SystemML is a flexible, scalable machine learning (ML) language written in Java. SystemML's distinguishing characteristics are: (1) algorithm customizability, (2) multiple execution modes, including Standalone, Hadoop Batch, and Spark Batch, and (3) automatic optimization.
ML algorithms in SystemML are specified in a high-level, declarative machine learning (DML) language. Algorithms can be expressed in either an R-like syntax or a Python-like syntax. DML includes linear algebra primitives, statistical functions, and additional constructs.
This high-level language significantly increases the productivity of data scientists as it provides (1) full flexibility in expressing custom analytics and (2) data independence from the underlying input formats and physical data representations.
SystemML computations can be executed in a variety of different modes. To begin with, SystemML can be operated in Standalone mode on a single machine, allowing data scientists to develop algorithms locally without need of a distributed cluster. Algorithms can be distributed across Hadoop or Spark. This flexibility allows the utilization of an organization's existing resources and expertise. In addition, SystemML can be operated via Java, Scala, and Python. SystemML also features an embedded API for scoring models.
Algorithms specified in DML are dynamically compiled and optimized based on data and cluster characteristics using rule-based and cost-based optimization techniques. The optimizer automatically generates hybrid runtime execution plans ranging from in-memory single-node execution to distributed computations on Spark or Hadoop. This ensures both efficiency and scalability. Automatic optimization reduces or eliminates the need to hand-tune distributed runtime execution plans and system configurations.
SystemML is built using Apache Maven. SystemML will build on Windows, Linux, or MacOS and requires Maven 3 and Java 7 (or higher). To build SystemML, run:
mvn clean package
SystemML features a comprehensive set of integration tests. To perform these tests, run:
cd system-ml
mvn verify
Note: these that these tests requires R to be installed and available as part of the PATH variable on the machine you are running these tests.
To install required packages for running integration tests, execute following command in R:
install.packages(c("batch", "bitops", "boot", "caTools", "data.table", "doMC", "doSNOW", "ggplot2", "glmnet", "lda", "Matrix", "matrixStats", "moments", "plotrix", "psych", "reshape", "topicmodels", "wordcloud", "methods"), dependencies=TRUE)
Known issue: package 'methods' is not available for R version 3.2.1. In which case, please downgrade R to version 3.1.1.
SystemML is distributed in several packages, including a standalone package. We'll operate in Standalone mode in this guide. If you built SystemML from source (mvn clean package) the standalone package got created under the system-ml/target folder. In order to follow the examples below, extract it to your working directory, i.e. ~/systemml-examples:
tar -xzf system-ml/target/system-ml*standalone.tar.gz -C ~/systemml-examples
cd ~/systemml-examples
The extracted package should have these contents:
algorithms/
docs/
lib/
log4j.properties
readme.txt
runStandaloneSystemML.bat*
runStandaloneSystemML.sh*
SystemML-config.xml
Standalone mode can be executed either on Mac/Unix using the runStandaloneSystemML.sh script or on Windows using the runStandaloneSystemML.bat batch file.
SystemML features a suite of algorithms that can be grouped into five broad categories: Descriptive Statistics, Classification, Clustering, Regression, and Matrix Factorization. Detailed descriptions of these algorithms can be found in the Algorithm Reference packaged with SystemML.
As an example of the capabilities and power of SystemML and DML, let's consider the Linear Regression algorithm. We require sets of data to train and test our model. To obtain this data, we can either use real data or generate data for our algorithm. The UCI Machine Learning Repository Datasets is one location for real data. Use of real data typically involves some degree of data wrangling. In the following example, we will use SystemML to generate random data to train and test our model.
This example consists of the following parts:
- Run DML Script to Generate Random Data
- Divide Generated Data into Two Sample Groups
- Split Label Column from First Sample
- Split Label Column from Second Sample
- Train Model on First Sample
- Test Model on Second Sample
SystemML is distributed in several packages, including a standalone package. We'll operate in Standalone mode in this example.
We can execute the genLinearRegressionData.dml script in Standalone mode using either the runStandaloneSystemML.sh or runStandaloneSystemML.bat file.
In this example, we'll generate a matrix of 1000 rows of 50 columns of test data, with sparsity 0.7. In addition to this, a 51st column consisting of labels will
be appended to the matrix.
./runStandaloneSystemML.sh algorithms/datagen/genLinearRegressionData.dml -nvargs numSamples=1000 numFeatures=50 maxFeatureValue=5 maxWeight=5 addNoise=FALSE b=0 sparsity=0.7 output=linRegData.csv format=csv perc=0.5
This generates the following files:
linRegData.csv # 1000 rows of 51 columns of doubles (50 data columns and 1 label column), csv format
linRegData.csv.mtd # metadata file
Next, we'll create two subsets of the generated data, each of size ~50%. We can accomplish this using the sample.dml script.
This script will randomly sample rows from the linRegData.csv file and place them into 2 files.
To do this, we need to create a csv file for the sv named argument (see sample.dml for more details),
which I called perc.csv. This file was generated in previous step and looks like:
0.5
0.5
This will create two sample groups of roughly 50 percent each.
Now, the sample.dml script can be run.
./runStandaloneSystemML.sh algorithms/utils/sample.dml -nvargs X=linRegData.csv sv=perc.csv O=linRegDataParts ofmt=csv
This script creates two partitions of the original data and places them in a linRegDataParts folder. The files created are
as follows:
linRegDataParts/1 # first partition of data, ~50% of rows of linRegData.csv, csv format
linRegDataParts/1.mtd # metadata
linRegDataParts/2 # second partition of data, ~50% of rows of linRegData.csv, csv format
linRegDataParts/2.mtd # metadata
The 1 file contains the first partition of data, and the 2 file contains the second partition of data.
An associated metadata file describes
the nature of each partition of data. If we open 1 and 2 and look at the number of rows, we can see that typically
the partitions are not exactly 50% but instead are close to 50%. However, we find that the total number of rows in the
original data file equals the sum of the number of rows in 1 and 2.
The next task is to split the label column from the first sample. We can do this using the splitXY.dml script.
./runStandaloneSystemML.sh algorithms/utils/splitXY.dml -nvargs X=linRegDataParts/1 y=51 OX=linRegData.train.data.csv OY=linRegData.train.labels.csv ofmt=csv
This splits column 51, the label column, off from the data. When done, the following files have been created.
linRegData.train.data.csv # training data of 50 columns, csv format
linRegData.train.data.csv.mtd # metadata
linRegData.train.labels.csv # training labels of 1 column, csv format
linRegData.train.labels.csv.mtd # metadata
We also need to split the label column from the second sample.
./runStandaloneSystemML.sh algorithms/utils/splitXY.dml -nvargs X=linRegDataParts/2 y=51 OX=linRegData.test.data.csv OY=linRegData.test.labels.csv ofmt=csv
This splits column 51 off the data, resulting in the following files:
linRegData.test.data.csv # test data of 50 columns, csv format
linRegData.test.data.csv.mtd # metadata
linRegData.test.labels.csv # test labels of 1 column, csv format
linRegData.test.labels.csv.mtd # metadata
Now, we can train our model based on the first sample. To do this, we utilize the LinearRegDS.dml (Linear Regression
Direct Solve) script. Note that SystemML also includes a LinearRegCG.dml (Linear Regression Conjugate Gradient) algorithm
for situations where the number of features is large.
./runStandaloneSystemML.sh algorithms/LinearRegDS.dml -nvargs X=linRegData.train.data.csv Y=linRegData.train.labels.csv B=betas.csv fmt=csv
This will generate the following files:
betas.csv # betas, 50 rows of 1 column, csv format
betas.csv.mtd # metadata
The LinearRegDS.dml script generates statistics to standard output similar to the following.
BEGIN LINEAR REGRESSION SCRIPT
Reading X and Y...
Calling the Direct Solver...
Computing the statistics...
AVG_TOT_Y,-2.160284487670675
STDEV_TOT_Y,66.86434576808432
AVG_RES_Y,-3.3127468704080085E-10
STDEV_RES_Y,1.7231785003947183E-8
DISPERSION,2.963950542926297E-16
PLAIN_R2,1.0
ADJUSTED_R2,1.0
PLAIN_R2_NOBIAS,1.0
ADJUSTED_R2_NOBIAS,1.0
PLAIN_R2_VS_0,1.0
ADJUSTED_R2_VS_0,1.0
Writing the output matrix...
END LINEAR REGRESSION SCRIPT
Now that we have our betas.csv, we can test our model with our second set of data.
To test our model on the second sample, we can use the GLM-predict.dml script. This script can be used for both
prediction and scoring. Here, we're using it for scoring since we include the Y named argument. Our betas.csv
file is specified as the B named argument.
./runStandaloneSystemML.sh algorithms/GLM-predict.dml -nvargs X=linRegData.test.data.csv Y=linRegData.test.labels.csv B=betas.csv fmt=csv
This generates the following statistics to standard output.
LOGLHOOD_Z,,FALSE,NaN
LOGLHOOD_Z_PVAL,,FALSE,NaN
PEARSON_X2,,FALSE,1.895530994504798E-13
PEARSON_X2_BY_DF,,FALSE,4.202951207327712E-16
PEARSON_X2_PVAL,,FALSE,1.0
DEVIANCE_G2,,FALSE,0.0
DEVIANCE_G2_BY_DF,,FALSE,0.0
DEVIANCE_G2_PVAL,,FALSE,1.0
LOGLHOOD_Z,,TRUE,NaN
LOGLHOOD_Z_PVAL,,TRUE,NaN
PEARSON_X2,,TRUE,1.895530994504798E-13
PEARSON_X2_BY_DF,,TRUE,4.202951207327712E-16
PEARSON_X2_PVAL,,TRUE,1.0
DEVIANCE_G2,,TRUE,0.0
DEVIANCE_G2_BY_DF,,TRUE,0.0
DEVIANCE_G2_PVAL,,TRUE,1.0
AVG_TOT_Y,1,,1.0069397725436522
STDEV_TOT_Y,1,,68.29092137526905
AVG_RES_Y,1,,-4.1450397073455047E-10
STDEV_RES_Y,1,,2.0519206226041048E-8
PRED_STDEV_RES,1,TRUE,1.0
PLAIN_R2,1,,1.0
ADJUSTED_R2,1,,1.0
PLAIN_R2_NOBIAS,1,,1.0
ADJUSTED_R2_NOBIAS,1,,1.0
We see that the STDEV_RES_Y value of the testing phase is of similar magnitude to the value obtained from the model training phase.
For convenience, we can encapsulate our DML invocations in a single script:
#!/bin/bash
./runStandaloneSystemML.sh algorithms/datagen/genLinearRegressionData.dml -nvargs numSamples=1000 numFeatures=50 maxFeatureValue=5 maxWeight=5 addNoise=FALSE b=0 sparsity=0.7 output=linRegData.csv format=csv perc=0.5
./runStandaloneSystemML.sh algorithms/utils/sample.dml -nvargs X=linRegData.csv sv=perc.csv O=linRegDataParts ofmt=csv
./runStandaloneSystemML.sh algorithms/utils/splitXY.dml -nvargs X=linRegDataParts/1 y=51 OX=linRegData.train.data.csv OY=linRegData.train.labels.csv ofmt=csv
./runStandaloneSystemML.sh algorithms/utils/splitXY.dml -nvargs X=linRegDataParts/2 y=51 OX=linRegData.test.data.csv OY=linRegData.test.labels.csv ofmt=csv
./runStandaloneSystemML.sh algorithms/LinearRegDS.dml -nvargs X=linRegData.train.data.csv Y=linRegData.train.labels.csv B=betas.csv fmt=csv
./runStandaloneSystemML.sh algorithms/GLM-predict.dml -nvargs X=linRegData.test.data.csv Y=linRegData.test.labels.csv B=betas.csv fmt=csv
In this example, we've seen a small part of the capabilities of SystemML. For more detailed information, please consult the SystemML Algorithm Reference and SystemML Language Reference.