Debian Codegraph @

CHANGES

raw
1 3.0.0
2 -----
3 Removed trove dependency.
4 Changed license to ASL.
5 Re-organized package structure to support up coming work.
6 Added perceptron classifier.
7
8 2.5.3
9 -----
10 Fixed bug in GIS class, a cutoff value was not passed on to the GISTrainer
11 Fixed bug in GISTrainer, a cutoff value compare was incorrect
12
13 2.5.2
14 -----
15 Wrapped model reading input streams with BufferedInputStream for performance gain.
16
17 2.5.1
18 -----
19 Fixed bugs with real-valued feature support.
20 Added unit tests for these bugs.
21
22 2.5.0
23 -----
24 Added support for real-valued features with RealBasicEventStream,
25 RealValueFileEventStream, OnePassRealValueDataIndexer, and
26 TwoPassRealValueDataIndexer classes.
27
28 Added support for priors with the Prior interface. By default the
29 package has been using a uniform prior when no other information is
30 known, now you can create other priors.
31
32 Set TwoPassDataIndexer to use UTF8 encoding.
33
34 Made GISModel thread safe for calls to the eval() method.
35
36 Refactored GISModel and GISTrainer to use common eval() method.
37
38 2.4.0
39 -----
40 Updated parameter datatype to re-use data structure storing which outcomes
41 are found with that predicate/context in GISModel. (Per suggestion by
42 Richard Northedge).
43
44 Made changes to support above change in GISTrainer, GISModelReader,
45 GISModelWriter, and OldFormatGISModelReader.
46
47 Changed smoothing boolean to be parameter instead of static variable in GIS.
48
49 Update sample code to reflect above change.
50
51 Extracted FileEventStream from TwoPassDataIndexer.
52
53 Improved javadoc.
54
55 2.3.0
56 -----
57 Added and updated javadoc including package descriptions.
58 Made new trove type to improve efficiency.
59
60 2.2.0
61 -----
62 Added TwoPassDataIndexer to use a temorpy file for storing the events. This
63 allows much larger models to be trained in a smaller amount of memory with no
64 noticible decrease in performance for large models.
65
66 Made minor additions to interface to allow more flexible access to model
67 outcomes.
68
69 2.1.0 (Major bug fixes)
70 -----
71
72 Fixed some bugs with the updating off the corrections paramater.
73 Namely the expected value needed to check if a particular context was
74 avalable with the outcome seen in training and if not add a term to
75 the expected value of the correction constant. (Tom)
76
77 Smoothing initial value wasn't in the log domain. (Tom)
78
79 Loglikelihood update as returned by nextIteration wasn't in the right
80 place so the loglikelihood value it returned was incorrect. (Tom)
81
82 Added cast to integer division in model update. (Tom, fixing bug
83 pointed out by Zhang Le)
84
85
86 2.0 (Major improvements)
87 ---
88 Fixed bug where singleton events are dropped. (Tom)
89
90 Added eval method so distribution could be passed in rathar then
91 allocated durring each call. Left old interface in place but modified
92 it to use the new eval method. Also made numfeats a class level
93 variable. (Tom)
94
95 Fixed cases where parameters which only occured with a single output
96 weren't getting updated. Ended up getting rid of pabi and cfvals
97 structures. These have been replaced with the data for a single event,
98 double[] modelDistribution, and this is used to update the modifiers
99 for a single event and then updated for each additional event. This
100 change made it easier to initialize the modleDistribution to the
101 uniform distribution which was necessary to fix teh above problem.
102 Also moved the computation of modelDistribution into it's own routine
103 which is name eval and is almost exactly the same as GISModel.eval w/o
104 doing the context string to integer mappings. (Tom)
105
106 Made correction constant non-optional. When the events all have the same
107 number of contexts then the model tries to make the expected value of the
108 correction constant nearly 0. This is needed because while the number of
109 contexts may be same it is very unlikly that all context occur with all
110 outcomes. (Tom)
111
112 Made nextIteration return a double which is the log-likelihood from
113 the previous itteration. At some point there isn't enough accuracy in
114 a double to make further iterations useful so the routine may stop
115 prematurly when the decrease in log-likelihood is too small. (Tom)
116
117
118 1.2.10
119 ------
120 Fixed minor bug (found by Arno Erpenbeck) in BasicContextGenerator: it
121 was retaining whitespace in the contextual predicates. (Jason)
122
123 Added error message to TrainEval's eval() method. (Jason)
124
125
126 1.2.9 (Bug fix release)
127 _____
128
129 Modified the cutoff loop in DataIndexer to use the increment() method
130 of TObjectIntHashMap. (Jason)
131
132 Fixed a bug (found by Chieu Hai Leong) in which the correctionConstant
133 of GISModel was an int that was used in division. Now,
134 correctionConstant is a double. (Jason)
135
136
137 1.2.8
138 _____
139
140 Modified GISTrainer to use the new increment() and adjustValue()
141 methods available in Trove 0.1.4 hashmaps. (Jason)
142
143 Set up the GISTrainer to use an initial capacity and load factor for
144 the big hashmaps it uses. The initial capacity is half the number of
145 outcomes, and the load factor is 0.9. (Jason)
146
147 (opennlp.maxent.DataIndexer) Do not index events with 0 active features.
148 (Eric)
149
150 Upgraded trove dependency to 0.1.1 (includes TIntArrayList, with reset())
151 (Eric)
152
153 (opennlp.maxent.DataIndexer)
154 Refactored event count computation so that the cutoff can be applied while
155 events are read. This obviates the need for a separate pass over the
156 predicates between event count computation and indexing. It also saves
157 memory by reducing the amount of temporary data needed and by avoiding
158 creation of instances of the Counter class. the applyCutoff() method
159 was no longer needed and so is gone. (Eric)
160
161 (opennlp.maxent.DataIndexer)
162 Made the event count computation + cutoff application also handle the
163 assignment of unique indexes to predicates that "make the cut." This
164 saves a fair amount of time in the indexing process. (Eric)
165
166 (opennlp.maxent.DataIndexer)
167 Refactored the indexing implementation so that TIntArrayLists are
168 (re-)used for constructing the array of predicate references associated
169 with each ComparableEvent. Using the TIntArrayList instead of an
170 ArrayList of Integers dramatically reduces the amount of garbage
171 produced during indexing; it's also smaller. (Eric)
172
173 (opennlp.maxent.DataIndexer)
174 removed toIntArray() method, since TIntArrayList provides the same
175 behavior without the cost of a loop over a List of Integers (Eric)
176
177 (opennlp.maxent.DataIndexer)
178 changed indexing Maps to TObjectIntHashMaps to save space in several
179 places. (Eric)
180
181 1.2.6
182 -----
183 Summary: efficiency improvements for model training.
184
185 Removed Colt dependency in favor of GNU Trove. (Eric)
186
187 Refactored index() method in DataIndexer so that only one pass over the
188 list of events is needed. This saves time (of course) and also space,
189 since it's no longer necessary to allocate temporary data structures to
190 share data between two loops. (Eric)
191
192 Refactored sorting/merging algorithm for ComparableEvents so that
193 merging can be done in place. This makes it possible to merge without
194 copying duplicate events into sublists and so improves the indexer's
195 ability to work on large data sets with a reasonable amount of memory.
196 There is still more to be done in this department, however. (Eric)
197
198 The output directory of the build structure is now "output" instead of
199 "build". (Jason)
200
201 1.2.4
202 -----
203 Added options for doing *very* simple smoothing, in which we 'observe'
204 features that we didn't actually see in the training data. Seems to
205 improve performance for models with small data sets and only a few
206 outcomes, though it conversely appears to degrade those with lots of
207 data and lots of outcomes.
208
209 Added BasicEventStream and BasicContextGenerator classes which assume
210 that the contextual predicates and outcomes are just sitting pretty in
211 lines. This allows the events to be stored in a file and then read in
212 for training without scanning around producing the events everytime.
213
214 Added sample application "sports" to help with testing model behavior
215 and to act as an example to help newbies use the toolkit and build
216 their own maxent applications.
217
218 Fixed bug in TrainEval in which the number of iterations and the
219 cutoff were swapped in the call to train the model.
220
221 PerlHelp and BasicEnglishAffixes classes were moved out of Maxent so
222 that gnu-regexp.jar is no longer needed.
223
224
225 1.2.2
226 -----
227 Added "exe" to build.xml to create stand alone jar files. Also added
228 structure and "homepage" target so that it is easier to keep homepage
229 up-to-date.
230
231 Added IntegerPool, which manages a pool of shareable, read-only Integer
232 objects in a fixed range. This is used in several places where Integer
233 objects were previously created and then GCed.
234
235 In various places, used Collections API features to speed things up.
236 For example, java.util.List.toArray() will do a System.arraycopy, if
237 given a big enough array to copy into. This is, therefore, much
238 faster.
239
240 Added getAllOutcomes() method in GISModel to show the String names of
241 all outcomes matched up with their normalized probabilities.
242
243
244 1.2.0
245 -----
246 Changed license to LGPL.
247
248 Added build.xml file to be used with the build tool Jakarta Ant.
249
250 Work sponsored by Electric Knowledge:
251 Added BasicEnglishAffixes class to perform basic morphological
252 stemming for English.
253
254 Fixed a bug pointed out by Chieu Hai Leong in which the model would
255 not train properly in situations where the number of features was
256 constant (and therefore, no correction features need to be computed).
257
258 The top level package name has changed from quipu to opennlp. Thus,
259 what was "quipu.maxent" is now "opennlp.maxent". See
260 http://www.opennlp.com for more details.
261
262 Lots of little tweaks to reduce memory consumption.
263
264 Several changes sponsored by eTranslate Inc:
265
266 * The new opennlp.maxent.io subpackage. The input/output system for
267 models is now designed to facilitate storing and retrieving of
268 models in different formats. At present, GISModelReaders and
269 GISModelWriters have been supplied for reading and writing plain
270 text and binary files and either can be gzipped or
271 uncompressed. The OldFormatGISModelReader can be used to read old
272 models (from Maxent 1.0), and also provides command line
273 functionality to convert old models to a new format. Also,
274 SuffixSensitiveGISModelReader and SuffixSensitiveGISModelWriter
275 have been provided to allow models to be stored and retrieved
276 appropriated based on the suffixes on their file names.
277
278 * Model training no longer automatically persists the new model to
279 disk. Instead it returns a GISModel object which can then be
280 persisted using an object from one of the GISModelWriter classes.
281
282 * Model training now relies on EventStreams rather than
283 EventCollectors. An EventStream is fed to the DataIndexer
284 directly without the developer needing to invoke the DataIndex
285 class explicitly. A good way to feed an EventStream the data it
286 needs to form events is to use a DataStream that can return
287 Objects from a data source in a format and os-independent
288 manner. An implementation of the DataStream interface for reading
289 plain text files line by line is provide in the
290 opennlp.maxent.PlainTextByLineDataStream class.
291
292 In order to retain backwards compatability, the
293 EventCollectorAsStream class is provided as a wrapper over the
294 EventCollectors used in Maxent 1.0.
295
296 * GISModel is now thread-safe. Thus, one maxent application can
297 service multiple evaluations in parallel with a single model.
298
299 * The opennlp.maxent.ModelReplacementManager class has been added to
300 allow a maxent application to replace its current maxent model
301 with a newly trained one in a thread-safe manner without stopping
302 the servicing of requests.
303
304 An alternative to the ModelReplacementManager is to use a
305 DomainToModelMap object to record the mapping between different
306 data domains to models which are optimized for them. This class
307 allows models to be swapped in a thread-safe manner as well.
308
309 * The GIS class now is a factory which invokes a new GISTrainer
310 whenever a new model is being trained. Since GISTrainer has only
311 local variables and methods, multiple models can be trained
312 simultaneously on different threads. With the previous
313 implementation, requests to train new models were forced to queue
314 up.
315
316 1.0
317 _____
318 Reworked the GIS algorithm to use efficient data structures
319 * Tables matching things like predicates, probabilities, correction
320 values to their outcomes now use OpenIntDoubleHashMaps.
321 * Several functions over OpenIntDoubleHashMaps are now defined,
322 and most of the work of the iteration loop is in fact done by
323 these.
324
325 Events with the same outcome and contextual predicates are collapsed
326 to reduce the number of tokens which must be iterated through in
327 several loops. The number of times each event "type" is seen is then
328 stored in an array (numTimesEventSeen) to provide the proper values.
329
330 GISModel uses less memory for models with many outcomes, and is much
331 faster on them as well. Performance is roughly the same for models
332 with only two outcomes.
333
334 More code documentation.
335
336 Fully compatible with models built using version 0.2.0.
337
338
339 0.2.0
340 _____
341 Initial release of fully functional maxent package.