This repository has been archived by the owner on Jun 25, 2024. It is now read-only.
forked from EdinburghNLP/Refresh
-
Notifications
You must be signed in to change notification settings - Fork 1
/
data_utils.py
356 lines (286 loc) · 15.7 KB
/
data_utils.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
####################################
# Author: Shashi Narayan
# Date: September 2016
# Project: Document Summarization
# H2020 Summa Project
####################################
"""
Document Summarization Modules and Models
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import tensorflow as tf
import random
import os
from my_flags import FLAGS
from model_utils import convert_logits_to_softmax, predict_topranked
# Special IDs
PAD_ID = 0
UNK_ID = 1
class Data:
def __init__(self, vocab_dict, data_type):
self.filenames = []
self.docs = []
self.titles = []
self.images = []
self.labels = []
self.rewards = []
self.weights = []
self.fileindices = []
self.data_type = data_type
# populate the data
self.populate_data(vocab_dict, data_type)
# Write to files
self.write_to_files(data_type)
def write_prediction_summaries(self, pred_logits, modelname, session=None):
print("Writing predictions and final summaries ...")
# Convert to softmax logits
pred_logits = convert_logits_to_softmax(pred_logits, session=session)
# Save Output Logits
np.save(FLAGS.train_dir+"/"+modelname+"."+self.data_type+"-prediction", pred_logits)
# Writing
pred_labels = predict_topranked(pred_logits, self.weights, self.filenames)
self.write_predictions(modelname+"."+self.data_type, pred_logits, pred_labels)
self.process_predictions_topranked(modelname+"."+self.data_type)
def write_predictions(self, file_prefix, np_predictions, np_labels):
foutput = open(FLAGS.train_dir+"/"+file_prefix+".predictions", "w")
for fileindex in self.fileindices:
filename = self.filenames[fileindex]
foutput.write(filename+"\n")
sentcount = 0
for sentpred, sentlabel in zip(np_predictions[fileindex], np_labels[fileindex]):
one_prob = sentpred[0]
label = sentlabel[0]
if sentcount < len(self.weights[fileindex]):
foutput.write(str(int(label))+"\t"+str(one_prob)+"\n")
else:
break
sentcount += 1
foutput.write("\n")
foutput.close()
def process_predictions_topranked(self, file_prefix):
predictiondata = open(FLAGS.train_dir+"/"+file_prefix+".predictions").read().strip().split("\n\n")
# print len(predictiondata)
summary_dirname = FLAGS.train_dir+"/"+file_prefix+"-summary-topranked"
os.system("mkdir "+summary_dirname)
for item in predictiondata:
# print(item)
itemdata = item.strip().split("\n")
# print len(itemdata)
filename = itemdata[0]
# print filename
# predictions file already have top three sentences marked
final_sentids = []
for sentid in range(len(itemdata[1:])):
label_score = itemdata[sentid+1].split()
if label_score[0] == "1":
final_sentids.append(sentid)
# Create final summary files
fileid = filename.split("-")[-1] # cnn-fileid, dailymail-fileid
summary_file = open(summary_dirname+"/"+fileid+".model", "w")
# Read Sents in the document : Always use original sentences
sent_filename = FLAGS.doc_sentence_directory + "/" + self.data_type +"/mainbody/"+fileid+".mainbody"
docsents = open(sent_filename).readlines()
# Top Ranked three sentences
selected_sents = [docsents[sentid] for sentid in final_sentids if sentid < len(docsents)]
# print(selected_sents)
summary_file.write("".join(selected_sents)+"\n")
summary_file.close()
def get_batch(self, startidx, endidx):
# This is very fast if you keep everything in Numpy
def process_to_chop_pad(orgids, requiredsize):
if (len(orgids) >= requiredsize):
return orgids[:requiredsize]
else:
padids = [PAD_ID] * (requiredsize - len(orgids))
return (orgids + padids)
# Numpy dtype
dtype = np.float16 if FLAGS.use_fp16 else np.float32
# For train, (endidx-startidx)=FLAGS.batch_size, for others its as specified
batch_docnames = np.empty((endidx-startidx), dtype="S60") # File ID of size "cnn-" or "dailymail-" with fileid of size 40
batch_docs = np.empty(((endidx-startidx), (FLAGS.max_doc_length + FLAGS.max_title_length + FLAGS.max_image_length), FLAGS.max_sent_length), dtype="int32")
batch_label = np.empty(((endidx-startidx), FLAGS.max_doc_length, FLAGS.target_label_size), dtype=dtype) # Single best oracle, used for JP models or accuracy estimation
batch_weight = np.empty(((endidx-startidx), FLAGS.max_doc_length), dtype=dtype)
batch_oracle_multiple = np.empty(((endidx-startidx), 1, FLAGS.max_doc_length, FLAGS.target_label_size), dtype=dtype)
batch_reward_multiple = np.empty(((endidx-startidx), 1), dtype=dtype)
batch_idx = 0
for fileindex in self.fileindices[startidx:endidx]:
# Document Names
batch_docnames[batch_idx] = self.filenames[fileindex]
# Document
doc_wordids = [] # [FLAGS.max_doc_length + FLAGS.max_title_length + FLAGS.max_image_length, FLAGS.max_sent_length]
for idx in range(FLAGS.max_doc_length):
thissent = []
if idx < len(self.docs[fileindex]):
thissent = self.docs[fileindex][idx][:]
thissent = process_to_chop_pad(thissent, FLAGS.max_sent_length) # [FLAGS.max_sent_length]
doc_wordids.append(thissent)
for idx in range(FLAGS.max_title_length):
thissent = []
if idx < len(self.titles[fileindex]):
thissent = self.titles[fileindex][idx][:]
thissent = process_to_chop_pad(thissent, FLAGS.max_sent_length) # [FLAGS.max_sent_length]
doc_wordids.append(thissent)
for idx in range(FLAGS.max_image_length):
thissent = []
if idx < len(self.images[fileindex]):
thissent = self.images[fileindex][idx][:]
thissent = process_to_chop_pad(thissent, FLAGS.max_sent_length) # [FLAGS.max_sent_length]
doc_wordids.append(thissent)
batch_docs[batch_idx] = np.array(doc_wordids[:], dtype="int32")
# Labels: Select the single best
labels_vecs = [[1, 0] if (item in self.labels[fileindex][0]) else [0, 1] for item in range(FLAGS.max_doc_length)]
batch_label[batch_idx] = np.array(labels_vecs[:], dtype=dtype)
# Weights
weights = process_to_chop_pad(self.weights[fileindex][:], FLAGS.max_doc_length)
batch_weight[batch_idx] = np.array(weights[:], dtype=dtype)
# Multiple Labels and rewards
labels_set = [] # FLAGS.num_sample_rollout, FLAGS.max_doc_length, FLAGS.target_label_size
reward_set = [] # FLAGS.num_sample_rollout, FLAGS.max_doc_length, FLAGS.target_label_size
for idx in range(FLAGS.num_sample_rollout):
thislabels = []
if idx < len(self.labels[fileindex]):
thislabels = [[1, 0] if (item in self.labels[fileindex][idx]) else [0, 1] for item in range(FLAGS.max_doc_length)]
reward_set.append(self.rewards[fileindex][idx])
else:
# Simply copy the best one
thislabels = [[1, 0] if (item in self.labels[fileindex][0]) else [0, 1] for item in range(FLAGS.max_doc_length)]
reward_set.append(self.rewards[fileindex][0])
labels_set.append(thislabels)
# Randomly Sample one oracle label
randidx_oracle = random.randint(0, (FLAGS.num_sample_rollout-1))
batch_oracle_multiple[batch_idx][0] = np.array(labels_set[randidx_oracle][:], dtype=dtype)
batch_reward_multiple[batch_idx] = np.array([reward_set[randidx_oracle]], dtype=dtype)
# increase batch count
batch_idx += 1
return batch_docnames, batch_docs, batch_label, batch_weight, batch_oracle_multiple, batch_reward_multiple
def shuffle_fileindices(self):
random.shuffle(self.fileindices)
def write_to_files(self, data_type):
full_data_file_prefix = FLAGS.train_dir + "/" + FLAGS.data_mode + "." + data_type
print("Writing data files with prefix (.filename, .doc, .title, .image, .label, .weight, .rewards): %s"%full_data_file_prefix)
ffilenames = open(full_data_file_prefix+".filename", "w")
fdoc = open(full_data_file_prefix+".doc", "w")
ftitle = open(full_data_file_prefix+".title", "w")
fimage = open(full_data_file_prefix+".image", "w")
flabel = open(full_data_file_prefix+".label", "w")
fweight = open(full_data_file_prefix+".weight", "w")
freward = open(full_data_file_prefix+".reward", "w")
for filename, doc, title, image, label, weight, reward in zip(self.filenames, self.docs, self.titles, self.images, self.labels, self.weights, self.rewards):
ffilenames.write(filename+"\n")
fdoc.write("\n".join([" ".join([str(item) for item in itemlist]) for itemlist in doc])+"\n\n")
ftitle.write("\n".join([" ".join([str(item) for item in itemlist]) for itemlist in title])+"\n\n")
fimage.write("\n".join([" ".join([str(item) for item in itemlist]) for itemlist in image])+"\n\n")
flabel.write("\n".join([" ".join([str(item) for item in itemlist]) for itemlist in label])+"\n\n")
fweight.write(" ".join([str(item) for item in weight])+"\n")
freward.write(" ".join([str(item) for item in reward])+"\n")
ffilenames.close()
fdoc.close()
ftitle.close()
fimage.close()
flabel.close()
fweight.close()
freward.close()
def populate_data(self, vocab_dict, data_type):
full_data_file_prefix = FLAGS.preprocessed_data_directory + "/" + FLAGS.data_mode + "." + data_type
print("Data file prefix (.doc, .title, .image, .label.multipleoracle): %s"%full_data_file_prefix)
# Process doc, title, image, label
doc_data_list = open(full_data_file_prefix+".doc").read().strip().split("\n\n")
title_data_list = open(full_data_file_prefix+".title").read().strip().split("\n\n")
image_data_list = open(full_data_file_prefix+".image").read().strip().split("\n\n")
label_data_list = open(full_data_file_prefix+".label.multipleoracle").read().strip().split("\n\n")
print("Data sizes: %d %d %d %d"%(len(doc_data_list), len(title_data_list), len(image_data_list), len(label_data_list)))
print("Reading data (no padding to save memory) ...")
doccount = 0
for doc_data, title_data, image_data, label_data in zip(doc_data_list, title_data_list, image_data_list, label_data_list):
doc_lines = doc_data.strip().split("\n")
title_lines = title_data.strip().split("\n")
image_lines = image_data.strip().split("\n")
label_lines = label_data.strip().split("\n")
filename = doc_lines[0].strip()
if ((filename == title_lines[0].strip()) and (filename == image_lines[0].strip()) and (filename == label_lines[0].strip())):
# Put filename
self.filenames.append(filename)
# Doc
thisdoc = []
for line in doc_lines[1:FLAGS.max_doc_length+1]:
thissent = [int(item) for item in line.strip().split()]
thisdoc.append(thissent)
self.docs.append(thisdoc)
# Title
thistitle = []
for line in title_lines[1:FLAGS.max_title_length+1]:
thissent = [int(item) for item in line.strip().split()]
thistitle.append(thissent)
self.titles.append(thistitle)
# Image
thisimage = []
for line in image_lines[1:FLAGS.max_image_length+1]:
thissent = [int(item) for item in line.strip().split()]
thisimage.append(thissent)
self.images.append(thisimage)
# Weights
originaldoclen = int(label_lines[1].strip())
thisweight = [1 for item in range(originaldoclen)][:FLAGS.max_doc_length]
self.weights.append(thisweight)
# Labels (multiple oracles and preestimated rewards)
thislabel = []
thisreward = []
for line in label_lines[2:FLAGS.num_sample_rollout+2]:
thislabel.append([int(item) for item in line.split()[:-1]])
thisreward.append(float(line.split()[-1]))
self.labels.append(thislabel)
self.rewards.append(thisreward)
else:
print("Some problem with %s.* files. Exiting!"%full_data_file_prefix)
exit(0)
if doccount%10000==0:
print("%d ..."%doccount)
doccount += 1
# Set Fileindices
self.fileindices = range(len(self.filenames))
class DataProcessor:
def prepare_news_data(self, vocab_dict, data_type="training"):
data = Data(vocab_dict, data_type)
return data
def prepare_vocab_embeddingdict(self):
# Numpy dtype
dtype = np.float16 if FLAGS.use_fp16 else np.float32
vocab_dict = {}
word_embedding_array = []
# Add padding
vocab_dict["_PAD"] = PAD_ID
# Add UNK
vocab_dict["_UNK"] = UNK_ID
# Read word embedding file
wordembed_filename = FLAGS.pretrained_wordembedding
print("Reading pretrained word embeddings file: %s"%wordembed_filename)
embed_line = ""
linecount = 0
with open(wordembed_filename, "r") as fembedd:
for line in fembedd:
if linecount == 0:
vocabsize = int(line.split()[0])
# Initiate fixed size empty array
word_embedding_array = np.empty((vocabsize, FLAGS.wordembed_size), dtype=dtype)
else:
linedata = line.split()
vocab_dict[linedata[0]] = linecount + 1
embeddata = [float(item) for item in linedata[1:]][0:FLAGS.wordembed_size]
word_embedding_array[linecount-1] = np.array(embeddata, dtype=dtype)
if linecount%100000 == 0:
print(str(linecount)+" ...")
linecount += 1
print("Read pretrained embeddings: %s"%str(word_embedding_array.shape))
print("Size of vocab: %d (_PAD:0, _UNK:1)"%len(vocab_dict))
vocabfilename = FLAGS.train_dir+"/vocab.txt"
print("Writing vocab file: %s"%vocabfilename)
foutput = open(vocabfilename,"w")
vocab_list = [(vocab_dict[key], key) for key in vocab_dict.keys()]
vocab_list.sort()
vocab_list = [item[1] for item in vocab_list]
foutput.write("\n".join(vocab_list)+"\n")
foutput.close()
return vocab_dict, word_embedding_array