|
1
|
|
|
from fuel.transformers import Transformer |
|
2
|
|
|
|
|
3
|
|
|
|
|
4
|
|
|
class Window(Transformer): |
|
5
|
|
|
"""Return pairs of source and target windows from a stream. |
|
6
|
|
|
|
|
7
|
|
|
This data stream wrapper takes as an input a data stream outputting |
|
8
|
|
|
sequences of potentially varying lengths (e.g. sentences, audio tracks, |
|
9
|
|
|
etc.). It then returns two sliding windows (source and target) over |
|
10
|
|
|
these sequences. |
|
11
|
|
|
|
|
12
|
|
|
For example, to train an n-gram model set `source_window` to n, |
|
13
|
|
|
`target_window` to 1, no offset, and `overlapping` to false. This will |
|
14
|
|
|
give chunks [1, N] and [N + 1]. To train an RNN you often want to set |
|
15
|
|
|
the source and target window to the same size and use an offset of 1 |
|
16
|
|
|
with overlap, this would give you chunks [1, N] and [2, N + 1]. |
|
17
|
|
|
|
|
18
|
|
|
Parameters |
|
19
|
|
|
---------- |
|
20
|
|
|
offset : int |
|
21
|
|
|
The offset from the source window where the target window starts. |
|
22
|
|
|
source_window : int |
|
23
|
|
|
The size of the source window. |
|
24
|
|
|
target_window : int |
|
25
|
|
|
The size of the target window. |
|
26
|
|
|
overlapping : bool |
|
27
|
|
|
If true, the source and target windows overlap i.e. the offset of |
|
28
|
|
|
the target window is taken to be from the beginning of the source |
|
29
|
|
|
window. If false, the target window offset is taken to be from the |
|
30
|
|
|
end of the source window. |
|
31
|
|
|
data_stream : :class:`.DataStream` instance |
|
32
|
|
|
The data stream providing sequences. Each example is assumed to be |
|
33
|
|
|
an object that supports slicing. |
|
34
|
|
|
target_source : str, optional |
|
35
|
|
|
This data stream adds a new source for the target words. By default |
|
36
|
|
|
this source is 'targets'. |
|
37
|
|
|
|
|
38
|
|
|
""" |
|
39
|
|
|
def __init__(self, offset, source_window, target_window, |
|
40
|
|
|
overlapping, data_stream, target_source='targets', **kwargs): |
|
41
|
|
|
if not data_stream.produces_examples: |
|
42
|
|
|
raise ValueError('the wrapped data stream must produce examples, ' |
|
43
|
|
|
'not batches of examples.') |
|
44
|
|
|
if len(data_stream.sources) > 1: |
|
45
|
|
|
raise ValueError('{} expects only one source' |
|
46
|
|
|
.format(self.__class__.__name__)) |
|
47
|
|
|
|
|
48
|
|
|
super(Window, self).__init__(data_stream, produces_examples=True, |
|
49
|
|
|
**kwargs) |
|
50
|
|
|
self.sources = self.sources + (target_source,) |
|
51
|
|
|
|
|
52
|
|
|
self.offset = offset |
|
53
|
|
|
self.source_window = source_window |
|
54
|
|
|
self.target_window = target_window |
|
55
|
|
|
self.overlapping = overlapping |
|
56
|
|
|
|
|
57
|
|
|
self.sentence = [] |
|
58
|
|
|
self._set_index() |
|
59
|
|
|
|
|
60
|
|
|
def _set_index(self): |
|
61
|
|
|
"""Set the starting index of the source window.""" |
|
62
|
|
|
self.index = 0 |
|
63
|
|
|
# If offset is negative, target window might start before 0 |
|
64
|
|
|
self.index = -min(0, self._get_target_index()) |
|
65
|
|
|
|
|
66
|
|
|
def _get_target_index(self): |
|
67
|
|
|
"""Return the index where the target window starts.""" |
|
68
|
|
|
return (self.index + self.source_window * (not self.overlapping) + |
|
69
|
|
|
self.offset) |
|
70
|
|
|
|
|
71
|
|
|
def _get_end_index(self): |
|
72
|
|
|
"""Return the end of both windows.""" |
|
73
|
|
|
return max(self.index + self.source_window, |
|
74
|
|
|
self._get_target_index() + self.target_window) |
|
75
|
|
|
|
|
76
|
|
|
def get_data(self, request=None): |
|
77
|
|
|
if request is not None: |
|
78
|
|
|
raise ValueError |
|
79
|
|
|
while not self._get_end_index() <= len(self.sentence): |
|
80
|
|
|
self.sentence, = next(self.child_epoch_iterator) |
|
81
|
|
|
self._set_index() |
|
82
|
|
|
source = self.sentence[self.index:self.index + self.source_window] |
|
83
|
|
|
target = self.sentence[self._get_target_index(): |
|
84
|
|
|
self._get_target_index() + self.target_window] |
|
85
|
|
|
self.index += 1 |
|
86
|
|
|
return (source, target) |
|
87
|
|
|
|
|
88
|
|
|
|
|
89
|
|
|
class NGrams(Window): |
|
90
|
|
|
"""Return n-grams from a stream. |
|
91
|
|
|
|
|
92
|
|
|
This data stream wrapper takes as an input a data stream outputting |
|
93
|
|
|
sentences. From these sentences n-grams of a fixed order (e.g. bigrams, |
|
94
|
|
|
trigrams, etc.) are extracted and returned. It also creates a |
|
95
|
|
|
``targets`` data source. For each example, the target is the word |
|
96
|
|
|
immediately following that n-gram. It is normally used for language |
|
97
|
|
|
modeling, where we try to predict the next word from the previous *n* |
|
98
|
|
|
words. |
|
99
|
|
|
|
|
100
|
|
|
.. note:: |
|
101
|
|
|
|
|
102
|
|
|
Unlike the :class:`Window` stream, the target returned by |
|
103
|
|
|
:class:`NGrams` is a single element instead of a window. |
|
104
|
|
|
|
|
105
|
|
|
Parameters |
|
106
|
|
|
---------- |
|
107
|
|
|
ngram_order : int |
|
108
|
|
|
The order of the n-grams to output e.g. 3 for trigrams. |
|
109
|
|
|
data_stream : :class:`.DataStream` instance |
|
110
|
|
|
The data stream providing sentences. Each example is assumed to be |
|
111
|
|
|
a list of integers. |
|
112
|
|
|
target_source : str, optional |
|
113
|
|
|
This data stream adds a new source for the target words. By default |
|
114
|
|
|
this source is 'targets'. |
|
115
|
|
|
|
|
116
|
|
|
""" |
|
117
|
|
|
def __init__(self, ngram_order, *args, **kwargs): |
|
118
|
|
|
super(NGrams, self).__init__( |
|
119
|
|
|
0, ngram_order, 1, False, *args, **kwargs) |
|
120
|
|
|
|
|
121
|
|
|
def get_data(self, *args, **kwargs): |
|
122
|
|
|
source, target = super(NGrams, self).get_data(*args, **kwargs) |
|
123
|
|
|
return (source, target[0]) |
|
124
|
|
|
|
mila-udem /
fuel
