|
1
|
|
|
import os |
|
2
|
|
|
|
|
3
|
|
|
from fuel.datasets import TextFile |
|
4
|
|
|
from fuel.utils import find_in_data_path |
|
5
|
|
|
|
|
6
|
|
|
|
|
7
|
|
|
class OneBillionWord(TextFile): |
|
8
|
|
|
"""Google's One Billion Word benchmark. |
|
9
|
|
|
|
|
10
|
|
|
This monolingual corpus contains 829,250,940 tokens (including sentence |
|
11
|
|
|
boundary markers). The data is split into 100 partitions, one of which |
|
12
|
|
|
is the held-out set. This held-out set is further divided into 50 |
|
13
|
|
|
partitions. More information about the dataset can be found in |
|
14
|
|
|
[CMSG14]. |
|
15
|
|
|
|
|
16
|
|
|
.. [CSMG14] Ciprian Chelba, Tomas Mikolov, Mike Schuster, Qi Ge, and |
|
17
|
|
|
Thorsten Brants, *One Billion Word Benchmark for Measuring Progress |
|
18
|
|
|
in Statistical Language Modeling*, `arXiv:1312.3005 [cs.CL] |
|
19
|
|
|
<http://arxiv.org/abs/1312.3005>`. |
|
20
|
|
|
|
|
21
|
|
|
Parameters |
|
22
|
|
|
---------- |
|
23
|
|
|
which_set : 'training' or 'heldout' |
|
24
|
|
|
Which dataset to load. |
|
25
|
|
|
which_partitions : list of ints |
|
26
|
|
|
For the training set, valid values must lie in [1, 99]. For the |
|
27
|
|
|
heldout set they must be in [0, 49]. |
|
28
|
|
|
vocabulary : dict |
|
29
|
|
|
A dictionary mapping tokens to integers. This dictionary is |
|
30
|
|
|
expected to contain the tokens ``<S>``, ``</S>`` and ``<UNK>``, |
|
31
|
|
|
representing "start of sentence", "end of sentence", and |
|
32
|
|
|
"out-of-vocabulary" (OoV). The latter will be used whenever a token |
|
33
|
|
|
cannot be found in the vocabulary. |
|
34
|
|
|
preprocess : function, optional |
|
35
|
|
|
A function that takes a string (a sentence including new line) as |
|
36
|
|
|
an input and returns a modified string. A useful function to pass |
|
37
|
|
|
could be ``str.lower``. |
|
38
|
|
|
|
|
39
|
|
|
See :class:`TextFile` for remaining keyword arguments. |
|
40
|
|
|
|
|
41
|
|
|
""" |
|
42
|
|
|
def __init__(self, which_set, which_partitions, dictionary, **kwargs): |
|
43
|
|
|
if which_set not in ('training', 'heldout'): |
|
44
|
|
|
raise ValueError |
|
45
|
|
|
if which_set == 'training': |
|
46
|
|
|
if not all(partition in range(1, 100) |
|
47
|
|
|
for partition in which_partitions): |
|
48
|
|
|
raise ValueError |
|
49
|
|
|
files = [find_in_data_path(os.path.join( |
|
50
|
|
|
'1-billion-word', 'training-monolingual.tokenized.shuffled', |
|
51
|
|
|
'news.en-{:05d}-of-00100'.format(partition))) |
|
52
|
|
|
for partition in which_partitions] |
|
53
|
|
|
else: |
|
54
|
|
|
if not all(partition in range(50) |
|
55
|
|
|
for partition in which_partitions): |
|
56
|
|
|
raise ValueError |
|
57
|
|
|
files = [find_in_data_path(os.path.join( |
|
58
|
|
|
'1-billion-word', 'heldout-monolingual.tokenized.shuffled', |
|
59
|
|
|
'news.en.heldout-{:05d}-of-00050'.format(partition))) |
|
60
|
|
|
for partition in which_partitions] |
|
61
|
|
|
super(OneBillionWord, self).__init__(files, dictionary, **kwargs) |
|
62
|
|
|
|
mila-udem /
fuel
