/rgs-struct

Primary LanguageScala

Randomized Greedy Search for Structured Prediction

Related Publication

Randomized Greedy Search for Structured Prediction: Amortized Inference and Learning
Chao Ma, F A Rezaur Rahman Chowdhury, Aryan Deshwal, Md Rakibul Islam, Janardhan Rao Doppa, and Dan Roth

Requirement Packages

The project is coded in both Java and Scala. The Scala code is only employed in the NLP applications. Please include Java 8 JRE library and Scala 2.11 library in project or building path.

You also need to install the Xgboost JVM-Package. Xgboost is only used for an alternamtive approach to learn the initialization classifier $h$. Note that Xgboost is not mandatory because you can always employ Logistic Regression model for $h$.

All other libraries can be downloaded from here.

Directory Tree of the Project

The root of the project contains four folders:

.
|   +-- datasets
|   +-- logistic_models
|   +-- rgs-proj
|   +-- sl-config
  • datasets: The sub directory contains all data set files (We will describe more details about the Data Files section).
  • rgs-proj: The source code directory that contains all Java and Scala files.
  • logistic_models: The models files of the initialization classifier $h$ in RGS($\alpha$) (See more details of IID Classifier paragraph in section 3 in the paper).
  • sl-config: The configuration files of structured SVM learner for different task and datasets.

Data Files

Quick Start

The main java file to run the project is experiment.RndLocalSearchExperiment.java.

The following is the description of running option:

-name: Dataset name. Candidate names: {HW_SMALL, HW_LARGE, NETTALK_STREE, NETTALK_PHONEME, YEAST, ENRON, BIBTEX, BOOKMARKS, ...(more detailed data file description will be provided in dataset section) } -startyp = Initialization state generation type. Candidate values: { UNIFORM_INIT (corresponds to RGS(0)), LOGISTIC_INIT (corresponds to RGS(1.0)), ALPHA_INIT (corresponds to RGS($\alpha$)) } -initAlpha: The alpha value for RGS($\alpha$). This is only appilcable when starttyp is set to be ALPHA_INIT.

-restartTrain: Number of random restart for training. -restartTest: Number of random restart for testing. -mlType: Multi-label loss function type. Candidate values: { HAMMING_LOSS, EXMPF1_LOSS, EXMPACC_LOSS } -debug: Output debug info in logistic regression or not.

-cfgPath: SSVM config file path (See more details about Illinois-SL). -svmModelPath: SSVM model file path. -logsPath: Logistic regression model path. This is model file of initialization classifier $h$ in RGS($\alpha$).

-usePairFeat: Use binary feature when it is true. -useTernFeat: Use tenery feature when it is true. -useQuadFeat: Use quatery feature when it is true.

-runEvalLearn: Do E-learning when it is true. -eIter: Number of iterations of E-learning.

-doCostCache Do cost weight caching? Once the cost weight was cached, during testing the system will directly load the existing weight vector instead of re-running the training.

For example, if you want run the system on multi-label dataset Yeast, you can compile the project to a runable jar file rgs.jar with , and run with follow command:

java -jar rgs.jar -name YEAST -startyp LOGISTIC_INIT -restartTrain 1 -restartTest 1 -usePairFeat true -useTernFeat false -useQuadFeat false -mlType EXMPF1_LOSS -runEvalLearn false -doCostCache false

Contact
Chao Ma (machao@engr.orst.edu, nkg114mc@hotmail.com)
Oregon State University.