|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
import logging as loggers |
|
5
|
|
|
import numpy as np |
|
6
|
|
|
from deepy.dataset import Dataset |
|
7
|
|
|
from deepy.utils import FLOATX |
|
8
|
|
|
|
|
9
|
|
|
logging = loggers.getLogger(__name__) |
|
10
|
|
|
|
|
11
|
|
|
class LMDataset(Dataset): |
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
def __init__(self, vocab, train_path, valid_path, history_len=-1, char_based=False, max_tokens=999, min_tokens=0, sort=True): |
|
15
|
|
|
""" |
|
16
|
|
|
Generate data for training with RNN |
|
17
|
|
|
:type vocab: vocab.Vocab |
|
18
|
|
|
""" |
|
19
|
|
|
assert history_len == -1 or history_len >= 1 |
|
20
|
|
|
self.vocab = vocab |
|
21
|
|
|
self.history_len = history_len |
|
22
|
|
|
self.char_based = char_based |
|
23
|
|
|
self.sort = sort |
|
24
|
|
|
|
|
25
|
|
|
self.min_tokens = min_tokens |
|
26
|
|
|
self.max_tokens = max_tokens |
|
27
|
|
|
|
|
28
|
|
|
self._train_set = self.read_data(train_path) |
|
29
|
|
|
self._valid_set = self.read_data(valid_path) |
|
30
|
|
|
|
|
31
|
|
|
def train_set(self): |
|
32
|
|
|
return self._train_set |
|
33
|
|
|
|
|
34
|
|
|
def valid_set(self): |
|
35
|
|
|
return self._valid_set |
|
36
|
|
|
|
|
37
|
|
|
def read_data(self, path): |
|
38
|
|
|
data = [] |
|
39
|
|
|
sent_count = 0 |
|
40
|
|
|
for line in open(path).xreadlines(): |
|
41
|
|
|
line = line.strip() |
|
42
|
|
|
wc = len(line) if self.char_based else line.count(" ") + 1 |
|
43
|
|
|
if wc < self.min_tokens or wc > self.max_tokens: |
|
44
|
|
|
continue |
|
45
|
|
|
sent_count += 1 |
|
46
|
|
|
sequence = [self.vocab.sent_index] |
|
47
|
|
|
tokens = line if self.char_based else line.split(" ") |
|
48
|
|
|
for w in tokens: |
|
49
|
|
|
sequence.append(self.vocab.index(w)) |
|
50
|
|
|
sequence.append(self.vocab.sent_index) |
|
51
|
|
|
|
|
52
|
|
|
if self.history_len == -1: |
|
53
|
|
|
# Full sentence |
|
54
|
|
|
data.append(self.convert_to_data(sequence)) |
|
55
|
|
|
else: |
|
56
|
|
|
# trunk by trunk |
|
57
|
|
|
for begin in range(0, len(sequence), self.history_len): |
|
58
|
|
|
trunk = sequence[begin: begin + self.history_len + 1] |
|
59
|
|
|
if len(trunk) > 1: |
|
60
|
|
|
data.append(self.convert_to_data(trunk)) |
|
61
|
|
|
if self.sort: |
|
62
|
|
|
data.sort(key=lambda x: len(x[1])) |
|
63
|
|
|
logging.info("loaded from %s: %d sentences, %d data pieces " % (path, sent_count, len(data))) |
|
64
|
|
|
return data |
|
65
|
|
|
|
|
66
|
|
|
def convert_to_data(self, seq): |
|
67
|
|
|
assert len(seq) >= 2 |
|
68
|
|
|
input_indices = seq[:-1] |
|
69
|
|
|
target_indices = seq[1:] |
|
70
|
|
|
return input_indices, target_indices |
|
71
|
|
|
|
|
72
|
|
|
|
|
73
|
|
|
|
uaca /
deepy
