-
Notifications
You must be signed in to change notification settings - Fork 25
/
util.py
524 lines (380 loc) · 17.4 KB
/
util.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
import codecs
import operator
import numpy as np
import random
import math
import codecs
import sys
from collections import defaultdict
def load_vocab(corpus, word_minfreq, dummy_symbols):
idxword, idxchar = [], []
wordxid, charxid = defaultdict(int), defaultdict(int)
word_freq, char_freq = defaultdict(int), defaultdict(int)
wordxchar = defaultdict(list)
def update_dic(symbol, idxvocab, vocabxid):
if symbol not in vocabxid:
idxvocab.append(symbol)
vocabxid[symbol] = len(idxvocab) - 1
for line_id, line in enumerate(codecs.open(corpus, "r", "utf-8")):
for word in line.strip().split():
word_freq[word] += 1
for char in line.strip():
char_freq[char] += 1
#add in dummy symbols into dictionaries
for s in dummy_symbols:
update_dic(s, idxword, wordxid)
update_dic(s, idxchar, charxid)
#remove low fequency words/chars
def collect_vocab(vocab_freq, idxvocab, vocabxid):
for w, f in sorted(vocab_freq.items(), key=operator.itemgetter(1), reverse=True):
if f < word_minfreq:
break
else:
update_dic(w, idxvocab, vocabxid)
collect_vocab(word_freq, idxword, wordxid)
collect_vocab(char_freq, idxchar, charxid)
#word id to [char ids]
dummy_symbols_set = set(dummy_symbols)
for wi, w in enumerate(idxword):
if w in dummy_symbols:
wordxchar[wi] = [wi]
else:
for c in w:
wordxchar[wi].append(charxid[c] if c in charxid else charxid[dummy_symbols[2]])
return idxword, wordxid, idxchar, charxid, wordxchar
def only_symbol(word):
for c in word:
if c.isalpha():
return False
return True
def remove_punct(string):
return " ".join("".join([ item for item in string if (item.isalpha() or item == " ") ]).split())
def load_data(corpus, wordxid, idxword, charxid, idxchar, (pad_symbol, end_symbol, unk_symbol)):
nwords = [] #number of words for each line
nchars = [] #number of chars for each line
word_data = [] #data[doc_id][0][line_id] = list of word ids; data[doc_id][1][line_id] = list of [char_ids]
char_data = [] #data[line_id] = list of char ids
rhyme_data = [] #list of ( target_word, [candidate_words], target_word_line_id ); word is a list of characters
def word_to_char(word):
if word in set([pad_symbol, end_symbol, unk_symbol]):
return [ wordxid[word] ]
else:
return [ charxid[item] if item in charxid else charxid[unk_symbol] for item in word ]
for doc in codecs.open(corpus, "r", "utf-8"):
word_lines, char_lines = [[], []], []
last_words = []
#reverse the order of lines and words as we are generating from end to start
for line in reversed(doc.strip().split(end_symbol)):
if len(line.strip()) > 0:
word_seq = [ wordxid[item] if item in wordxid else wordxid[unk_symbol] \
for item in reversed(line.strip().split()) ] + [wordxid[end_symbol]]
char_seq = [ word_to_char(item) for item in reversed(line.strip().split()) ] + [word_to_char(end_symbol)]
word_lines[0].append(word_seq)
word_lines[1].append(char_seq)
char_lines.append([ charxid[item] if item in charxid else charxid[unk_symbol] \
for item in remove_punct(line.strip())])
nwords.append(len(word_lines[0][-1]))
nchars.append(len(char_lines[-1]))
last_words.append(line.strip().split()[-1])
if len(word_lines[0]) == 14: #14 lines for sonnets
word_data.append(word_lines)
char_data.extend(char_lines)
last_words = last_words[2:] #remove couplets (since they don't always rhyme)
for wi, w in enumerate(last_words):
rhyme_data.append( (word_to_char(w), [ word_to_char(item)
for item_id, item in enumerate(last_words[(wi/4)*4:(wi/4+1)*4]) if item_id != (wi%4) ], (11-wi)) )
return word_data, char_data, rhyme_data, nwords, nchars
def print_stats(partition, word_data, rhyme_data, nwords, nchars):
print partition, "statistics:"
print " Number of documents =", len(word_data)
print " Number of rhyme examples =", len(rhyme_data)
print " Total number of word tokens =", sum(nwords)
print " Mean/min/max words per line = %.2f/%d/%d" % (np.mean(nwords), min(nwords), max(nwords))
print " Total number of char tokens =", sum(nchars)
print " Mean/min/max chars per line = %.2f/%d/%d" % (np.mean(nchars), min(nchars), max(nchars))
def init_embedding(model, idxword):
word_emb = []
for vi, v in enumerate(idxword):
if v in model:
word_emb.append(model[v])
else:
word_emb.append(np.random.uniform(-0.5/model.vector_size, 0.5/model.vector_size, [model.vector_size,]))
return np.array(word_emb)
def pad(lst, max_len, pad_symbol):
if len(lst) > max_len:
print "\nERROR: padding"
print "length of list greater than maxlen; list =", lst, "; maxlen =", max_len
raise SystemExit
return lst + [pad_symbol] * (max_len - len(lst))
def get_vowels():
return set(["a", "e", "i", "o", "u"])
def coverage_mask(char_ids, idxchar):
vowels = get_vowels()
return [ float(idxchar[c] in vowels) for c in char_ids ]
def flatten_list(l):
return [item for sublist in l for item in sublist]
def create_word_batch(data, batch_size, lines_per_doc, nlines_per_batch, pad_symbol, end_symbol, unk_symbol, shuffle_data):
docs_per_batch = len(data) / batch_size
batches = []
doc_ids = range(len(data))
if shuffle_data:
random.shuffle(doc_ids)
if lines_per_doc % nlines_per_batch != 0:
print "\nERROR:"
print "lines_per_doc (%d) %% nlines_per_batch (%d) must equal 0" % (lines_per_doc, nlines_per_batch)
raise SystemExit
for i in range(docs_per_batch):
for j in range(lines_per_doc / nlines_per_batch):
docs = []
doc_lens = []
doc_lines = []
x = []
y = []
xchar = []
xchar_lens = []
hist = []
hist_lens = []
for k in range(batch_size):
d = doc_ids[i*batch_size+k]
wordseq = flatten_list(data[d][0][j*nlines_per_batch:(j+1)*nlines_per_batch])
charseq = flatten_list(data[d][1][j*nlines_per_batch:(j+1)*nlines_per_batch])
histseq = flatten_list(data[d][0][:j*nlines_per_batch])
x.append([end_symbol] + wordseq[:-1])
y.append(wordseq)
docs.append(d)
doc_lens.append(len(wordseq))
doc_lines.append(range(j*nlines_per_batch, (j+1)*nlines_per_batch))
xchar.append([[end_symbol]] + charseq[:-1])
xchar_lens.append( [1] + [ len(item) for item in charseq[:-1] ])
hist.append(histseq if len(histseq) > 0 else [unk_symbol])
hist_lens.append(len(histseq) if len(histseq) > 0 else 1)
#pad the data
word_pad_len = max(doc_lens)
char_pad_len = max(flatten_list(xchar_lens))
hist_pad_len = max(hist_lens)
for k in range(batch_size):
x[k] = pad(x[k], word_pad_len, pad_symbol)
y[k] = pad(y[k], word_pad_len, pad_symbol)
xchar_lens[k].extend( [1]*(word_pad_len-len(xchar[k])) ) #add len for pad symbols
xchar[k] = pad(xchar[k], word_pad_len, [pad_symbol]) #pad the word lengths
xchar[k] = [pad(item, char_pad_len, pad_symbol) for item in xchar[k]] #pad the characters
hist[k] = pad(hist[k], hist_pad_len, pad_symbol)
batches.append((x, y, docs, doc_lens, doc_lines, xchar, xchar_lens, hist, hist_lens))
return batches
def create_char_batch(data, batch_size, pad_symbol, pentameter, idxchar, shuffle_data):
batches = []
batch_len = len(data) / batch_size
if shuffle_data:
random.shuffle(data)
for i in range(batch_len):
enc_x = []
enc_xlen = []
for j in range(batch_size):
enc_x.append(data[i*batch_size+j])
enc_xlen.append(len(data[i*batch_size+j]))
xlen_max = max(enc_xlen)
cov_mask = np.zeros((batch_size, xlen_max)) #coverage mask
for j in range(batch_size):
enc_x[j] = pad(enc_x[j], xlen_max, pad_symbol)
cov_mask[j] = coverage_mask(enc_x[j], idxchar)
batches.append((enc_x, enc_xlen, cov_mask))
return batches
def create_rhyme_batch(data, batch_size, pad_symbol, wordxchar, num_neg, shuffle_data):
if shuffle_data:
random.shuffle(data)
batches = []
for i in range(len(data) / batch_size):
x, xid, c = [], [], []
xlen, clen = [], []
for j in range(batch_size):
x.append(data[i*batch_size+j][0])
xid.append(data[i*batch_size+j][2])
xlen.append(len(data[i*batch_size+j][0]))
for context in data[i*batch_size+j][1]:
c.append(context)
clen.append(len(context))
for _ in range(num_neg):
c.append(wordxchar[random.randrange(3, len(wordxchar))])
clen.append(len(c[-1]))
#merging target and context words
# (first batch_size = target words; following batch_size*(3+num_neg) = context words)
xc = x + c
xclen = xlen + clen
xclen_max = max(xclen)
#pad the target words and context words
for xci, xcv in enumerate(xc):
xc[xci] = pad(xcv, xclen_max, pad_symbol)
batches.append((xc, xclen, xid))
return batches
def print_lm_attention(bi, b, attentions, idxword, cf):
print "\n", "="*100
for ex in range(cf.batch_size)[-1:]:
xword = [ idxword[item] for item in b[1][ex] ]
hword = [ idxword[item] for item in b[7][ex] ]
print "\nBatch ID =", bi
print "Example =", ex
print "x_word =", " ".join(xword)
print "hist_word=", " ".join(hword)
for xi, x in enumerate(xword):
print "\nWord =", x
print "\tSum dist =", sum(attentions[ex][xi])
attn_dist_sort = np.argsort(-attentions[ex][xi])
print "\t",
for hi in attn_dist_sort[:5]:
print ("[%d]%s:%.3f " % (hi, hword[hi], attentions[ex][xi][hi])),
print
def print_pm_attention(b, batch_size, costs, logits, attentions, mius, idxchar):
print "\n", "="*100
for ex in range(batch_size)[-10:]:
print "\nSentence =", ex
print "x =", b[0][ex]
print "x len =", b[1][ex]
print "x char=", "".join(idxchar[item] for item in b[0][ex])
print "losses =", costs[ex]
print "pentameter output =", logits[ex]
print "coverage mask =", b[2][ex]
for attni, attn in enumerate(attentions):
print "attention at time step", attni, ":"
print "\tmiu_p =", mius[attni][ex] * (b[1][ex] - 1.0)
for xid in reversed(np.argsort(attn[ex])):
if attn[ex][xid] > 0.05:
print "\t%.3f %d %s" % (attn[ex][xid], xid, (idxchar[b[0][ex][xid]]))
def print_rm_attention(b, batch_size, num_context, attentions, pad_id, idxchar):
print "\n", "="*100
for exid in range(batch_size)[-10:]:
print "\nTarget word =", "".join([idxchar[item] for item in b[0][exid] if item != pad_id])
for ci, c in enumerate(b[0][(exid*num_context+batch_size):((exid+1)*num_context+batch_size)]):
print "\t", ("%.2f" % attentions[exid][ci]), "=", "".join([idxchar[item] for item in c if item != pad_id])
def get_word_stress(cmu, word):
stresses = set([])
def valid(stress):
for sti in range(len(stress)-1):
if abs(int(stress[sti]) - int(stress[sti+1])) != 1:
return False
return True
if word in cmu:
for res in cmu[word]:
stress = ""
for syl in res:
if syl[-1] == "0":
stress += "0"
elif syl[-1] == "1" or syl[-1] == "2":
stress += "1"
if valid(stress):
stresses.add(stress)
return stresses
def update_stress_accs(accs, word_len, score):
word_buckets = [4,8,float("inf")]
for wi, wb in enumerate(word_buckets):
if word_len <= wb:
accs[wi].append(score)
break
def eval_stress(accs, cmu, attns, pentameter, batch, idxchar, charxid, pad_symbol, cf):
attn_threshold = 0.2
for ex in range(cf.batch_size):
chars = [idxchar[item] for item in batch[ex]]
space_ids = [-1] + [i for i, ch in enumerate(chars) if ch == " "] + \
[batch[ex].index(charxid[pad_symbol]) if charxid[pad_symbol] in batch[ex] else len(batch[ex])]
for spi, sp in enumerate(space_ids[:-1]):
start = sp+1
end = space_ids[spi+1]
word = "".join(chars[start:end])
gold_stress = get_word_stress(cmu, word)
sys_stress = ""
if len(gold_stress) == 0:
continue
for attni, attn in enumerate(attns):
for ch in range(start, end):
if attn[ex][ch] >= attn_threshold:
sys_stress += str(pentameter[attni])
break
update_stress_accs(accs, (end-start), float(sys_stress in gold_stress))
def eval_rhyme(pr, thresholds, cmu, attns, b, idxchar, charxid, pad_symbol, cf, cmu_rhyme=None, cmu_norhyme=None,
em_vocab=None, em_theta=None):
def syllable_to_rhyme(syllable):
stresses = set(["0", "1", "2"])
r = []
for s in reversed(syllable):
if s[-1] in stresses:
r.append(s[:-1])
break
else:
r.append(s)
return "-".join(list(reversed(r)))
def get_rhyme(words):
rhymes = []
for word in words:
word_rhymes = set([])
if word in cmu:
for res in cmu[word]:
word_rhymes.add(syllable_to_rhyme(res))
rhymes.append(word_rhymes)
return rhymes
def rhyme_score(target_rhymes, context_rhymes):
if len(target_rhymes) == 0 or len(context_rhymes) == 0:
return None
else:
for tr in target_rhymes:
if tr in context_rhymes:
return 1.0
return 0.0
def last_syllable(word):
i = len(word)
for c in reversed(word):
i -= 1
if c in get_vowels():
break
return word[i:]
def em_rhyme_score(x, y):
if x in em_vocab and y in em_vocab:
xi = em_vocab.index(x)
yi = em_vocab.index(y)
return max(em_theta[xi][yi], em_theta[yi][xi])
return 0.0
num_c = 3 + cf.rm_neg
for ex in range(cf.batch_size):
target = "".join([idxchar[item] for item in b[0][ex][:b[1][ex]]])
context = []
for ci, c in enumerate(b[0][(ex*num_c+cf.batch_size):(ex*num_c+cf.batch_size+3)]):
context.append("".join([idxchar[item] for item in c[:b[1][ex*num_c+cf.batch_size+ci]]]))
target_rhyme = get_rhyme([target])[0]
context_rhyme = get_rhyme(context)
for t in thresholds:
for ci, c in enumerate(context):
score = rhyme_score(target_rhyme, context_rhyme[ci])
system_score = attns[ex][ci] if em_vocab == None else em_rhyme_score(target, context[ci])
#precision
if system_score >= t:
if score != None:
pr[t][0].append(score)
#recall
if score == 1.0:
pr[t][1].append(float(system_score >= t))
if cmu_rhyme != None and cmu_norhyme != None:
if score == 1.0:
if (target, context[ci]) not in cmu_rhyme and (context[ci], target) not in cmu_rhyme:
cmu_rhyme[(target, context[ci])] = system_score
elif score == 0.0:
if (target, context[ci]) not in cmu_norhyme and (context[ci], target) not in cmu_norhyme:
cmu_norhyme[(target, context[ci])] = system_score
def collect_rhyme_pattern(rhyme_pattern, attentions, b, batch_size, num_context, idxchar, pad_id):
def get_context_line_id(target_line_id, context_id):
p = target_line_id % 4
q = 3-context_id
if q <= p:
q -= 1
return q
#print "\n", "="*100
for exid in range(batch_size):
target_line_id = b[2][exid]
for ci, c in enumerate(b[0][(exid*num_context+batch_size):((exid+1)*num_context+batch_size)][:3]):
context_line_id = get_context_line_id(target_line_id, ci)
rhyme_pattern[target_line_id][context_line_id].append(attentions[exid][ci])
def postprocess_sentence(line):
cleaned_sent = ""
for w in line.strip().split():
spacing = " "
if w.startswith("'") or only_symbol(w):
spacing = ""
cleaned_sent += spacing + w
return cleaned_sent.strip()