.000001); # $tol=0.000001; max_iteration = ifdef ($maxi, 0); # $maxi=0; regularization = ifdef ($reg, 0.000001); # $reg=0.000001;
X = read (fileX); y = read (fileY);
n = nrow (X); m = ncol (X); ones_n = matrix (1, rows = n, cols = 1); zero_cell = matrix (0, rows = 1, cols = 1);
Introduce the intercept, shift and rescale the columns of X if needed
m_ext = m; if (intercept_status == 1 | intercept_status == 2) # add the intercept column { X = append (X, ones_n); m_ext = ncol (X); }
scale_lambda = matrix (1, rows = m_ext, cols = 1); if (intercept_status == 1 | intercept_status == 2) { scale_lambda [m_ext, 1] = 0; }
if (intercept_status == 2) # scale-&-shift X columns to mean 0, variance 1 { # Important assumption: X [, m_ext] = ones_n avg_X_cols = t(colSums(X)) / n; var_X_cols = (t(colSums (X ^ 2)) - n * (avg_X_cols ^ 2)) / (n - 1); is_unsafe = ppred (var_X_cols, 0.0, “<=”); scale_X = 1.0 / sqrt (var_X_cols * (1 - is_unsafe) + is_unsafe); scale_X [m_ext, 1] = 1; shift_X = - avg_X_cols * scale_X; shift_X [m_ext, 1] = 0; } else { scale_X = matrix (1, rows = m_ext, cols = 1); shift_X = matrix (0, rows = m_ext, cols = 1); }
Henceforth, if intercept_status == 2, we use “X %*% (SHIFT/SCALE TRANSFORM)”
instead of “X”. However, in order to preserve the sparsity of X,
we apply the transform associatively to some other part of the expression
in which it occurs. To avoid materializing a large matrix, we rewrite it:
ssX_A = (SHIFT/SCALE TRANSFORM) %*% A --- is rewritten as:
ssX_A = diag (scale_X) %*% A;
ssX_A [m_ext, ] = ssX_A [m_ext, ] + t(shift_X) %*% A;
tssX_A = t(SHIFT/SCALE TRANSFORM) %*% A --- is rewritten as:
tssX_A = diag (scale_X) %% A + shift_X %% A [m_ext, ];
lambda = scale_lambda * regularization; beta_unscaled = matrix (0, rows = m_ext, cols = 1);
if (max_iteration == 0) { max_iteration = m_ext; } i = 0;
BEGIN THE CONJUGATE GRADIENT ALGORITHM
r = - t(X) %*% y;
if (intercept_status == 2) { r = scale_X * r + shift_X %*% r [m_ext, ]; }
p = - r; norm_r2 = sum (r ^ 2); norm_r2_initial = norm_r2; norm_r2_target = norm_r2_initial * tolerance ^ 2;
while (i < max_iteration & norm_r2 > norm_r2_target) { if (intercept_status == 2) { ssX_p = scale_X * p; ssX_p [m_ext, ] = ssX_p [m_ext, ] + t(shift_X) %*% p; } else { ssX_p = p; }
q = t(X) %*% (X %*% ssX_p);
if (intercept_status == 2) {
q = scale_X * q + shift_X %*% q [m_ext, ];
}
q = q + lambda * p;
a = norm_r2 / sum (p * q);
beta_unscaled = beta_unscaled + a * p;
r = r + a * q;
old_norm_r2 = norm_r2;
norm_r2 = sum (r ^ 2);
p = -r + (norm_r2 / old_norm_r2) * p;
i = i + 1;
}
END THE CONJUGATE GRADIENT ALGORITHM
if (intercept_status == 2) { beta = scale_X * beta_unscaled; beta [m_ext, ] = beta [m_ext, ] + t(shift_X) %*% beta_unscaled; } else { beta = beta_unscaled; }
Output statistics
avg_tot = sum (y) / n; ss_tot = sum (y ^ 2); ss_avg_tot = ss_tot - n * avg_tot ^ 2; var_tot = ss_avg_tot / (n - 1); y_residual = y - X %*% beta; avg_res = sum (y_residual) / n; ss_res = sum (y_residual ^ 2); ss_avg_res = ss_res - n * avg_res ^ 2;
R2_temp = 1 - ss_res / ss_avg_tot R2 = matrix(R2_temp, rows=1, cols=1) write(R2, "")
totalIters = matrix(i, rows=1, cols=1) write(totalIters, "")
Prepare the output matrix
if (intercept_status == 2) { beta_out = append (beta, beta_unscaled); } else { beta_out = beta; }
write (beta_out, fileB); """ {% endhighlight %}
Output:
None
Helper Methods
This cell contains helper methods to return Double and Int values from output generated by the MLContext API.
Cell: {% highlight scala %} // Helper functions import org.apache.sysml.api.MLOutput
def getScalar(outputs: MLOutput, symbol: String): Any = outputs.getDF(sqlContext, symbol).first()(1)
def getScalarDouble(outputs: MLOutput, symbol: String): Double = getScalar(outputs, symbol).asInstanceOf[Double]
def getScalarInt(outputs: MLOutput, symbol: String): Int = getScalarDouble(outputs, symbol).toInt {% endhighlight %}
Output: {% highlight scala %} import org.apache.sysml.api.MLOutput getScalar: (outputs: org.apache.sysml.api.MLOutput, symbol: String)Any getScalarDouble: (outputs: org.apache.sysml.api.MLOutput, symbol: String)Double getScalarInt: (outputs: org.apache.sysml.api.MLOutput, symbol: String)Int {% endhighlight %}
Convert DataFrame to Binary-Block Format
SystemML uses a binary-block format for matrix data representation. This cell explicitly converts the DataFrame data object to a binary-block features matrix and single-column label matrix, both represented by the JavaPairRDD[MatrixIndexes, MatrixBlock] datatype.
Cell: {% highlight scala %} // Imports import org.apache.sysml.api.MLContext import org.apache.sysml.runtime.instructions.spark.utils.{RDDConverterUtilsExt => RDDConverterUtils} import org.apache.sysml.runtime.matrix.MatrixCharacteristics;
// Create SystemML context val ml = new MLContext(sc)
// Convert data to proper format val mcX = new MatrixCharacteristics(numRows, numCols, 1000, 1000) val mcY = new MatrixCharacteristics(numRows, 1, 1000, 1000) val X = RDDConverterUtils.vectorDataFrameToBinaryBlock(sc, data, mcX, false, “features”) val y = RDDConverterUtils.dataFrameToBinaryBlock(sc, data.select(“label”), mcY, false) // val y = data.select(“label”)
// Cache val X2 = X.cache() val y2 = y.cache() val cnt1 = X2.count() val cnt2 = y2.count() {% endhighlight %}
Output: {% highlight scala %} import org.apache.sysml.api.MLContext import org.apache.sysml.runtime.instructions.spark.utils.{RDDConverterUtilsExt=>RDDConverterUtils} import org.apache.sysml.runtime.matrix.MatrixCharacteristics ml: org.apache.sysml.api.MLContext = org.apache.sysml.api.MLContext@38d59245 mcX: org.apache.sysml.runtime.matrix.MatrixCharacteristics = [10000 x 1000, nnz=-1, blocks (1000 x 1000)] mcY: org.apache.sysml.runtime.matrix.MatrixCharacteristics = [10000 x 1, nnz=-1, blocks (1000 x 1000)] X: org.apache.spark.api.java.JavaPairRDD[org.apache.sysml.runtime.matrix.data.MatrixIndexes,org.apache.sysml.runtime.matrix.data.MatrixBlock] = org.apache.spark.api.java.JavaPairRDD@b5a86e3 y: org.apache.spark.api.java.JavaPairRDD[org.apache.sysml.runtime.matrix.data.MatrixIndexes,org.apache.sysml.runtime.matrix.data.MatrixBlock] = org.apache.spark.api.java.JavaPairRDD@56377665 X2: org.apache.spark.api.java.JavaPairRDD[org.apache.sysml.runtime.matrix.data.MatrixIndexes,org.apache.sysml.runtime.matrix.data.MatrixBlock] = org.apache.spark.api.java.JavaPairRDD@650f29d2 y2: org.apache.spark.api.java.JavaPairRDD[org.apache.sysml.runtime.matrix.data.MatrixIndexes,org.apache.sysml.runtime.matrix.data.MatrixBlock] = org.apache.spark.api.java.JavaPairRDD@334857a8 cnt1: Long = 10 cnt2: Long = 10 {% endhighlight %}
Train using SystemML Linear Regression Algorithm
Now, we can train our model using the SystemML linear regression algorithm. We register the features matrix X and the label matrix y as inputs. We register the beta_out matrix, R2, and totalIters as outputs.
Cell: {% highlight scala %} // Register inputs & outputs ml.reset()
ml.registerInput(“X”, X, numRows, numCols) ml.registerInput(“y”, y, numRows, 1) // ml.registerInput(“y”, y) ml.registerOutput(“beta_out”) ml.registerOutput(“R2”) ml.registerOutput(“totalIters”)
// Run the script val start = System.currentTimeMillis() val outputs = ml.executeScript(linearReg) val trainingTime = (System.currentTimeMillis() - start).toDouble / 1000.0
// Get outputs val B = outputs.getDF(sqlContext, “beta_out”).sort(“ID”).drop(“ID”) val r2 = getScalarDouble(outputs, “R2”) val iters = getScalarInt(outputs, “totalIters”) val trainingTimePerIter = trainingTime / iters {% endhighlight %}
Output: {% highlight scala %} start: Long = 1444672090620 outputs: org.apache.sysml.api.MLOutput = org.apache.sysml.api.MLOutput@5d2c22d0 trainingTime: Double = 1.176 B: org.apache.spark.sql.DataFrame = [C1: double] r2: Double = 0.9677079547216473 iters: Int = 12 trainingTimePerIter: Double = 0.09799999999999999 {% endhighlight %}
SystemML Linear Regression Summary Statistics
SystemML linear regression summary statistics are displayed by this cell.
Cell: {% highlight scala %} // Print statistics println(s“R2: ${r2}”) println(s“Iterations: ${iters}”) println(s“Training time per iter: ${trainingTimePerIter} seconds”) B.describe().show() {% endhighlight %}
Output: {% highlight scala %} R2: 0.9677079547216473 Iterations: 12 Training time per iter: 0.2334166666666667 seconds +-------+-------------------+ |summary| C1| +-------+-------------------+ | count| 1000| | mean| 0.0184500840658385| | stddev| 0.2764750319432085| | min|-0.5426068958986378| | max| 0.5225309861616542| +-------+-------------------+ {% endhighlight %}
Jupyter (PySpark) Notebook Example - Poisson Nonnegative Matrix Factorization
Here, we'll explore the use of SystemML via PySpark in a Jupyter notebook. This Jupyter notebook example can be nicely viewed in a rendered state on GitHub, and can be downloaded here to a directory of your choice.
From the directory with the downloaded notebook, start Jupyter with PySpark:
PYSPARK_DRIVER_PYTHON=jupyter PYSPARK_DRIVER_PYTHON_OPTS="notebook" $SPARK_HOME/bin/pyspark --master local[*] --jars $SYSTEMML_HOME/target/SystemML.jar --py-files $SYSTEMML_HOME/src/main/java/org/apache/sysml/api/python/SystemML.py --driver-class-path $SYSTEMML_HOME/target/SystemML.jar
This will open Jupyter in a browser:
We can then open up the SystemML-PySpark-Recommendation-Demo notebook:
Set up the notebook and download the data
{% highlight python %} %load_ext autoreload %autoreload 2 %matplotlib inline
import numpy as np import matplotlib.pyplot as plt plt.rcParams[‘figure.figsize’] = (10, 6) {% endhighlight %}
{% highlight python %} %%sh
Download dataset
curl -O http://snap.stanford.edu/data/amazon0601.txt.gz gunzip amazon0601.txt.gz {% endhighlight %}
Use PySpark to load the data in as a Spark DataFrame
{% highlight python %}
Load data
import pyspark.sql.functions as F dataPath = “amazon0601.txt”
X_train = (sc.textFile(dataPath) .filter(lambda l: not l.startswith(“#”)) .map(lambda l: l.split(“\t”)) .map(lambda prods: (int(prods[0]), int(prods[1]), 1.0)) .toDF((“prod_i”, “prod_j”, “x_ij”)) .filter(“prod_i < 500 AND prod_j < 500”) # Filter for memory constraints .cache())
max_prod_i = X_train.select(F.max(“prod_i”)).first()[0] max_prod_j = X_train.select(F.max(“prod_j”)).first()[0] numProducts = max(max_prod_i, max_prod_j) + 1 # 0-based indexing print(“Total number of products: {}”.format(numProducts)) {% endhighlight %}
Create a SystemML MLContext object
{% highlight python %}
Create SystemML MLContext
from SystemML import MLContext ml = MLContext(sc) {% endhighlight %}
Define a kernel for Poisson nonnegative matrix factorization (PNMF) in DML
{% highlight python %}
Define PNMF kernel in SystemML's DSL using the R-like syntax for PNMF
pnmf = """
data & args
X = read($X) X = X+1 # change product IDs to be 1-based, rather than 0-based V = table(X[,1], X[,2]) size = ifdef($size, -1) if(size > -1) { V = V[1:size,1:size] } max_iteration = as.integer($maxiter) rank = as.integer($rank)
n = nrow(V) m = ncol(V) range = 0.01 W = Rand(rows=n, cols=rank, min=0, max=range, pdf=“uniform”) H = Rand(rows=rank, cols=m, min=0, max=range, pdf=“uniform”) losses = matrix(0, rows=max_iteration, cols=1)
run PNMF
i=1 while(i <= max_iteration) {
update params
H = (H * (t(W) %% (V/(W%%H))))/t(colSums(W)) W = (W * ((V/(W%%H)) %% t(H)))/t(rowSums(H))
compute loss
losses[i,] = -1 * (sum(Vlog(W%%H)) - as.scalar(colSums(W)%*%rowSums(H))) i = i + 1; }
write outputs
write(losses, $lossout) write(W, $Wout) write(H, $Hout) """ {% endhighlight %}
Execute the algorithm
{% highlight python %}
Run the PNMF script on SystemML with Spark
ml.reset() outputs = ml.executeScript(pnmf, {“X”: X_train, “maxiter”: 100, “rank”: 10}, [“W”, “H”, “losses”]) {% endhighlight %}
Retrieve the losses during training and plot them
{% highlight python %}
Plot training loss over time
losses = outputs.getDF(sqlContext, “losses”) xy = losses.sort(losses.ID).map(lambda r: (r[0], r[1])).collect() x, y = zip(*xy) plt.plot(x, y) plt.xlabel(‘Iteration’) plt.ylabel(‘Loss’) plt.title(‘PNMF Training Loss’) {% endhighlight %}
