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
|
from caffe2.python import core
import caffe2.python.hypothesis_test_util as hu
import caffe2.python.serialized_test.serialized_test_util as serial
from hypothesis import given, settings
import hypothesis.strategies as st
import numpy as np
import unittest
@st.composite
def _tensor_splits(draw, add_axis=False):
"""Generates (axis, split_info, tensor_splits) tuples."""
tensor = draw(hu.tensor(min_value=4)) # Each dim has at least 4 elements.
axis = draw(st.integers(-len(tensor.shape), len(tensor.shape) - 1))
if add_axis:
# Simple case: get individual slices along one axis, where each of them
# is (N-1)-dimensional. The axis will be added back upon concatenation.
return (
axis,
np.ones(tensor.shape[axis], dtype=np.int32),
[
np.array(tensor.take(i, axis=axis))
for i in range(tensor.shape[axis])
]
)
else:
# General case: pick some (possibly consecutive, even non-unique)
# indices at which we will split the tensor, along the given axis.
splits = sorted(draw(
st.lists(elements=st.integers(0, tensor.shape[axis]), max_size=4)
) + [0, tensor.shape[axis]])
return (
axis,
np.array(np.diff(splits), dtype=np.int32),
[
tensor.take(range(splits[i], splits[i + 1]), axis=axis)
for i in range(len(splits) - 1)
],
)
class TestConcatSplitOps(serial.SerializedTestCase):
@serial.given(tensor_splits=_tensor_splits(),
**hu.gcs)
def test_concat(self, tensor_splits, gc, dc):
axis, _, splits = tensor_splits
op = core.CreateOperator(
"Concat",
['X_{}'.format(i) for i in range(len(splits))],
['concat_result', 'split_info'],
axis=axis
)
self.assertReferenceChecks(
gc, op, splits, lambda *splits: (
np.concatenate(splits, axis=axis),
np.array([a.shape[axis] for a in splits])
),
ensure_outputs_are_inferred=True,
)
self.assertDeviceChecks(dc, op, splits, [0, 1])
self.assertGradientChecks(
gc, op, splits, 0, [0],
ensure_outputs_are_inferred=True,
)
@given(tensor_splits=_tensor_splits(add_axis=True),
**hu.gcs)
@settings(deadline=10000)
def test_concat_add_axis(self, tensor_splits, gc, dc):
axis, _, splits = tensor_splits
op = core.CreateOperator(
"Concat",
['X_{}'.format(i) for i in range(len(splits))],
['concat_result', 'split_info'],
axis=axis,
add_axis=1
)
self.assertReferenceChecks(
gc, op, splits, lambda *splits: (
np.concatenate(
[np.expand_dims(a, axis) for a in splits],
axis=axis
),
np.array([1] * len(splits))
),
ensure_outputs_are_inferred=True,
)
self.assertDeviceChecks(dc, op, splits, [0, 1])
for i in range(len(splits)):
self.assertGradientChecks(
gc, op, splits, i, [0],
ensure_outputs_are_inferred=True,
)
@serial.given(tensor_splits=_tensor_splits(),
split_as_arg=st.booleans(),
**hu.gcs)
def test_split(self, tensor_splits, split_as_arg, gc, dc):
axis, split_info, splits = tensor_splits
split_as_arg = True
if split_as_arg:
input_names = ['input']
input_tensors = [np.concatenate(splits, axis=axis)]
kwargs = dict(axis=axis, split=split_info)
else:
input_names = ['input', 'split']
input_tensors = [np.concatenate(splits, axis=axis), split_info]
kwargs = dict(axis=axis)
op = core.CreateOperator(
"Split",
input_names,
['X_{}'.format(i) for i in range(len(split_info))],
**kwargs
)
def split_ref(input, split=split_info):
s = np.cumsum([0] + list(split))
return [
np.array(input.take(np.arange(s[i], s[i + 1]), axis=axis))
for i in range(len(split))
]
outputs_with_grad = range(len(split_info))
self.assertReferenceChecks(
gc, op, input_tensors, split_ref,
ensure_outputs_are_inferred=True,
)
self.assertDeviceChecks(dc, op, input_tensors, outputs_with_grad)
self.assertGradientChecks(
gc, op, input_tensors, 0, outputs_with_grad,
ensure_outputs_are_inferred=True,
)
@given(
inputs=hu.lengths_tensor(
dtype=np.float32,
min_value=1,
max_value=11,
allow_empty=True,
),
split_by_scaling_lengths=st.booleans(),
**hu.gcs
)
@settings(deadline=10000)
def test_split_by_lengths(self, inputs, split_by_scaling_lengths, gc, dc):
data, lengths = inputs
len_len = len(lengths)
def _find_factor_simple(x):
for i in [2, 3, 5, 7, 9, 11]:
if x % i == 0:
return i
return x
num_output = _find_factor_simple(len_len)
scaling_factor = 1
if split_by_scaling_lengths:
sum_len = sum(lengths)
sum_scaling_lengths = _find_factor_simple(sum_len)
if sum_scaling_lengths != sum_len and sum_scaling_lengths >= num_output:
scaling_lengths = [1] * (num_output - 1) + [sum_scaling_lengths - num_output + 1]
len_len = len(scaling_lengths)
lengths = np.array(scaling_lengths, dtype=np.int32)
scaling_factor = (sum_len // sum_scaling_lengths) if sum_scaling_lengths else 1
axis = 0
op = core.CreateOperator(
"SplitByLengths",
["data", "lengths"],
['X_{}'.format(i) for i in range(num_output)],
axis=axis,
use_scaling_lengths=split_by_scaling_lengths,
)
def split_by_lengths_ref(data, lengths, num_output=num_output, axis=0):
idxs = np.cumsum([0] + list(lengths)).astype(np.int32)
return [
np.array(
data.take(
np.arange(
scaling_factor * idxs[i * len_len // num_output],
scaling_factor * idxs[(i + 1) * len_len // num_output]
),
axis=axis
)
) for i in range(num_output)
]
outputs_with_grad = range(num_output)
input_tensors = [data, lengths]
self.assertReferenceChecks(
hu.cpu_do, op, input_tensors, split_by_lengths_ref)
self.assertDeviceChecks(dc, op, input_tensors, outputs_with_grad)
self.assertGradientChecks(
hu.cpu_do, op, input_tensors, 0, outputs_with_grad,
input_device_options={"lengths": hu.cpu_do})
if __name__ == "__main__":
unittest.main()
|
