-
Notifications
You must be signed in to change notification settings - Fork 2
/
prepare_cca.py
246 lines (189 loc) · 7.35 KB
/
prepare_cca.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
import collections
import json
import logging
import os
import random
import re
import h5py
import numpy as np
import math
import data_model
word_re = re.compile(r'([A-Za-z0-9_]+)')
def tokenize(text):
for match in word_re.finditer(text):
yield match.group(1)
def tokenize_letter(text):
for letter in text:
yield letter
class XTrackCCA(object):
def build(self, dialogs, out_file):
for dialog_ndx, dialog in enumerate(dialogs):
for msgs, state, actor in zip(dialog.messages,
dialog.states,
dialog.actors):
actor_is_system = actor == data_model.Dialog.ACTOR_SYSTEM
if actor_is_system:
msg_id = 0
else:
msg_id = random.choice(n_best_order)
msg, msg_score = msgs[msg_id]
true_msg, _ = msgs[0]
if actor == data_model.Dialog.ACTOR_USER:
self.msg_scores.append(np.exp(msg_score))
if not include_system_utterances and actor_is_system:
continue
else:
#msg_score = max(msg_score, -100)
#msg_score = np.exp(msg_score)
self._process_msg(msg, msg_score, state, last_state,
actor, seq,
oov_ins_p, word_drop_p, n_best_order,
f_dump_text,
true_msg, score_bins)
last_state = state
# Sanity check that all data elements are equal size.
data_lens = [len(seq[key]) for key in seq_data_keys]
assert data_lens[1:] == data_lens[:-1]
if len(seq['data']) > 0:
n_labels += len(seq['labels'])
if not split_dialogs:
self.sequences.append(seq)
else:
self.sequences.extend(self._split_dialog(seq))
f_dump_text.write('\n')
logging.info('There are in total %d labels in %d sequences.'
% (n_labels, len(self.sequences, )))
#if not self.stats:
# logging.info('Computing stats.')
# self._compute_stats('data_score', 'data_switch')
#logging.info('Normalizing.')
#self._normalize('data_score', 'data_switch')
if not self.based_on:
logging.info('Building token features.')
self._build_token_features()
def _build_token_features(self):
self.token_features = {}
for word, word_id in self.vocab.iteritems():
features = []
for slot in self.slots:
features.append(int(word in slot))
for cls in self.classes[slot]:
ftr_val = 0
for cls_part in cls.split():
if cls_part[0] == '@':
cls_part = cls_part[1:]
if word in cls_part:
ftr_val = 1
break
features.append(ftr_val)
self.token_features[word_id] = features
def _compute_stats(self, *vars):
score = {var: [] for var in vars}
for seq in self.sequences:
for var in vars:
score[var].extend(seq[var])
#import ipdb; ipdb.set_trace()
self.stats = {}
for var in vars:
mean = np.mean(score[var])
stddev = np.std(score[var])
self.stats[var] = {
'mean': mean,
'stddev': stddev
}
def _normalize(self, *vars):
for seq in self.sequences:
for var in vars:
res = seq[var]
for i in xrange(len(res)):
res[i] -= self.stats[var]['mean']
res[i] /= self.stats[var]['stddev'] + 1e-7
def get_token_ndx(self, token):
if token in self.vocab:
return self.vocab[token]
else:
if not self.vocab_fixed:
self.vocab[token] = res = len(self.vocab)
self.vocab_rev[self.vocab[token]] = token
return res
else:
logging.warning('Mapping to OOV: %s' % token)
return self.vocab['#OOV']
def state_to_label(self, state, slots):
res = []
for slot in slots:
res.append(self.state_to_label_for(state, slot))
return res
def state_to_label_for(self, state, slot):
if not state:
return self.classes[slot][self.null_class]
else:
value = state.get(slot)
if value:
food = value #next(tokenize(value))
if self.vocab_fixed:
if not food in self.classes[slot]:
res = self.classes[slot][self.null_class]
else:
res = self.classes[slot][food]
else:
if not food in self.classes[slot]:
self.classes[slot][food] = len(self.classes[slot])
res = self.classes[slot][food]
else:
res = self.classes[slot][self.null_class]
return res
def save(self, out_file):
with open(out_file, 'w') as f_out:
obj = {}
for attr in self.attrs_to_save:
obj[attr] = getattr(self, attr)
json.dump(obj, f_out, indent=4)
import matplotlib
matplotlib.use('Agg')
import seaborn as sns
import matplotlib.pyplot as plt
sns.set_palette("deep", desat=.6)
plt.figure()
plt.hist(self.msg_scores, [0.0, 0.3, 0.6, 0.95, 1.0])
plt.savefig(out_file + '.score.png')
plt.figure()
plt.hist(np.log(np.array(self.word_freq.values())))
plt.savefig(out_file + '.word_freqs.png')
with open(out_file + '.oov.txt', 'w') as f_out:
for word, freq in self.word_freq.most_common():
if freq < 5:
f_out.write(word + '\n')
#import ipdb; ipdb.set_trace()
@classmethod
def load(cls, in_file):
with open(in_file, 'r') as f_in:
data = json.load(f_in)
xtd = XTrackData2()
for attr in cls.attrs_to_save:
val = data[attr]
setattr(xtd, attr, val)
xtd._init_after_load()
return xtd
if __name__ == '__main__':
from utils import init_logging
init_logging('XTrack CCA')
random.seed(0)
from utils import pdb_on_error
pdb_on_error()
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--data_dir', required=True)
parser.add_argument('--out_file', required=True)
args = parser.parse_args()
dialogs = []
for f_name in sorted(os.listdir(args.data_dir), key=lambda x: int(x.split(
'.')[0])):
if f_name.endswith('.json'):
dialogs.append(
data_model.Dialog.deserialize(
open(os.path.join(args.data_dir, f_name)).read()
)
)
xcca = XTrackCCA()
xcca.build(dialogs=dialogs, out_file=args.out_file)