search

pmanion0 / cs224n-pa4

GitHub repository for CS224n's PA4 - Neural Network for Named Entity Recognition
0 stars 0 forks source link

readme

cs224n-pa4

Launcher

Launcher is used to run a single configuration. Use the following command to launch the program from the cs224n-pa4/java folder:

java -Xmx2g -cp "classes:extlib/*" main.Launcher <options>

The list of possible options, with their defaults, is provided below:

Option Type Default Description
-train String ../data/train Path to the training data
-test String ../data/test Path to the testing data
-wordvec String ../data/wordVectors.txt Path to the WordVector file
-vocab String ../data/vocab.txt Path to the Vocab listing file
-outfile String ../scored.out Output file location
-unkword String UUUNKKK Word in the vocab and
-entities , Delimited String O,LOC,MISC,ORG,PER List of possible output entities
-lambda Double 0.0 Lambda value for regularization
-learnrate Double 1e-4 Learning rate for SGD
-iters Int 10 Maximum number of iterations
-windowsize Int 3 Rolling window size to input to the NN
-wordvecdim Int 50 Dimension of the word vector
-learnwordvec Boolean false Update word vectors during training
-lowcaseall Boolean true Ignore case in the word vectors
-hiddendim , Delimited Int 10 Size of each hidden layer in the NN

TestConfigLauncher

TestConfigLauncher is used to run a series of tests that are described in a test configuration file. This file allows you to submit a number of closely related configurations to assess the impact of different settings. Use the following command to launch the program from the cs224n-pa4/java folder:

java -Xmx2g -cp "classes:extlib/*" main.TestConfigLauncher <test_config_file> <?output_dir>

 Parameter    | Required? | Description

------------------|-----------|------------ test_config_file | Yes | Path to a test configuration file (format described below) output_dir | No | Root directory to put all scored test outputs (labeled as test_1... test_N) along with a summary file of all tests


The test config file is parsed by the TestConfigurationReader. This reader begins by initializing a default configuration and proceeds to read all user defined settings until a test case delimiter ---- is encountered. At that point, a test run will be created with the settings as of that point in the file. This can be done repeatedly to define as many test cases as you would like.

The options are remembered in a rolling manner as the reader goes through the file, so any options defined in one test will carry over into all future tests unless they are explicitly overwritten.

Options are defined in <option>: <value> pairs in the file. White space for these settings are ignored, so make it look readable!

The first line and all test case delimiter lines must begin with ---- (no leading whitespace). However, you can add comments on the same line to describe the test configuration.

All options outlined in the Launcher documentation above are valid here, but make sure to exclude the - prefix, e.g. the command line option -iters should show up as iters below.

Structure:

---- Initial Setup
  <option1>: <value>
            ...
  <optionN>: <value>
---- Test Case #1 Created
  <option1>: <value>
            ...
  <optionN>: <value>
---- Test Case #2 Created

Example:

---- Define all general settings for the test
  train:      ../data/train
  test:       ../data/dev
  wordvec:    ../data/wordVectors.txt
  vocab:      ../data/vocab.txt
  lambda:     0.0
  learnrate:  0.001
  windowsize: 5
  iters:      100
---- Create the 1st test run with all options above ^
  lambda:     0.03
---- Create a 2nd similar run but now lambda = 0.03
  learnrate: 0.01
  learnrate: 0.001
---- Create a 3rd similar run but now learnrate = 0.001 (lambda is still 0.03)

Notice above that for the 3rd run, the first learnrate setting is overridden by the second learnrate call!