USER MANUAL OF FELIX 0.2 - The Stanford University InfoLab
Contents
1. If this predicate with key constraint only appears in non recursive rules F will consider mapping it to a logistic regression operator Conditional Random Field F will assign a predicate to a conditional random field operator if a predicate has a key constraint same syntax as Logistic Regression and appears in non recursive rules or rules with linear chain recursion of this form wgt weight2 11 12 wgt label2 sid widi 11 wid2 wid1 1 gt label2 sid wid2 12 word seqid wordid word feature wordid feature weight label feature double seqTime seqid date tcrt word word dwinner word date dloser word date label seqid wordid label tcoref wordid wordid tcoref_map wordid wordid tcoref t1 t1 teoref ti t2 gt tcoref t2 ti tcoref ti t2 tcoref t2 t3 gt teoref el 3 20 word seqid1 wordid1 word1 word seqid2 wordid2 word2 tcrt wordi w tcrt word2 w gt tcoref wordidi wordid2 wgt word seq id word feature id feature weight label feature wgt gt label seq id label 5 label seq word WIN tcoref map word word1 word seqi wordi text1 seqTime seq date gt dwinner texti date 5 label seq word LOS tcoref map word word1 word seqi wordi text1 seqTime seq date dloser texti date 5 dwinner ti date gt dloser ti date Figure 1 Sample Input to F prog mln seqTime 48 Dec 12 2006 word 48 132. 0 min 2 484 sec gt gt gt Optimize DMOs for LR Operator of label seqidl wordid2 label3 with 3 Relevant Clauses Partitioned by label seqidl wordid2 label3 round seqidl 1 0 4 3 Uses 0 min 2 459 sec gt gt gt Optimize DMOs for LR Operator of label seqidl wordid2 label3 with 3 Relevant Clauses Partitioned by label seqidl wordid2 label3 round seqidl 1 0 4 0 Uses 0 min 2 232 sec Figure 4 Screencast of Felix 5 Input Options 5 1 Local Files This is the standard way to input data to Felix Supply MLN program files with i and evidence files with e See Figures 1 and 2 for examples 5 2 Relational Database Tables Another option for supplying evidence to F is through relational database tables To use this option first edit the predicate declaration in the MLN program file to specify the name of the relational database table word seqid wordid word db my schema my table In this example the predicate word is supplied via the relational database table my table in schema my schema The database table also needs to follow several F specific options before it can be used Each table requires two columns The first column must be named truth with type boolean The value of this column specifies whether each tuple is true or false The second reguired column must be named prior with type real or double The value of this co
3. intermediate data could blow up easily F avoids the explosion of data movement between operators with a novel cost based materialization strategy that utilizes the RDBMS s optimizer e Multi core Parallelism F will partition the input data and run sub operators on different por tions concurrently to exploit multi core CPUs For further technical details about F please read our technical report 2 or visit our website at http research cs wisc edu hazy felix This manual provides essential information about how to install set up and execute F Hereinafter we assume the readers are familiar with T 3 The current version of F only implements inference Users interested in weight learning should refer to T instead 2 Installation 2 1 Installing Prerequisites F has almost the same setup procedure as T Please visit T s website http research cs wisc edu hazy tuffy doc for details 2 2 Configuring Felix Download the latest version of F from After unpacking you should see a configuration file named felix conf This file contains necessary information F will use for execution Please refer to T s website http research cs wisc edu hazy tuffy doc for explanations of this configuration file 3 Usage Example In this section we use an example to demonstrate how to run F Input Inthestandardcase F s input is exactly the sameasT Again seeT s documentation for details We assume the following three fi
4. 10 GreenBay feature 1310 U00 GreenBay weight LOS U00 GreenBay 0 0698738377461130 weight NON U00 GreenBay 0 4092761660611854 weight WIN U00 GreenBay 0 4791500038073283 tcrt GreenBay GreenBayPackers Figure 2 Sample Input to F evidence db dwinner teamCluster date dloser teamCluster date Figure 3 Sample Input to F query db Correlation Clustering F will assign a predicate to a correlation clustering operator if a predicate has hard constraints associating with it which specify the three properties of equivalent relations ocoref a a reflexive ocoref a b ocoref b c gt ocoref a c transitive ocoref a b gt ocoref b a symmetric MLN F will assign the remaining predicates to the T operator 4 2 Special Predicates Besides standard MLN syntax used by T F also supports some special syntax The motivation of these special syntax are to ease the development of MLN programs If the program contains a predicate of the form P that is processed by the CC operator and in addition there is a predicate of the form P_map then the second predicate will also be filled by the CC operator The content is a linear representation of the clustering result P_map x y lt gt y is a representative element of the cluster x belongs to 1 When dumping the result of P F actually dumps P map s result because the former can be quadratic in the size of domain An
5. Operator of label seqidl wordid2 label3 with 3 Relevant Clauses Partitioned by label seqidl wordid2 label3 round segidl 1 0 4 1 LR Operator of label seqidl wordid2 label3 with 3 Relevant Clauses Partitioned by label seqidl wordid2 label3 round seqidl 1 0 4 3 LR Operator of label seqidl wordid2 label3 with 3 Relevant Clauses Partitioned by label seqidl wordid2 lLabel3 round seqidl 1 0 4 0 TUFFY Operator of dwinner wordl date2 dloser wordl date2 with 3 Relevant Clauses gt gt gt Parsing evidence file evidence db gt gt gt Storing evidence gt gt gt Optimize DMOs for COREF Operator of tcoref wordidl wordid2 with 4 Relevant Clauses Uses 0 min 7 sec gt gt gt Start Running COREF Operator of tcoref wordidl wordid2 with 4 Relevant Clauses gt gt gt COREF Operator of tcoref wordidl wordid2 with 4 Relevant Clauses uses 0 min 38 860 sec gt gt gt Dumping results for tcoref wordidl wordid2 1 45 gt gt gt Optimize DMOs for LR Operator of label seqidl wordid2 label3 with 3 Relevant Clauses Partitioned by label seqidl wordid2 label3 round seqidl 1 0 4 2 Uses 0 min 9 877 sec gt gt gt Optimize DMOs for LR Operator of label seqidl wordid2 label3 with 3 Relevant Clauses Partitioned by label seqidl wordid2 label3 round seqidl 1 0 4 1 Uses
6. USER MANUAL OF FELIX 0 2 Feng Niu Christopher R Jude Shavlik Josh Slauson Ce Zhang University of Wisconsin Madison leonn chrisre shavlik czhang slauson cs wisc edu August 22 2011 Contents 1 Overview 2 Installation 2 1 Instale Prerequistes Lus ced eed ar a ee a re Ee e He SE 22 TXOMHSUEHE E ann vee a as ee EE OE te en E 3 Usage Example 4 Compilation Mo MM bn ten eenn TEE RE Sl es EE EE EF 42 Special Predicates 22s sn Ah bees eh quede ERA pod rie PR Ee Rote ew d 5 Input Options SL Local Piles OE EE AO ei a Se deat EE ee ee A 52 Relational Database Tables s issus eR Romx ko Roe do RR eee ve 5 3 BlahBlah feature extraction language eee 6 Command Options 1 Overview F is a relational optimizer for statistical inference Recent years have seen a surge of sophisticated statistical frameworks with the relational data model via SOL logic like languages Examples include MLNs and PRMs While this movement has demonstrated significant quality advantage in numerous applications on small datasets efficiency and scalability have been a critical challenge to their deployment in Enterprise settings Felix addresses such performance challenges with an operator based approach where each operator performs a statistical algorithm with relational input output The key observations underlying Felix are 1 Task Decomposition Conventional approaches to statistical relational inference are monolithic in that not only do th
7. cal Hadoop This is useful for debugging program files with print statements mapreduce MapReduce address e g hdfs localhost 9001 nDD Number of iterations in dual decomposition nReduce Number of reducers to use in Hadoop WordCount word count hdfs hdfs localhost 9000 my_directory my_file xml 1 MAPINIT 1 import re JO OMAP xml title on inputvf for m in re finditer lt title gt lt title gt inputv title m group 1 for k in title split felixio collect k 1 yO OREDUCE on key values sum 0 for v in values sum sum int v felixio_push key sum JO Popular word WordCount word c c gt 100 gt Popular word Figure 5 Sample MLN program using BlahBlah XML WordLength word len hdfs hdfs localhost 9000 my_directory my_file 1 MAP on inputv for m in inputv split felixio_collect m len m JO LongWord word WordLength word len len gt 10 gt LongWord word Figure 6 Sample MLN program using BlahBlah text 10 References 1 P Domingos and D Lowd Markov Logic An Interface Layer for Artificial Intelligence 2009 2 2 F Niu C Zhang C R and J Shavlik Felix Scaling Inference for Markov Logic with an Operator based Approach http http arxiv org abs 1108 0294v1 Technical Report 2011 3 3 M Richardson and P Domingos Markov logic networks Machine Learning 2006 2 11
8. example of using P_map can be found in Figure 1 File Edit View Terminal Tabs Help czhang sebastian 1 java jar felix jar i prog mln e evidence db queryFile query db r output out Welcome to Felix 0 1 Database schema felix sebastian cs wisc edu czhang 20132 Current directory afs cs wisc edu u c z czhang Desktop Exp nfl ps nfl demo Temporary directory tmp tuffy workspace felix sebastian cs wisc edu czhang 20132 gt gt gt Connecting to RDBMS at jdbc postgresql localhost 5432 tuffydb gt gt gt Parsing program file prog mln gt gt gt Parsing query file query db gt gt gt Parsing Operators Add lr operator for label Add coref operator for tcoref Add mln operator for dwinner wordl date2 dloser wordl date2 gt gt gt The following operator is not decomposable COREF Operator of tcoref wordidl wordid2 with 4 Relevant Clauses gt gt gt Decomposing the following operator into 4 parts LR Operator of label seqidl wordid2 label3 with 3 Relevant Clauses gt gt gt The following operator is not decomposable TUFFY Operator of dwinner wordl date2 dloser wordl date2 with 3 Relevant Clauses gt gt gt Serialized Execution Plan COREF Operator of tcoref wordidl wordid2 with 4 Relevant Clauses LR Operator of label seqidl wordid2 label3 with 3 Relevant Clauses Partitioned by label seqidl wordid2 label3 round seqidl 1 0 4 2 LR
9. ey express complex tasks in a unified language they also attempt to solve the tasks by running a generic algorithm e g random walk greedy or sampling over the entire program This often results in poor quality and or miserable scalability On the other hand a complex task often consists of subtasks that have specialized algorithms with high quality and high efficiency For example many text analytics tasks can be decomposed into a handful of primitives such as classification clustering and ranking And so a natural question is Why not combine the ease and flexibility of the unified language with the high quality and efficiency of specialized statistical algorithms That is what F does 2 Data Partitioning Parallelism is king of scalability The king is happy when the data can be par titioned in embarrassingly obvious ways However the king is often puzzled in front of complex statistical tasks the point of such tasks after all is to embrace rich correlations F is set to find ways to identify parallelization opportunities in complex statistical tasks automatically In this prototype implementation of F we take Markov logic 3 1 as the unified language to ex press sophisticated statistical tasks Markov logic adds weights to first order logic rules and has rigorous semantics based on the exponential models It has been successfully applied to a wide range of appli cations including information extraction entity res
10. les are provided e prog mln an MLN program file e evidence db a file containing evidence tuples e query db a file specifying the queries Figures 1 2 3 show the content of these files See Section 5 for other input options You can run F with the following command http research cs wisc edu hazy tuffy java jar felix jar i prog mln e evidence db queryFile query db r out db F will decompose the input MLN and assign the rules to the following three operators e a Correlation Clustering Operator on tcoref e a Logistics Regression Operator on label eaT Operator on dwinner and dloser One possible output of this program will be dwinner GreenBay Dec 10 2006 F will generate Hou four output files out dloser out dwinner out label and out tcoref one for each open predicate Figure 4 shows a screencast of F running this program 4 Compilation 4 1 Operators F implements a best effort compiler to detect subtasks in the given MLN programs This section will show some syntactical patterns F uses to discover specialized subtasks Users who want exactly these operators can follow these patterns Logistic Regression F will discover a predicate that can be solved as logistic regression by finding key constraints in predicate declaration label seqid wordid label F will also tries to parse hard constraint specifying key constraints like label s w 1 1 11 gt label s w 11
11. lumn specifies the probability of each tuple The predicate arguments also need to follow some conventions For each predicate argument the database table needs a column starting with the type of the argument followed by the argument index starting at 1 The type of each predicate argument column must be string unless the argument is specified with the double or int_ option truth prior seqidi wordid2 word3 5 3 BlahBlah feature extraction language The final option for supplying evidence to F is through its feature extraction language called Blah Blah BlahBlah makes use of Hadoop and Python to easily extract features from unstructured and semi structured data on HDFS To use this option first edit the predicate declaration in the MLN program file The first line is similar to using a relational database table However instead of a schema and table a filepath on HDFS is specified See Figures 5 and 6 for example programs Within the predicate declaration there can be three F functions containing arbitrary python code MAPINIT This function can be used for initializing variables or importing packages Code here will only be run once on each node in the Hadoop cluster This function is optional MAP This function contains the Map code For XML files use lt xml tag gt where tag is name of the XML tag which contains relevant data Use felixio collect to generate data to be used in the Reduce function This fu
12. nction is required REDUCE This function contains the Reduce code It combines the data generated by the Map function and outputs it in the format specified by the predicate declaration with felixio_push The number of arguments sent to felixio_push should be exactly the same with the target relation each invocation of felixio_push will generate a tuple in the target relation This function is optional If not specified F will use a default reducer which will output each key value pair generated by the mapper Please note that the default reducer can only be used for predicates with exactly two arguments Notes e Python packages Currently F allows all python packages included in Jython s original Lib folder to be imported e Third party java code packages Arbitrary java code packages can be run following Jython s gram mar e Debugging Use the local option for debugging feature extraction programs Doing so displays output from print statements Required Options See Section 6 for details e auxSchema e hdfs e mapreduce nReduce 6 Command Options F has several options on top of what T has auxSchema Schema used for saving BlahBlah results e dd Run in dual decomposition mode explain Print out the execution plan of F without actually executing the program gp Use Greenplum instead of PostgreSQL hdfs HDFS address e g hdfs localhost 9000 local Connect to a lo
13. olution text mining and natural language processing Nevertheless state of the art implementations of Markov logic e g A 1 and our own T 2 are unaware of the above optimization opportunities that motivated F and so suffer from suboptimal quality and scalability F takes as input a Markov logic program with evidence data and outputs the same kinds of predic tions as what A and T would output given the same input if they can run that is Internally F performs the following steps 1 Compilation identifies specialized subtasks in the program and assign them to a predefined set of statistical operators including logistic regression LR conditional random field CRF correlation clustering CC and T the generic MLN inference engine 2 Optimization applies cost based data materialization strategies to ensure the efficiency of data movement between operators 3 Execution partitions the data with linear programming so that one operator can process the data in parallel Some technical highlights in F 0 2 lnttp alchemy cs washington edu http research cs wisc edu hazy tuffy e Three Specialized Operators F will discover Logistic Regression for classification Conditional Random Field for sequential labeling and Correlation Clustering for clustering or coreference subtasks in the input program and solve them using specialized algorithms Viterbi etc e Data Movement Optimization Explicitly materializing
Download Pdf Manuals
Related Search