forked from PaddlePaddle/PaddleNLP
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcollate.py
333 lines (282 loc) · 12.2 KB
/
collate.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
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import paddle
__all__ = [
'Stack',
'Pad',
'Tuple',
'Dict',
]
class Stack(object):
"""
Stacks the input data samples to construct the batch. The N input samples
must have the same shape/length and will be stacked to construct a batch.
Args:
axis (int, optional): The axis in the result data along which the input
data are stacked. Default: 0.
dtype (str|numpy.dtype, optional): The value type of the output. If it
is set to None, the type of input data is used. Default: None.
"""
def __init__(self, axis=0, dtype=None):
self._axis = axis
self._dtype = dtype
def __call__(self, data):
"""
Batchifies the input data by stacking.
Args:
data (list[numpy.ndarray]): The input data samples. It is a list.
Each element is a numpy.ndarray or list.
Returns:
numpy.ndarray: Stacked batch data.
Example:
.. code-block:: python
from paddlenlp.data import Stack
a = [1, 2, 3, 4]
b = [3, 4, 5, 6]
c = [5, 6, 7, 8]
result = Stack()([a, b, c])
'''
[[1, 2, 3, 4],
[3, 4, 5, 6],
[5, 6, 7, 8]]
'''
"""
data = np.stack(data, axis=self._axis).astype(
self._dtype) if self._dtype else np.stack(data, axis=self._axis)
return data
class Pad(object):
"""
Pads the input data samples to the largest length at `axis`.
Args:
pad_val (float|int, optional): The padding value. Default: 0.
axis (int, optional): The axis to pad the arrays. The arrays will be
padded to the largest length at `axis`. For example, assume the
input arrays have shape (10, 8, 5), (6, 8, 5), (3, 8, 5) and the
axis is 0. Each input will be padded into (10, 8, 5) and then
stacked to form the final output, which has shape (3, 10, 8, 5).
Default: 0.
ret_length (bool|numpy.dtype, optional): If it is bool, indicate whether
to return the valid length in the output, and the data type of
returned length is int32 if True. If it is numpy.dtype, indicate the
data type of returned length. Default: None.
dtype (numpy.dtype, optional): The value type of the output. If it is
set to None, the input data type is used. Default: None.
pad_right (bool, optional): Whether the padding direction is right-side.
If True, it indicates we pad to the right side, while False indicates
we pad to the left side. Default: True.
"""
def __init__(self,
pad_val=0,
axis=0,
ret_length=None,
dtype=None,
pad_right=True):
self._pad_val = pad_val
self._axis = axis
self._ret_length = ret_length
self._dtype = dtype
self._pad_right = pad_right
def __call__(self, data):
"""
Batchifies the input data by padding. The input will be padded to the
largest dimension at `axis` and then stacked to form the final output.
In addition, the function will output the original dimensions at the
`axis` if `ret_length` is not None or False.
Args:
data (list[numpy.ndarray|list]): The input data samples. It is a
list. Each element is a numpy.ndarray or list.
Returns:
numpy.ndarray|tuple[numpy.ndarray]: If `ret_length` is False, it
is a numpy.ndarray representing the padded batch data and the
shape is (N, …). Otherwise, it is a tuple, besides the padded batch
data, the tuple also includes a numpy.ndarray representing original
length at `axis` of all input samples, which shaped `(N,)`.
Example:
.. code-block:: python
from paddlenlp.data import Pad
a = [1, 2, 3, 4]
b = [5, 6, 7]
c = [8, 9]
result = Pad(pad_val=0)([a, b, c])
'''
[[1, 2, 3, 4],
[5, 6, 7, 0],
[8, 9, 0, 0]]
'''
"""
# return data itself for rare unexpected cases when 1-D array is passed to Pad
if not isinstance(data[0], list) and not isinstance(
data[0], np.ndarray):
return np.asarray(
data,
dtype=self._dtype if self._dtype is not None else np.int64)
arrs = [np.asarray(ele) for ele in data]
original_length = [ele.shape[self._axis] for ele in arrs]
max_size = max(original_length)
ret_shape = list(arrs[0].shape)
ret_shape[self._axis] = max_size
ret_shape = (len(arrs), ) + tuple(ret_shape)
ret = np.full(
shape=ret_shape,
fill_value=self._pad_val,
dtype=arrs[0].dtype if self._dtype is None else self._dtype)
for i, arr in enumerate(arrs):
if arr.shape[self._axis] == max_size:
ret[i] = arr
else:
slices = [slice(None) for _ in range(arr.ndim)]
if self._pad_right:
slices[self._axis] = slice(0, arr.shape[self._axis])
else:
slices[self._axis] = slice(max_size - arr.shape[self._axis],
max_size)
if slices[self._axis].start != slices[self._axis].stop:
slices = [slice(i, i + 1)] + slices
ret[tuple(slices)] = arr
if self._ret_length:
return ret, np.asarray(
original_length,
dtype="int32") if self._ret_length == True else np.asarray(
original_length, self._ret_length)
else:
return ret
class Tuple(object):
"""
Wraps multiple batchify functions together. The input functions will be applied
to the corresponding input fields.
Each sample should be a list or tuple containing multiple fields. The i'th
batchify function stored in Tuple will be applied on the i'th field.
For example, when data sample is (nd_data, label), you can wrap two batchify
functions using `Tuple(DataBatchify, LabelBatchify)` to batchify nd_data and
label correspondingly.
Args:
fn (callable|list[callable]|tuple[callable]): The batchify functions to
wrap. It is a callable function or a list/tuple of callable functions.
args (tuple[callable]): The additional batchify functions to wrap.
"""
def __init__(self, fn, *args):
if isinstance(fn, (list, tuple)):
assert len(args) == 0, 'Input pattern not understood. The input of Tuple can be ' \
'Tuple(A, B, C) or Tuple([A, B, C]) or Tuple((A, B, C)). ' \
'Received fn=%s, args=%s' % (str(fn), str(args))
self._fn = fn
else:
self._fn = (fn, ) + args
for i, ele_fn in enumerate(self._fn):
assert callable(
ele_fn
), 'Batchify functions must be callable! type(fn[%d]) = %s' % (
i, str(type(ele_fn)))
def __call__(self, data):
"""
Batchifies data samples by applying each function on the corresponding
data field, and each data field is produced by stacking the field data
of samples.
Args:
data (list|tuple): The samples to batchfy. Each sample in list/tuple
should contain `N` fields.
Returns:
tuple: A tuple composed of results from all including batchifying
functions.
Example:
.. code-block:: python
from paddlenlp.data import Stack, Pad, Tuple
data = [
[[1, 2, 3, 4], [1]],
[[5, 6, 7], [0]],
[[8, 9], [1]],
]
batchify_fn = Tuple(Pad(pad_val=0), Stack())
ids, label = batchify_fn(data)
'''
ids:
[[1, 2, 3, 4],
[5, 6, 7, 0],
[8, 9, 0, 0]]
label: [[1], [0], [1]]
'''
"""
assert len(data[0]) == len(self._fn),\
'The number of attributes in each data sample should contain' \
' {} elements'.format(len(self._fn))
ret = []
for i, ele_fn in enumerate(self._fn):
result = ele_fn([ele[i] for ele in data])
if isinstance(result, (tuple, list)):
ret.extend(result)
else:
ret.append(result)
return tuple(ret)
class Dict(object):
"""
Wraps multiple batchify functions together. The input functions will be
applied to the corresponding input fields.
Each sample should be a dict containing multiple fields. Each batchify
function with key stored in `Dict` will be applied on the field which has
the same key.
For example, when data sample is {'tokens': tokens, 'labels': labels}, you
can wrap two batchify functions using
`Dict({'tokens': DataBatchify, 'labels': LabelBatchify})` to batchify tokens
and labels correspondingly.
Args:
fn (dict): The batchify functions to wrap. It is a dict, which values is
callable functions.
"""
def __init__(self, fn):
assert isinstance(fn, (dict)), 'Input pattern not understood. The input of Dict must be a dict with key of input column name and value of collate_fn ' \
'Received fn=%s' % (str(fn))
self._fn = fn
for col_name, ele_fn in self._fn.items():
assert callable(
ele_fn
), 'Batchify functions must be callable! type(fn[%d]) = %s' % (
col_name, str(type(ele_fn)))
def __call__(self, data):
"""
Batchifies data samples by applying each function on the corresponding
data field, and each data field is produced by stacking the field data
with the same key as batchify functions of all samples.
Args:
data (list[dict]|tuple[dict]): The samples to batchfy. Each sample
in list/tuple is a dict with `N` key-values.
Returns:
tuple: A tuple composed of results from all including batchifying
functions.
Example:
.. code-block:: python
from paddlenlp.data import Stack, Pad, Dict
data = [
{'labels':[1], 'token_ids':[1, 2, 3, 4]},
{'labels':[0], 'token_ids':[5, 6, 7]},
{'labels':[1], 'token_ids':[8, 9]},
]
batchify_fn = Dict({'token_ids':Pad(pad_val=0), 'labels':Stack()})
ids, label = batchify_fn(data)
'''
ids:
[[1, 2, 3, 4],
[5, 6, 7, 0],
[8, 9, 0, 0]]
label: [[1], [0], [1]]
'''
"""
ret = []
for col_name, ele_fn in self._fn.items():
result = ele_fn([ele[col_name] for ele in data])
if isinstance(result, (tuple, list)):
ret.extend(result)
else:
ret.append(result)
return tuple(ret)