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
|
import caffe2.python.hypothesis_test_util as hu
import caffe2.python.serialized_test.serialized_test_util as serial
import numpy as np
from caffe2.python import core, workspace
from hypothesis import given, settings, strategies as st
def batched_boarders_and_data(
data_min_size=5,
data_max_size=10,
examples_min_number=1,
examples_max_number=4,
example_min_size=1,
example_max_size=3,
dtype=np.float32,
elements=None,
):
dims_ = st.tuples(
st.integers(min_value=data_min_size, max_value=data_max_size),
st.integers(min_value=examples_min_number, max_value=examples_max_number),
st.integers(min_value=example_min_size, max_value=example_max_size),
)
return dims_.flatmap(
lambda dims: st.tuples(
hu.arrays(
[dims[1], dims[2], 2],
dtype=np.int32,
elements=st.integers(min_value=0, max_value=dims[0]),
),
hu.arrays([dims[0]], dtype, elements),
)
)
@st.composite
def _tensor_splits(draw):
lengths = draw(st.lists(st.integers(1, 5), min_size=1, max_size=10))
batch_size = draw(st.integers(1, 5))
element_pairs = [
(batch, r) for batch in range(batch_size) for r in range(len(lengths))
]
perm = draw(st.permutations(element_pairs))
perm = perm[:-1] # skip one range
ranges = [[(0, 0)] * len(lengths) for _ in range(batch_size)]
offset = 0
for pair in perm:
ranges[pair[0]][pair[1]] = (offset, lengths[pair[1]])
offset += lengths[pair[1]]
data = draw(
st.lists(
st.floats(min_value=-1.0, max_value=1.0), min_size=offset, max_size=offset
)
)
key = draw(st.permutations(range(offset)))
return (
np.array(data).astype(np.float32),
np.array(ranges),
np.array(lengths),
np.array(key).astype(np.int64),
)
@st.composite
def _bad_tensor_splits(draw):
lengths = draw(st.lists(st.integers(4, 6), min_size=4, max_size=4))
batch_size = 4
element_pairs = [
(batch, r) for batch in range(batch_size) for r in range(len(lengths))
]
perm = draw(st.permutations(element_pairs))
ranges = [[(0, 0)] * len(lengths) for _ in range(batch_size)]
offset = 0
# Inject some bad samples depending on the batch.
# Batch 2: length is set to 0. This way, 25% of the samples are empty.
# Batch 0-1: length is set to half the original length. This way, 50% of the
# samples are of mismatched length.
for pair in perm:
if pair[0] == 2:
length = 0
elif pair[0] <= 1:
length = lengths[pair[1]] // 2
else:
length = lengths[pair[1]]
ranges[pair[0]][pair[1]] = (offset, length)
offset += length
data = draw(
st.lists(
st.floats(min_value=-1.0, max_value=1.0), min_size=offset, max_size=offset
)
)
key = draw(st.permutations(range(offset)))
return (
np.array(data).astype(np.float32),
np.array(ranges),
np.array(lengths),
np.array(key).astype(np.int64),
)
def gather_ranges(data, ranges):
lengths = []
output = []
for example_ranges in ranges:
length = 0
for range in example_ranges:
assert len(range) == 2
output.extend(data[range[0] : range[0] + range[1]])
length += range[1]
lengths.append(length)
return output, lengths
def gather_ranges_to_dense(data, ranges, lengths):
outputs = []
assert len(ranges)
batch_size = len(ranges)
assert len(ranges[0])
num_ranges = len(ranges[0])
assert ranges.shape[2] == 2
for i in range(num_ranges):
out = []
for j in range(batch_size):
start, length = ranges[j][i]
if not length:
out.append([0] * lengths[i])
else:
assert length == lengths[i]
out.append(data[start : start + length])
outputs.append(np.array(out))
return outputs
def gather_ranges_to_dense_with_key(data, ranges, key, lengths):
outputs = []
assert len(ranges)
batch_size = len(ranges)
assert len(ranges[0])
num_ranges = len(ranges[0])
assert ranges.shape[2] == 2
for i in range(num_ranges):
out = []
for j in range(batch_size):
start, length = ranges[j][i]
if not length:
out.append([0] * lengths[i])
else:
assert length == lengths[i]
key_data_list = zip(
key[start : start + length], data[start : start + length]
)
sorted_key_data_list = sorted(key_data_list, key=lambda x: x[0])
sorted_data = [d for (k, d) in sorted_key_data_list]
out.append(sorted_data)
outputs.append(np.array(out))
return outputs
class TestGatherRanges(serial.SerializedTestCase):
@given(boarders_and_data=batched_boarders_and_data(), **hu.gcs_cpu_only)
@settings(deadline=10000)
def test_gather_ranges(self, boarders_and_data, gc, dc):
boarders, data = boarders_and_data
def boarders_to_range(boarders):
assert len(boarders) == 2
boarders = sorted(boarders)
return [boarders[0], boarders[1] - boarders[0]]
ranges = np.apply_along_axis(boarders_to_range, 2, boarders)
self.assertReferenceChecks(
device_option=gc,
op=core.CreateOperator(
"GatherRanges", ["data", "ranges"], ["output", "lengths"]
),
inputs=[data, ranges],
reference=gather_ranges,
)
@given(tensor_splits=_tensor_splits(), **hu.gcs_cpu_only)
@settings(deadline=10000)
def test_gather_ranges_split(self, tensor_splits, gc, dc):
data, ranges, lengths, _ = tensor_splits
self.assertReferenceChecks(
device_option=gc,
op=core.CreateOperator(
"GatherRangesToDense",
["data", "ranges"],
["X_{}".format(i) for i in range(len(lengths))],
lengths=lengths,
),
inputs=[data, ranges, lengths],
reference=gather_ranges_to_dense,
)
@given(tensor_splits=_tensor_splits(), **hu.gcs_cpu_only)
def test_gather_ranges_with_key_split(self, tensor_splits, gc, dc):
data, ranges, lengths, key = tensor_splits
self.assertReferenceChecks(
device_option=gc,
op=core.CreateOperator(
"GatherRangesToDense",
["data", "ranges", "key"],
["X_{}".format(i) for i in range(len(lengths))],
lengths=lengths,
),
inputs=[data, ranges, key, lengths],
reference=gather_ranges_to_dense_with_key,
)
def test_shape_and_type_inference(self):
with hu.temp_workspace("shape_type_inf_int32"):
net = core.Net("test_net")
net.ConstantFill([], "ranges", shape=[3, 5, 2], dtype=core.DataType.INT32)
net.ConstantFill([], "values", shape=[64], dtype=core.DataType.INT64)
net.GatherRanges(["values", "ranges"], ["values_output", "lengths_output"])
(shapes, types) = workspace.InferShapesAndTypes([net], {})
self.assertEqual(shapes["values_output"], [64])
self.assertEqual(types["values_output"], core.DataType.INT64)
self.assertEqual(shapes["lengths_output"], [3])
self.assertEqual(types["lengths_output"], core.DataType.INT32)
@given(tensor_splits=_bad_tensor_splits(), **hu.gcs_cpu_only)
@settings(deadline=10000)
def test_empty_range_check(self, tensor_splits, gc, dc):
data, ranges, lengths, key = tensor_splits
workspace.FeedBlob("data", data)
workspace.FeedBlob("ranges", ranges)
workspace.FeedBlob("key", key)
def getOpWithThreshold(
min_observation=2, max_mismatched_ratio=0.5, max_empty_ratio=None
):
return core.CreateOperator(
"GatherRangesToDense",
["data", "ranges", "key"],
["X_{}".format(i) for i in range(len(lengths))],
lengths=lengths,
min_observation=min_observation,
max_mismatched_ratio=max_mismatched_ratio,
max_empty_ratio=max_empty_ratio,
)
workspace.RunOperatorOnce(getOpWithThreshold())
workspace.RunOperatorOnce(
getOpWithThreshold(max_mismatched_ratio=0.3, min_observation=50)
)
with self.assertRaises(RuntimeError):
workspace.RunOperatorOnce(
getOpWithThreshold(max_mismatched_ratio=0.3, min_observation=5)
)
with self.assertRaises(RuntimeError):
workspace.RunOperatorOnce(
getOpWithThreshold(min_observation=50, max_empty_ratio=0.01)
)
if __name__ == "__main__":
import unittest
unittest.main()
|
