-
Notifications
You must be signed in to change notification settings - Fork 3
/
lstm_context.py
537 lines (432 loc) · 20.8 KB
/
lstm_context.py
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
# -*- coding: utf-8 -*-
import sys
import random,json,codecs
from collections import OrderedDict
from collections import defaultdict
from itertools import count
import dynet as dy
import numpy as np
import os,argparse,re,codecs
from os import listdir
from os.path import isfile, join
# set the seed
random.seed(2823274491)
#filename_to_load = 'epoch_99-with-context/langtagger_model_embdim20_hiddim40_lyr2_e99_trainloss0.818105488535_trainprec99.4418968681_valprec97.6117062976.model'
filename_to_load = 'epoch_99/langtagger_model_embdim20_hiddim40_lyr2_e99_trainloss1.63745336175_trainprec99.9087639857_valprec98.594762855.model'
START_EPOCH = 0
# argument parse
parser = argparse.ArgumentParser()
parser.add_argument('-hiddim', '-hiddendim', help='Size of the RNN hidden layer, default 100', default=40,
required=False)
parser.add_argument('-embeddim', '-embeddingdim', help='Size of the embeddings, default 50', default=20, required=False)
parser.add_argument('-layers', '-mlplayers', help='Number of MLP layers, can only be 2 or 3', default=2, required=False)
parser.add_argument('-bilstmlayers', '-lstmlayers', help='Number of BILSTM layers, default 2', default=2,
required=False)
parser.add_argument('-model', '-modeltoload', help='Filename of model to load', default='', required=False)
args = vars(parser.parse_known_args()[0])
# get the params
HIDDEN_DIM = int(args['hiddim'])
EMBED_DIM = int(args['embeddim'])
BILSTM_LAYERS = int(args['bilstmlayers'])
fDo_3_Layers = int(args['layers']) == 3
sLAYERS = '3' if fDo_3_Layers else '2'
Filename_to_log = 'postagger_log_embdim' + str(EMBED_DIM) + '_hiddim' + str(HIDDEN_DIM) + '_lyr' + sLAYERS + '.txt'
def log_message(message):
print message
with codecs.open(Filename_to_log, "a", encoding="utf8") as myfile:
myfile.write("\n" + message)
if args['model']:
filename_to_load = args['model']
START_EPOCH = int(re.search("_e(\d+)", filename_to_load).group(1)) + 1
log_message('EMBED_DIM: ' + str(EMBED_DIM))
log_message('HIDDEN_DIM: ' + str(HIDDEN_DIM))
log_message('BILSTM_LAYERS: ' + str(BILSTM_LAYERS))
log_message('MLP Layers: ' + sLAYERS)
if filename_to_load:
log_message('Loading model: ' + filename_to_load)
log_message('Starting epoch: ' + str(START_EPOCH))
def read_data(dir=''):
if not dir:
dir = 'lstm_training.json'
training_set = json.load(codecs.open(dir, "rb", encoding="utf-8"))
tags = ['a','m']
training_set = [[(w['w'], tags.index(w['l'])) for w in sec] for sec in training_set ]
return training_set
# Classes:
# 1] Vocabulary class (the dictionary for char-to-int)
# 2] WordEncoder (actually, it'll be a char encoder)
# 3] Simple character BiLSTM
# 4] MLP
# 5] ConfusionMatrix
class Vocabulary(object):
def __init__(self):
self.all_items = []
self.c2i = {}
def add_text(self, paragraph):
self.all_items.extend(paragraph)
def finalize(self, fAddBOS=True):
self.vocab = sorted(list(set(self.all_items)))
c2i_start = 1 if fAddBOS else 0
self.c2i = {c: i for i, c in enumerate(self.vocab, c2i_start)}
self.i2c = self.vocab
if fAddBOS:
self.c2i['*BOS*'] = 0
self.i2c = ['*BOS*'] + self.vocab
self.all_items = None
# debug
def get_c2i(self):
return self.c2i
def size(self):
return len(self.i2c)
def __getitem__(self, c):
return self.c2i.get(c, 0)
def getItem(self, i):
return self.i2c[i]
class WordEncoder(object):
def __init__(self, name, dim, model, vocab):
self.vocab = vocab
self.enc = model.add_lookup_parameters((vocab.size(), dim))
def __call__(self, char, DIRECT_LOOKUP=False):
char_i = char if DIRECT_LOOKUP else self.vocab[char]
return dy.lookup(self.enc, char_i)
class MLP:
def __init__(self, model, name, in_dim, hidden_dim, out_dim):
self.mw = model.add_parameters((hidden_dim, in_dim))
self.mb = model.add_parameters((hidden_dim))
if not fDo_3_Layers:
self.mw2 = model.add_parameters((out_dim, hidden_dim))
self.mb2 = model.add_parameters((out_dim))
if fDo_3_Layers:
self.mw2 = model.add_parameters((hidden_dim, hidden_dim))
self.mb2 = model.add_parameters((hidden_dim))
self.mw3 = model.add_parameters((out_dim, hidden_dim))
self.mb3 = model.add_parameters((out_dim))
def __call__(self, x):
W = dy.parameter(self.mw)
b = dy.parameter(self.mb)
W2 = dy.parameter(self.mw2)
b2 = dy.parameter(self.mb2)
mlp_output = W2 * (dy.tanh(W * x + b)) + b2
if fDo_3_Layers:
W3 = dy.parameter(self.mw3)
b3 = dy.parameter(self.mb3)
mlp_output = W3 * (dy.tanh(mlp_output)) + b3
return dy.softmax(mlp_output)
class BILSTMTransducer:
def __init__(self, LSTM_LAYERS, IN_DIM, OUT_DIM, model):
self.lstmF = dy.LSTMBuilder(LSTM_LAYERS, IN_DIM, (int)(OUT_DIM / 2), model)
self.lstmB = dy.LSTMBuilder(LSTM_LAYERS, IN_DIM, (int)(OUT_DIM / 2), model)
def __call__(self, seq):
"""
seq is a list of vectors (either character embeddings or bilstm outputs)
"""
fw = self.lstmF.initial_state()
bw = self.lstmB.initial_state()
outf = fw.transduce(seq)
outb = list(reversed(bw.transduce(reversed(seq))))
return [dy.concatenate([f, b]) for f, b in zip(outf, outb)]
class ConfusionMatrix:
def __init__(self, size, vocab):
self.matrix = np.zeros((size, size))
self.size = size
self.vocab = vocab
def __call__(self, x, y):
self.matrix[x, y] += 1
def to_html(self):
fp_matrix = np.sum(self.matrix, 1)
fn_matrix = np.sum(self.matrix, 0)
html = """
<html>
<head>
<script src="https://ajax.googleapis.com/ajax/libs/jquery/1.12.4/jquery.min.js"></script>
<script src="confused.js"></script>
<style>.good{background-color:green;color:white}.bad{background-color:red;color:white}table{table-layout:fixed}td{text-align:center;padding:10px;border:solid 1px black}</style>
</head>
<body><h2>A Confusing Matrix</h2><table>"""
first_row = "<tr><td></td>"
for i in range(self.size):
first_row += "<td data-col-head={}>{}</td>".format(i, self.vocab[i])
first_row += "<td>False Positives</td></tr>"
html += first_row
for i in range(self.size):
html += "<tr><td data-row-head={}>{}</td>".format(i, self.vocab[i])
for j in range(self.size):
classy = "good" if i == j else "bad"
opacity = self.matrix[i, j] / (np.mean(self.matrix[self.matrix > 0]))
if opacity < 0.2: opacity = 0.2
if opacity > 1.0: opacity = 1.0
html += "<td data-i={} data-j={} class=\"{}\" style=\"opacity:{}\">{}</td>".format(i, j, classy,
opacity,
self.matrix[i, j])
html += "<td>{}</td></tr>".format(round(100.0 * (fp_matrix[i] - self.matrix[i, i]) / fp_matrix[i], 2))
# add confusion table for each class
stats = {"precision": self.precision, "recall": self.recall, "F1": self.f1}
html += "<tr><td>False Negatives</td>"
for i in range(self.size):
html += "<td>{}</td>".format(round(100.0 * (fn_matrix[i] - self.matrix[i, i]) / fn_matrix[i], 2))
html += "</tr>"
for k, v in stats.items():
html += "<tr><td>{}</td>".format(k)
for j in range(self.size):
tp = self.matrix[j, j]
fp = fp_matrix[j] - tp
fn = fn_matrix[j] - tp
html += "<td>{}</td>".format(round(100 * v(tp, fp, fn), 2))
html += "</tr>"
html += "</table><h2>Table of Confusion</h2>"
total_tp = sum([self.matrix[i, i] for i in range(self.size)])
total_fp = np.sum(fp_matrix) - total_tp
total_fn = np.sum(fn_matrix) - total_tp
html += "<h3>Precision: {}</h3>".format(round(100 * self.precision(total_tp, total_fp, total_fn), 2))
html += "<h3>Recall: {}</h3>".format(round(100 * self.recall(total_tp, total_fp, total_fn), 2))
html += "<h3>F1: {}</h3>".format(round(100 * self.f1(total_tp, total_fp, total_fn), 2))
html += "</body></html>"
return html
def f1(self, tp, fp, fn):
return 2.0 * tp / (2.0 * tp + fp + fn) if tp + fp + fn != 0 else 0.0
def recall(self, tp, fp, fn):
return 1.0 * tp / (tp + fn) if tp + fn != 0 else 0.0
def precision(self, tp, fp, fn):
return 1.0 * tp / (tp + fp) if tp + fn != 0 else 0.0
def clear(self):
self.matrix = np.zeros((self.size, self.size))
def CalculateLossForDaf(daf, fValidation=False, fRunning=False):
dy.renew_cg()
tagged_daf = {"words": []}
# add a bos before and after
seq = ['*BOS*'] + list(' '.join([word for word, _ in daf])) + ['*BOS*']
# get all the char encodings for the daf
char_embeds = [let_enc(let) for let in seq]
# run it through the bilstm
char_bilstm_outputs = bilstm(char_embeds)
# now iterate and get all the separate word representations by concatenating the bilstm output
# before and after the word
word_bilstm_outputs = []
iLet_start = 0
for iLet, char in enumerate(seq):
# if it is a bos, check if it's at the end of the sequence
if char == '*BOS*':
if iLet + 1 == len(seq):
char = ' '
else:
continue
# if we are at a space, take this bilstm output and the one at the letter start
if char == ' ':
cur_word_bilstm_output = dy.concatenate([char_bilstm_outputs[iLet_start], char_bilstm_outputs[iLet]])
# add it in
word_bilstm_outputs.append(cur_word_bilstm_output)
# set the iLet_start ocunter to here
iLet_start = iLet
# safe-check, make sure word bilstm outputs length is the same as the daf
if len(word_bilstm_outputs) != len(daf):
log_message('Size mismatch!! word_bilstm_outputs: ' + str(len(word_bilstm_outputs)) + ', daf: ' + str(len(daf)))
prev_lang_lstm_state = prev_lang_lstm.initial_state().add_input(lang_enc('*BOS*'))
all_losses = []
lang_prec = 0.0
lang_items = 0
# now iterate through the bilstm outputs, and each word in the daf
for (word, gold_word_lang), bilstm_output in zip(daf, word_bilstm_outputs):
# create the mlp input, a concatenate of the bilstm output and of the prev pos output
mlp_input = dy.concatenate([bilstm_output, prev_lang_lstm_state.output()])
# run through the class mlp
lang_mlp_output = lang_mlp(mlp_input)
predicted_word_lang = lang_vocab.getItem(np.argmax(lang_mlp_output.npvalue()))
confidence = np.max(lang_mlp_output.npvalue()) / np.sum(lang_mlp_output.npvalue())
lang_prec += 1 if predicted_word_lang == gold_word_lang else 0
lang_items += 1
tagged_daf["words"].append(
{"word": word, "predicted_lang": predicted_word_lang, "confidence": confidence})
# if we aren't doing validation, calculate the loss
if not fValidation and not fRunning:
all_losses.append(-dy.log(dy.pick(lang_mlp_output, lang_vocab[gold_word_lang])))
word_pos_ans = gold_word_lang
# otherwise, set the answer to be the argmax
elif not fRunning and fValidation:
lang_conf_matrix(lang_vocab[predicted_word_lang], lang_vocab[gold_word_lang])
word_pos_ans = predicted_word_lang
else:
continue
# run through the prev-pos-mlp
prev_lang_lstm_state = prev_lang_lstm_state.add_input(lang_enc(word_pos_ans))
# prev_pos_lstm_state = prev_pos_lstm_state.add_input(pos_enc(''))
lang_prec = lang_prec / lang_items if lang_items > 0 else None
# class_prec = class_prec / class_items if class_items > 0 else None
if fValidation:
return lang_prec, tagged_daf
if fRunning:
return tagged_daf
total_loss = dy.esum(all_losses) if len(all_losses) > 0 else None
return total_loss, lang_prec
def run_network_on_validation(epoch_num):
val_lang_prec = 0.0
val_lang_items = 0
# iterate
num_daf_to_save = 10
daf_to_save = []
for idaf, word in enumerate(val_data):
lang_prec, tagged_daf = CalculateLossForDaf(word, fValidation=True)
# increment and continue
val_lang_prec += lang_prec
val_lang_items += 1
if epoch_num >= 0 and idaf % round(1.0 * len(val_data) / num_daf_to_save) == 0:
daf_to_save.extend(tagged_daf)
# divide
val_lang_prec = val_lang_prec / val_lang_items * 100 if val_lang_items > 0 else 0.0
# print the results
log_message('Validation: pos_prec: ' + str(val_lang_prec))
objStr = json.dumps(daf_to_save, indent=4, ensure_ascii=False)
if not os.path.exists('epoch_{}'.format(epoch_num)):
os.makedirs('epoch_{}'.format(epoch_num))
with open("epoch_{}/tagged.json".format(epoch_num), "w") as f:
f.write(objStr.encode('utf-8'))
return val_lang_prec
def print_tagged_corpus_to_html_table(lang_out):
str = u"<html><head><style>h1{text-align:center;background:grey}td{text-align:center}table{margin-top:20px;margin-bottom:20px;margin-right:auto;margin-left:auto;width:1200px}.aramaic{background-color:blue;color:white}.mishnaic{background-color:red;color:white}.ambiguous{background-color:yellow;color:black}</style><meta charset='utf-8'></head><body>"
for daf in lang_out:
str += u"<h1>DAF {}</h1>".format(daf)
str += u"<table>"
count = 0
while count < len(lang_out[daf]['words']):
row_obj = lang_out[daf]['words'][count:count+10]
row = u"<tr>"
for w in reversed(row_obj):
row += u"<td class='{}'>{}</td>".format('aramaic' if w['predicted_lang'] == 0 else 'mishnaic',w['word'])
row += u"</tr>"
#row_sef += u"<td>({}-{})</td></tr>".format(count,count+len(row_obj)-1)
str += row
count += 10
str += u"</table>"
str += u"</body></html>"
return str
# read in all the data
all_data = list(read_data())
random.shuffle(all_data)
# train val will be split up 100-780
percent_training = 0.2
split_index = int(round(len(all_data) * percent_training))
train_data = all_data[split_index:]
val_data = all_data[:split_index]
# create the vocabulary
let_vocab = Vocabulary()
lang_vocab = Vocabulary()
lang_tags = ['a','m']
# iterate through all the dapim and put everything in the vocabulary
for sec in all_data:
let_vocab.add_text([c for w,_ in sec for c in w])
let_vocab.add_text([u' '])
lang_vocab.add_text([l for _,l in sec])
let_vocab.finalize()
lang_vocab.finalize(False)
lang_conf_matrix = ConfusionMatrix(len(lang_tags), lang_tags)
log_message('pos: ' + str(len(lang_tags)))
log_message('let: ' + str(let_vocab.size()))
# debug - write out the vocabularies
# write out to files the pos vocab and the letter vocab
with codecs.open('let_vocab.txt', 'w', encoding='utf8') as f:
for let, id in let_vocab.get_c2i().items():
f.write(str(id) + ' : ' + let + '\n')
with codecs.open('lang_vocab.txt', 'w', encoding='utf8') as f:
for lang, id in lang_vocab.get_c2i().items():
f.write(str(id) + ' : ' + str(lang) + '\n')
# to save on memory space, we will clear out all_data from memory
all_data = None
# create the model and all it's parameters
model = dy.Model()
# create the word encoders (an encoder for the chars for the bilstm, and an encoder for the prev-pos lstm)
let_enc = WordEncoder("letenc", EMBED_DIM, model, let_vocab)
lang_enc = WordEncoder("langenc", EMBED_DIM, model, lang_vocab)
# the BiLSTM for all the chars, take input of embed dim, and output of the hidden_dim minus the embed_dim because we will concatenate
# with output from a separate bilstm of just the word
bilstm = BILSTMTransducer(BILSTM_LAYERS, EMBED_DIM, HIDDEN_DIM, model)
prev_lang_lstm = dy.LSTMBuilder(BILSTM_LAYERS, EMBED_DIM, EMBED_DIM, model)
# now the class mlp, it will take input of 2*HIDDEN_DIM (A concatenate of the before and after the word) + EMBED_DIM from the prev-pos
# output of 2, unknown\talmud
lang_mlp = MLP(model, "classmlp", 2 * HIDDEN_DIM + EMBED_DIM, HIDDEN_DIM, 2)
# the trainer
trainer = dy.AdamTrainer(model)
# if we are loading in a model
if filename_to_load:
model.load(filename_to_load)
train_test = False
if train_test:
run_network_on_validation(START_EPOCH - 1)
lang_conf_matrix.clear()
# train!
for epoch in range(START_EPOCH, 100):
last_loss, last_lang_prec = 0.0, 0.0
total_loss, total_lang_prec = 0.0, 0.0
total_lang_items = 0
# shuffle the train data
random.shuffle(train_data)
items_seen = 0
# iterate
for daf in train_data:
# calculate the loss & prec
loss, lang_prec = CalculateLossForDaf(daf, fValidation=False)
# forward propagate
total_loss += loss.value() if loss else 0.0
# back propagate
if loss: loss.backward()
trainer.update()
# increment the prec variable
total_lang_prec += lang_prec
total_lang_items += 1
items_seen += 1
# breakpoint?
breakpoint = 100
if items_seen % breakpoint == 0:
last_loss = total_loss / breakpoint
last_lang_prec = total_lang_prec / total_lang_items * 100
log_message("Words processed: " + str(items_seen) + ", loss: " + str(last_loss) + ', lang_prec: ' + str(
last_lang_prec))
total_loss, total_lang_prec = 0.0, 0.0
total_lang_items = 0
log_message('Finished epoch ' + str(epoch))
val_lang_prec = run_network_on_validation(epoch)
if not os.path.exists('epoch_{}'.format(epoch)):
os.makedirs('epoch_{}'.format(epoch))
filename_to_save = 'epoch_' + str(epoch) + '/langtagger_model_embdim' + str(EMBED_DIM) + '_hiddim' + str(
HIDDEN_DIM) + '_lyr' + sLAYERS + '_e' + str(epoch)
filename_to_save += '_trainloss' + str(last_loss) + '_trainprec' + str(last_lang_prec) + '_valprec' + str(
val_lang_prec) + '.model'
model.save(filename_to_save)
f = open("epoch_{}/conf_matrix_e{}.html".format(epoch, epoch), 'w')
f.write(lang_conf_matrix.to_html())
f.close()
lang_conf_matrix.clear()
else:
#tag all of shas!
cal_matcher_path = '../../../dibur_hamatchil/dh_source_scripts/cal_matcher_output'
mesechtot_names = ['Berakhot','Shabbat','Eruvin','Pesachim']
for mesechta in mesechtot_names:
mesechta_path = '{}/{}/lang_naive_talmud'.format(cal_matcher_path,mesechta)
if not os.path.exists('{}/{}/lang_tagged_context'.format(cal_matcher_path, mesechta)):
os.makedirs('{}/{}/lang_tagged_context'.format(cal_matcher_path, mesechta))
if not os.path.exists('{}/{}/html_lang_tagged_context'.format(cal_matcher_path, mesechta)):
os.makedirs('{}/{}/html_lang_tagged_context'.format(cal_matcher_path, mesechta))
def sortdaf(fname):
daf = fname.split('lang_naive_talmud_')[1].split('.json')[0]
daf_int = int(daf[:-1])
amud_int = 1 if daf[-1] == 'b' else 0
return daf_int*2 + amud_int
files = [f for f in listdir(mesechta_path) if isfile(join(mesechta_path, f))]
files.sort(key=sortdaf)
html_out = OrderedDict()
for i_f,f_name in enumerate(files):
cal_matcher_out = json.load(codecs.open('{}/{}'.format(mesechta_path,f_name), "rb", encoding="utf-8"))
daf = cal_matcher_out['words']
daf = [(w['word'],'') for w in daf] #prepare for CalculateLossForDaf
lang_out = CalculateLossForDaf(daf, fRunning=True)
fp = codecs.open("{}/{}/lang_tagged_context/{}.json".format(cal_matcher_path,mesechta,f_name), "wb", encoding='utf-8')
json.dump(lang_out, fp, indent=4, encoding='utf-8', ensure_ascii=False)
fp.close()
daf = f_name.split('lang_naive_talmud_')[1].split('.json')[0]
html_out[daf] = lang_out
if i_f % 10 == 0:
print '{}/{}'.format(mesechta,f_name)
html = print_tagged_corpus_to_html_table(html_out)
fp = codecs.open("{}/{}/html_lang_tagged_context/{}.html".format(cal_matcher_path, mesechta, daf), "wb",
encoding='utf-8')
fp.write(html)
fp.close()
html_out = OrderedDict()