-
Notifications
You must be signed in to change notification settings - Fork 388
/
dataset.py
283 lines (227 loc) · 9.46 KB
/
dataset.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
"""Contains custom skorch Dataset and CVSplit."""
from functools import partial
from numbers import Number
import warnings
import numpy as np
from sklearn.model_selection import ShuffleSplit
from sklearn.model_selection import StratifiedKFold
from sklearn.model_selection import StratifiedShuffleSplit
from sklearn.model_selection import check_cv
import torch
import torch.utils.data
from skorch.utils import flatten
from skorch.utils import is_pandas_ndframe
from skorch.utils import check_indexing
from skorch.utils import multi_indexing
from skorch.utils import to_numpy
def _apply_to_data(data, func, unpack_dict=False):
"""Apply a function to data, trying to unpack different data
types.
"""
apply_ = partial(_apply_to_data, func=func, unpack_dict=unpack_dict)
if isinstance(data, dict):
if unpack_dict:
return [apply_(v) for v in data.values()]
return {k: apply_(v) for k, v in data.items()}
elif isinstance(data, (list, tuple)):
try:
# e.g.list/tuple of arrays
return [apply_(x) for x in data]
except TypeError:
return func(data)
return func(data)
def get_len(data):
lens = [_apply_to_data(data, len, unpack_dict=True)]
lens = list(flatten(lens))
len_set = set(lens)
if len(len_set) != 1:
raise ValueError("Dataset does not have consistent lengths.")
return list(len_set)[0]
def uses_placeholder_y(ds):
"""If ``ds`` is a ``skorch.dataset.Dataset`` or a
``skorch.dataset.Dataset`` nested inside a
``torch.utils.data.Subset`` and uses
y as a placeholder, return ``True``."""
if isinstance(ds, torch.utils.data.Subset):
return uses_placeholder_y(ds.dataset)
return isinstance(ds, Dataset) and hasattr(ds, "y") and ds.y is None
class Dataset(torch.utils.data.Dataset):
# pylint: disable=anomalous-backslash-in-string
"""General dataset wrapper that can be used in conjunction with
PyTorch :class:`~torch.utils.data.DataLoader`.
The dataset will always yield a tuple of two values, the first
from the data (``X``) and the second from the target (``y``).
However, the target is allowed to be ``None``. In that case,
:class:`.Dataset` will currently return a dummy tensor, since
:class:`~torch.utils.data.DataLoader` does not work with
``None``\s.
:class:`.Dataset` currently works with the following data types:
* numpy ``array``\s
* PyTorch :class:`~torch.Tensor`\s
* pandas NDFrame
* a dictionary of the former three
* a list/tuple of the former three
Parameters
----------
X : see above
Everything pertaining to the input data.
y : see above or None (default=None)
Everything pertaining to the target, if there is anything.
length : int or None (default=None)
If not ``None``, determines the length (``len``) of the data.
Should usually be left at ``None``, in which case the length is
determined by the data itself.
"""
def __init__(
self,
X,
y=None,
device=None,
length=None,
):
# TODO: Remove warning in release 0.4
if device is not None:
warnings.warn(
"device is no longer needed by Dataset and will be ignored.",
DeprecationWarning)
self.X = X
self.y = y
self.X_indexing = check_indexing(X)
self.y_indexing = check_indexing(y)
self.X_is_ndframe = is_pandas_ndframe(X)
if length is not None:
self._len = length
return
# pylint: disable=invalid-name
len_X = get_len(X)
if y is not None:
len_y = get_len(y)
if len_y != len_X:
raise ValueError("X and y have inconsistent lengths.")
self._len = len_X
def __len__(self):
return self._len
def transform(self, X, y):
# pylint: disable=anomalous-backslash-in-string
"""Additional transformations on ``X`` and ``y``.
By default, they are cast to PyTorch :class:`~torch.Tensor`\s.
Override this if you want a different behavior.
Note: If you use this in conjuction with PyTorch
:class:`~torch.utils.data.DataLoader`, the latter will call
the dataset for each row separately, which means that the
incoming ``X`` and ``y`` each are single rows.
"""
# pytorch DataLoader cannot deal with None so we use 0 as a
# placeholder value. We only return a Tensor with one value
# (as opposed to ``batchsz`` values) since the pytorch
# DataLoader calls __getitem__ for each row in the batch
# anyway, which results in a dummy ``y`` value for each row in
# the batch.
y = torch.Tensor([0]) if y is None else y
return X, y
def __getitem__(self, i):
X, y = self.X, self.y
if self.X_is_ndframe:
X = {k: X[k].values.reshape(-1, 1) for k in X}
Xi = multi_indexing(X, i, self.X_indexing)
yi = multi_indexing(y, i, self.y_indexing)
return self.transform(Xi, yi)
class CVSplit(object):
"""Class that performs the internal train/valid split on a dataset.
The ``cv`` argument here works similarly to the regular sklearn ``cv``
parameter in, e.g., ``GridSearchCV``. However, instead of cycling
through all splits, only one fixed split (the first one) is
used. To get a full cycle through the splits, don't use
``NeuralNet``'s internal validation but instead the corresponding
sklearn functions (e.g. ``cross_val_score``).
We additionally support a float, similar to sklearn's
``train_test_split``.
Parameters
----------
cv : int, float, cross-validation generator or an iterable, optional
(Refer sklearn's User Guide for cross_validation for the various
cross-validation strategies that can be used here.)
Determines the cross-validation splitting strategy.
Possible inputs for cv are:
- None, to use the default 3-fold cross validation,
- integer, to specify the number of folds in a ``(Stratified)KFold``,
- float, to represent the proportion of the dataset to include
in the validation split.
- An object to be used as a cross-validation generator.
- An iterable yielding train, validation splits.
stratified : bool (default=False)
Whether the split should be stratified. Only works if ``y`` is
either binary or multiclass classification.
random_state : int, RandomState instance, or None (default=None)
Control the random state in case that ``(Stratified)ShuffleSplit``
is used (which is when a float is passed to ``cv``). For more
information, look at the sklearn documentation of
``(Stratified)ShuffleSplit``.
"""
def __init__(
self,
cv=5,
stratified=False,
random_state=None,
):
self.stratified = stratified
self.random_state = random_state
if isinstance(cv, Number) and (cv <= 0):
raise ValueError("Numbers less than 0 are not allowed for cv "
"but CVSplit got {}".format(cv))
self.cv = cv
def _is_stratified(self, cv):
return isinstance(cv, (StratifiedKFold, StratifiedShuffleSplit))
def _is_float(self, x):
if not isinstance(x, Number):
return False
return not float(x).is_integer()
def _check_cv_float(self):
cv_cls = StratifiedShuffleSplit if self.stratified else ShuffleSplit
return cv_cls(test_size=self.cv, random_state=self.random_state)
def _check_cv_non_float(self, y):
return check_cv(
self.cv,
y=y,
classifier=self.stratified,
)
def check_cv(self, y):
"""Resolve which cross validation strategy is used."""
y_arr = None
if self.stratified:
# Try to convert y to numpy for sklearn's check_cv; if conversion
# doesn't work, still try.
try:
y_arr = to_numpy(y)
except (AttributeError, TypeError):
y_arr = y
if self._is_float(self.cv):
return self._check_cv_float()
return self._check_cv_non_float(y_arr)
def _is_regular(self, x):
return (x is None) or isinstance(x, np.ndarray) or is_pandas_ndframe(x)
def __call__(self, dataset, y=None, groups=None):
bad_y_error = ValueError(
"Stratified CV requires explicitely passing a suitable y.")
if (y is None) and self.stratified:
raise bad_y_error
cv = self.check_cv(y)
if self.stratified and not self._is_stratified(cv):
raise bad_y_error
# pylint: disable=invalid-name
len_dataset = get_len(dataset)
if y is not None:
len_y = get_len(y)
if len_dataset != len_y:
raise ValueError("Cannot perform a CV split if dataset and y "
"have different lengths.")
args = (np.arange(len_dataset),)
if self._is_stratified(cv):
args = args + (to_numpy(y),)
idx_train, idx_valid = next(iter(cv.split(*args, groups=groups)))
dataset_train = torch.utils.data.Subset(dataset, idx_train)
dataset_valid = torch.utils.data.Subset(dataset, idx_valid)
return dataset_train, dataset_valid
def __repr__(self):
# pylint: disable=useless-super-delegation
return super(CVSplit, self).__repr__()