-
Notifications
You must be signed in to change notification settings - Fork 116
/
ChangeLog
642 lines (381 loc) · 20.1 KB
/
ChangeLog
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
2016-06-26 Giuseppe Attardi <[email protected]>
* deepnl/corpus.py: moved from __future__ import print_function to top.
* deepnl/embeddings.py: idem.
2016-04-23 Giuseppe Attardi <[email protected]>
* deepnl/words.pyx (LmWorker.__init__): fixed.
* deepnl/*.py: use print function for Python 3 compatibility.
2016-03-07 Giuseppe Attardi <[email protected]>
* deepnl/words.pyx (train.train_worker): get left_context from
self_prepadding.
2016-03-05 Giuseppe Attardi <attardi@R730-gpu>
* deepnl/reader.py (TweetReader.__init__): specify fileds for text and polarity.
(TweetReader.create_vocabulary): fixed indentation.
(TweetReader.read): added optional polarity dictionary.
2016-02-03 Antonio Valerio Miceli Barone < github.com/Avmb >
* deepnl/HPCA.cpp: renamed to deepnl/HPCA_impl.cpp in order to avoid a
conflict with deepnl/hpca.cpp in case-insensitive file systems.
* deepnl/hpca.pyx: updated to reflect the above change.
* setup.py: idem.
2016-01-06 Giuseppe Attardi <[email protected]>
* deepnl/trainer.pyx (train): initialize AdaGrad accumulators by adaEps.
* deepnl/extractors.pyx (Extractor.adaGradInit): initialize with adaEps.
* bin/dl-conv.py (main): added option --eps.
2015-12-24 Giuseppe Attardi <[email protected]>
* deepnl/networkconv.pyx (backpropagate): use MultiMarginCriterion
also for computing gradients.
2015-12-18 Giuseppe Attardi <[email protected]>
* deepnl/utils.py (create_feature_tables): removed.
* bin/dl-sentiwords.py (create_trainer): consider parameter --window
as left/right context, consistently with word2vec.
* bin/dl-conv.py (main): idem.
* bin/dl-ner.py (main): idem.
* bin/dl-pos.py (main): idem.
* bin/dl-words.py (main): idem.
2015-11-18 Giuseppe Attardi <[email protected]>
* bin/dl-conv.py (main): added options text_field and label_field
to select fields from tsv input files.
* deepnl/network.pyx (ConvolutionalNetwork.backpropagate):
in np.where condition > -1.
* deepnl/networkconv.pyx (ConvolutionalNetwork.backpropagate): use
MultiMarginCriterion
in np.where condition > -1.
* deepnl/classifier.pyx (Classifier.predict: use argmax(vars.output)
2015-11-04 Giuseppe Attardi <[email protected]>
* deepnl/network.pyx (Network.backpropagate): grads.hiddenx_bias
instead of vars.hidden_bias.
2015-10-14 Giuseppe Attardi <[email protected]>
* bin/dl-sentiwords.py (main): pass minOccurr and vocab_size to reader.create_vocabulary.
2015-10-08 Giuseppe Attardi <[email protected]>
* deepnl/networkseq.pyx (SeqParameters.initialize): put back high to 1.0.
2015-10-07 Giuseppe Attardi <[email protected]>
* deepnl/trainerconv.pyx (ConvTrainer): holds a Classifier, which
saves to file after training.
* deepnl/utils.py (import_class): added.
2015-10-05 Giuseppe Attardi <[email protected]>
* deepnl/sentiwords.pyx (SentimentTrainer._train_pair_s): compute
also grads.input.
2015-10-04 Giuseppe Attardi <[email protected]>
* deepnl/tagger.pyx (Tagger.tag_sequence): return list of tags.
* deepnl/extractors.pyx (save): save also fields and use them in load.
* bin/dl-ner.py (main): pass whole tokens to tagger.
* deepnl/networkseq.pyx (_viterbi): preallocate prev_score_and_trans.
* deepnl/word_dictionary.py (WordDictionary.__setitem__): do not
insert normalized word if already present.
2015-10-02 Giuseppe Attardi <[email protected]>
* deepnl/networkseq.pyx (SeqParameters.initialize): high = 0.01
2015-10-01 Giuseppe Attardi <[email protected]>
* deepnl/network.pyx (Variables.addSquare): moved here from Gradients.
* deepnl/trainer.pyx (Trainer): turned adaEps, adaRo etc. into
Trainer attributes.
2015-09-29 Giuseppe Attardi <[email protected]>
* deepnl/network.pxd: use int32 instead of int64 which is weird
default for Python and Numpy.
2015-09-27 Giuseppe Attardi <[email protected]>
* deepnl/extractors.pyx (Converter): fully reworked: see below.
* deepnl/extractors.pyx (Converter.convert): It passes to each
extractor a sentence, which could be either a list of words or a list of
tokens. It then calls Extractor.extract() passing either the token
field index to use or -1 for the whole token.
* deepnl/extractors.pyx (Extractor.extract): the default behavior
is to perform a lookup in the extractor dictionary of the proper
word.
The method is specialized in derived classes.
Extractors are never called on padding tokens, which are dealt instead
by methods get_padding_left(), get_padding_right()
* deepnl/word_dictionary.py (WordDictionary.__init__): default is 'senna'.
(WordDictionary.check): eliminated.
(WordDictionary): turned left_padding, right_padding and rare into
indices, rather than strings. Strings are only needed at initialization.
* bin/dl-ner.py (main): variant None in POS extractor (no normalization).
2015-09-26 Giuseppe Attardi <[email protected]>
* deepnl/extractors.pyx (GazetteerExtractor.extract): no clobbering of words.
2015-09-25 Giuseppe Attardi <[email protected]>
* bin/dl-ner.py (main): fixed AttributeExtractor initialization.
* deepnl/extractors.pxd (class AttributeExtractor): added.
* deepnl/trainerconv.pyx (_train_epoch): fixed nn.variables() size.
* deepnl/reader.py (ClassifyReader): label_field -2, text_field -1.
2015-09-24 Giuseppe Attardi <[email protected]>
* bin/dl-ner.py (main): added option --pos to use POS tag features.
2015-09-22 Giuseppe Attardi <[email protected]>
* deepnl/tagger.pyx (_tag_sequence): lookup whole sentence at once.
* deepnl/networkseq.pyx (SequenceNetwork._calculate_gradients_sll): added parameter skipErr.
* deepnl/network.pyx (skipErr): set to 0.01 (was 1e-4). This might impact speed.
* deepnl/extractors.pyx (normalize): added.
2015-09-21 Giuseppe Attardi <[email protected]>
* deepnl/word_dictionary.py (WordDictionary.__setitem__): update
words as well.
* deepnl/extractors.pyx (save_vectors): use self.dict.words.
* deepnl/word_dictionary.py (WordDictionary.__init__): words are
normalized before insertion.
2015-09-19 Giuseppe Attardi <[email protected]>
* deepnl/extractors.pyx (merge): avoid getdefault().
* deepnl/word_dictionary.py (WordDictionary.__setitem__): added.
(WordDictionary._normalize): added.
* bin/dl-ner.py (main): use %(default)s in help string.
* bin/dl-pos.py (main): idem.
2015-09-18 Giuseppe Attardi <[email protected]>
* deepnl/word_dictionary.py (WordDictionary.size): added.
2015-08-24 Giuseppe Attardi <[email protected]>
* deepnl/trainer.pyx (Trainer.rain): stop after max_no_progress
epochs without progress.
* deepnl/network.pyx: use backward activation function hardath_back.
* deepnl/math.pyx (hardtanh_back): backward function for hardtanh
activation layer, that produces gradients of input by multiplying
by gradiets of output.
* deepnl/tagger.pyx: cpdef tag_sequence.
* bin/dl-words.py: commented out profiling.
* bin/dl-pos.py (main): updated similarly to dl-ner.py.
* deepnl/reader.py (PosReader.__init__): handle form and tag fields.
* deepnl/sentiwords.pyx (train): modified implementation of
AdaGrad.
* deepnl/extractors.pyx (Converter.upgrade): perform properly
AdaGrad, using a table of accumulated squares in each extractor.
Added method clearAdagrad() to initialize them, which should be
called at the beginning of training.
2015-08-23 Giuseppe Attardi <[email protected]>
* bin/dl-ner.py (prefix): default to 5.
2015-08-21 Giuseppe Attardi <[email protected]>
* deepnl/trainer.pyx (_train_epoch): add padding to sentence.
* deepnl/tagger.pyx (_tag_sequence): sentence is padded.
* bin/dl-ner.py (main): added options --eps and --ro.
* deepnl/network.pyx (__init__): Initialization protocol for
class networks now requires that Parameters object be passed, for
avoiding double initialization in derived classes.
* bin/dl-ner.py (main): default for vocab-size is 0.
* deepnl/networkseq.pyx: fixed self.p.transitions.
2015-08-20 Giuseppe Attardi <[email protected]>
* deepnl/trainer.pyx (save_vectors): pass variant.
* deepnl/network.pyx: use multiclass hinge loss.
* deepnl/networkconv.pyx: fixed hinge calculation.
* deepnl/math.pyx: revert to np.ndarray since memoryviews don't
have max operator.
* bin/dl-ner.py (main): SuffixExtractor.create() is just
SuffixExtractor().
* deepnl/sentiwords.pyx (train): invoke epoch_report().
* deepnl/extractors.pyx (save_vocabulary): restored.
2015-08-16 Giuseppe Attardi <[email protected]>
* deepnl/network.pxd ((l1_decay, l2_decay, momentum, adaRo,
adaEps): training parameters declared as globals.
* deepnl/trainer.pyx (Trainer.__init__): grouped optional parameters in
single options dictionary.
Network must be passed as argument, hence scripts dl-ner.py,
dl-pos.py etc. have been modified to create it and then pass it,
as well as to group options into dictionary.
* bin/dl-ner.py (main): added argument vocab-size.
* bin/dl-conv.py (main): grouped parser arguments.
* bin/dl-ner.py (main): grouped parser arguments.
2015-08-15 Giuseppe Attardi <[email protected]>
* deepnl/network.pxd (Network): moved error to class Trainer.
2015-08-14 Giuseppe Attardi <[email protected]>
* deepnl/math.pyx: started moving to using Memoryviews.
* deepnl/words.pyx (LmTrainer.update): call nn.p.update().
* deepnl/trainer.pyx (Trainer.update): use AdaGrad.
* deepnl/network.pyx (Parameters.copy): added.
2015-08-13 Giuseppe Attardi <[email protected]>
* deepnl/networkconv.pyx: revised.
* deepnl/network.pyx (Network.run): renamed to forward.
* deepnl/networkconv.pyx (ConvolutionalNetwork.run): renamed to forward.
* deepnl/networkconv.pyx: use multiclass hinge function.
2015-08-10 Giuseppe Attardi <[email protected]>
* deepnl/networkconv.pyx (ConvParameters): similar to Parameters.
* deepnl/networkconv.pyx (ConvolutionaleNetwor): use ConvParameters field.
* deepnl/networkseq.pyx (SeqParameters): added.
* deepnl/networkseq.pyx (SeqquenceNetwor): use SeqParameters field.
* deepnl/network.pyx (Parameters): Parameters turned into a field
of Network.
Added methods save, load and update.
* deepnl/network.pyx (Network.update): moved to Parameters.
* bin/dl-sentiwords.py: save vocabulary also whe using word2vec.
2015-08-09 Giuseppe Attardi <[email protected]>
* deepnl/reader.py (TweetReader.read): drop empty tweets.
2015-08-06 Giuseppe Attardi <[email protected]>
* deepnl/networkconv.pyx: completed.
2015-07-22 Giuseppe Attardi <[email protected]>
* deepnl/extractors.pyx (GazetteerExtractor): use Trie for ngrams.
* deepnl/utils.py (Trie): for storing and searching ngrams.
2015-07-20 Giuseppe Attardi <[email protected]>
* deepnl/extractors.pyx (extract): use trie of ngrams for gazetteers.
2015-07-17 Giuseppe Attardi <[email protected]>
* bin/dl-ner.py (main): use GazetteerExtractor.build()
* deepnl/extractors.pyx (build): build gazetteers with ngrams from
sentences.
2015-07-16 Giuseppe Attardi <[email protected]>
* deepnl/extractors.pyx (build): optionally put suffixes in lower case.
2015-07-16 Giuseppe Attardi <[email protected]>
* bin/dl-ner.py (main): create affixes and gazetteer if not supplied.
Create additional vocabulary from corpus.
Added parameter --vocab-size to specify size of created vocabulary.
Insert libdir to the front of sys.path, to give it precedence over
installed library.
* deepnl/extractors.pyx (AffixExtractorwrite): added.
2015-07-15 Giuseppe Attardi <[email protected]>
* deepnl/reader.py (TaggerReader.__init__): added formField, tagField
* deepnl/ner_tagger.py (ToIOBES.__iter__): added tagField
2015-07-14 giuseppe attardi <[email protected]>
* deepnl/embeddings.py (Word2Vec.save): added.
2015-07-14 Giuseppe Attardi <[email protected]>
* deepnl/extractors.pyx (create): preserve insertion order.
* bin/dl-ner.py (main): extend provided vocabulary from words in
corpus.
(main): create the gazetteers from the training corpus.
2015-06-29 Giuseppe Attardi <[email protected]>
* deepnl/extractors.pyx (save_vocabulary): preserve order by ID.
* deepnl/sentiwords.pyx (_train_pair_s): fixed computing
grads.output_weights.
2015-06-28 Giuseppe Attardi <[email protected]>
* deepnl/sentiwords.pyx (_update): fixed AdaGrad step.
2015-06-25 Giuseppe Attardi <[email protected]>
* deepnl/sentiwords.pyx (itertrie): yield unigram
2015-06-24 Giuseppe Attardi <[email protected]>
* deepnl/extractors.pyx (__init__): load vectors also when vocab
is supplied.
2015-06-17 Giuseppe Attardi <[email protected]>
* deepnl/extractors.pyx (merge): fixed iteration on vocab.
* bin/dl-sentiwords.py: made train a positional argument
2015-05-17 Giuseppe Attardi <[email protected]>
* deepnl/reader.py (TweetReader.read): fixed bug causing skipping
positive tweets.
(TweetReader.read): keep also neutral tweets.
2015-05-16 Giuseppe Attardi <[email protected]>
* deepnl/words.pyx (LmWorker._extract_window): moved to LmTrainer
for sharing with SentimentTrainer.
* deepnl/trainer.pyx (Trainer.update): moved here from Tagger.
* deepnl/network.pyx (Network): modularized dealing with variables
and gradients. Each network type provides methods variables() and
gradients() to allocate the variables or gradients it uses.
* deepnl/math.pyx (hardtanh, hardtanhd): use nditer.
* deepnl/networkseq.pyx (_calculate_gradients_sll): dropped case
for null self.transitions.
avoid allocations by using out parameter or [:] assignment.
* deepnl/networkseq.pyx (update): dropped case for null
self.transitions.
2015-05-14 Giuseppe Attardi <[email protected]>
* deepnl/embeddings.py (Word2Vec.read_vectors): renamed load().
* deepnl/extractors.pyx (__init__): use Word2Vec.
* deepnl/embeddings.py: dropped Gensim class since Gensim dropped
support for its own file format.
(Word2Vect): renamed to Word2Vec.
2015-05-10 Giuseppe Attardi <[email protected]>
* deepnl/corpus.py (TsvReader): added.
* deepnl/math.pyx (hardtanh, hardtanhd): use np.nditer.
(tanhe): use slicing to avoid copying.
2015-04-28 Giuseppe Attardi <[email protected]>
* bin/dl-words.py (main): missing default to param words.
(main): converter.add().
2015-04-24 Giuseppe Attardi <[email protected]>
* bin/dl-ner.py (main): pass tokens to SuffixExtractor()
* deepnl/word_dictionary.py (WordDictionary.__init__): check for None,
since index might be 0.
2015-04-23 Giuseppe Attardi <[email protected]>
* deepnl/embeddings.py (Gensim.read_vectors): do not create weights for
special symbols here.
* deepnl/word_dictionary.py (WordDictionary.__init__): padding for
word2vect.
* deepnl/extractors.pyx (AffixExtractor.__init__): accept wordlist as
argument.
2015-04-24 Giuseppe Attardi <[email protected]>
* deepnl/extractors.pyx (Embeddings.__init__): add vectors for
special symbols.
2015-04-23 Giuseppe Attardi <[email protected]>
* deepnl/word_dictionary.py (WordDictionary.__contains__):
lowercase only for SENNA.
(WordDictionary.__getitem__): check is not None since idx might be
None.
* bin/dl-pos.py (main): pass word list to SuffixExtractor.
2015-04-02 Giuseppe Attardi <[email protected]>
* deepnl/networkseq.pyx: softmax by column requires axis=0.
2015-03-29 Giuseppe Attardi <[email protected]>
* deepnl/trainer.pyx (_train_epoch): invoke Converter.update().
* deepnl/extractors.pyx (Converter.update): added.
2015-03-26 Giuseppe Attardi <[email protected]>
* deepnl/extractors.pyx (create): added lowcase option.
2015-01-29 Giuseppe Attardi <[email protected]>
* deepnl/trainer.pyx (MovingAverage): added.
2015-01-27 Giuseppe Attardi <[email protected]>
* bin/dl-sentiwords.py: added.
2015-01-08 Giuseppe Attardi <[email protected]>
* deepnl/reader.py (TextReader.read): removed any logger commands:
logging should be done by callers.
2015-01-04 Giuseppe Attardi <[email protected]>
* deepnl/extractors.py (Extractor): feature are now extracted by Extractor
objects, each one dealing with one kind of features, e.g. embeddings,
pos, caps, etc.
Each one is responsible for saving and loading its own data to a
single model file.
2015-01-03 Giuseppe Attardi <[email protected]>
* deepnl/srl/srl_reader.py (read_conll): split fields at \t.
(SRLReader._codify_sentences): use SentencexsConverter instead of
TokenConverter.
* deepnl/attributes.py (TokenConverter): renamed SentenceConverter since
it needs to operate at the sentence level, e.g. the NER converter
needs to lookup ngrams in the gazetteers.
* deepnl/reader.py (Reader.create_converter): return converter rather
than assign it to self.
* deepnl/taggers.py (load_data): pass parameter path instead of using
property sys.paths.
* deepnl/network.pyx: method load_features() does the job previously done
in load_data() in taggers.py.
(load_from_file): renamed to load_weights.
* deepnl/corpus.py: reader/writer for various corpus formats.
2014-12-30 Giuseppe Attardi <[email protected]>
* deepnl/networklm.pyx: split half_window into left_context and
right_context since they might not be necessarily the same.
* deepnl/srl/srl_reader.py: moved here relevant parts from nlpnet/read_data.py
* nlpnet/read_data.py: removed.
2014-12-26 Giuseppe Attardi <[email protected]>
* deepnl/word_dictionary.py (WordDictionary): make senna the default variant
(WordDictionary): just use a single padding for both left and right
2014-12-23 Giuseppe Attardi <[email protected]>
* metadata.py (Metadata.save_to_file): use file names from config.py
* config.py: added sslm portion.
Use consistent naming convenion of prefix task nae rather than postfix.
* deepnl/reader.py (TextReader.codify_sentences): convert_sentences -> convert
* deepnl/network.pyx (hardtanhe): derivative of hardtanh in terms of its
output.
* deepnl/networklm.pyx (_train_pair): use hardtanhe instead of hardtanhd.
2014-12-22 Giuseppe Attardi <[email protected]>
* deepnl/networklm.pyx (_new_random_pool): generate a simple list, not a
list of singletons.
2014-12-20 Giuseppe Attardi <[email protected]>
* nlpnet/networkSent.pyx: network for sentiment-specific word
embeddings.
2014-05-06 Giuseppe Attardi <[email protected]>
* deepnl/utils.py (generate_feature_vectors): use random.uniform.
* deepnl/network.pyx (_viterbi): fixed bug.
2014-05-05 Giuseppe Attardi <[email protected]>
* deepnl/network.pyx (_calculate_gradients_sll): use softmax(delta[-1]).
* deepnl/ner/ner_reader.py (noGazetteer): for cases where there is on
entry for a class in the gazetteer dictionary.
2014-05-03 Giuseppe Attardi <[email protected]>
* deepnl/network.pyx (softmax): added softmax function, for computing
softmax directly, without resorting to logsumexp.
2014-05-02 Giuseppe Attardi <[email protected]>
* deepnl/network.pyx (_calculate_delta): avoid transpose by using
logadd = logsumexp(delta[token - 1][:,np.newaxis] + transitions, 0)
instead of
logadd = logsumexp(delta[token - 1] + transitions_T, 1)
2014-04-23 Giuseppe Attardi <[email protected]>
* deepnl/networkconv.pyx (_convolution_lookup): use lookup().
* deepnl/networklm.pyx (_train_pair): use lookup().
* deepnl/network.pyx (lookup): added, so that run() method can get the
input directly and we need not to store the input in the network.
* deepnl/networkconv.pyx (_tag_sentence): use logsumexp to avoid numeric
instability.
* deepnl/networklm.pyx: class LanguageModel derived from Network so that
it inherits attributes and methods, including run().
* deepnl/arguments.py (get_args): set default to 0.01 for --lf and -lt.
(get_args): made option --gold required.
2014-04-22 Giuseppe Attardi <[email protected]>
* bin/nlpnet-train.py (saver): use random.seed() if you want to
ensure replicability.
2014-04-19 Giuseppe Attardi <[email protected]>
* bin/nlpnet-train.py (create_network): moved initialization of
transitions to Network.create_new().
2014-04-17 Giuseppe Attardi <[email protected]>
* bin/nlpnet-train.py (create_network): initialize transition
randomly rather than 0.
2014-03-10 Giuseppe Attardi <[email protected]>
* bin/nlpnet-tag.py (_print_tagged_ner): handle ner output.
* bin/nlpnet-train.py (create_reader): added case for ner.
(create_network): similarly
(train): first argument is text_reader, not reader