forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
boolean_mask_test.py
405 lines (358 loc) · 16.1 KB
/
boolean_mask_test.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
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from caffe2.proto import caffe2_pb2
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 assume, given, settings
import hypothesis.strategies as st
import numpy as np
class TestBooleanMaskOp(serial.SerializedTestCase):
@given(x=hu.tensor1d(min_len=1,
max_len=100,
elements=hu.floats(min_value=0.5, max_value=1.0)),
**hu.gcs_cpu_only)
@settings(deadline=1000)
def test_boolean_mask_gradient(self, x, gc, dc):
op = core.CreateOperator("BooleanMask",
["data", "mask"],
"masked_data")
mask = np.random.choice(a=[True, False], size=x.shape[0])
expected_gradient = np.copy(mask).astype(int)
self.assertDeviceChecks(dc, op, [x, mask], [0])
self.assertGradientChecks(gc, op, [x, mask], 0, [0])
@given(x=hu.tensor1d(min_len=1,
max_len=5,
elements=hu.floats(min_value=0.5, max_value=1.0)),
**hu.gcs)
@settings(deadline=1000)
def test_boolean_mask(self, x, gc, dc):
op = core.CreateOperator("BooleanMask",
["data", "mask"],
"masked_data")
mask = np.random.choice(a=[True, False], size=x.shape[0])
def ref(x, mask):
return (x[mask],)
self.assertReferenceChecks(gc, op, [x, mask], ref)
self.assertDeviceChecks(dc, op, [x, mask], [0])
@given(x=hu.tensor1d(min_len=1,
max_len=5,
elements=hu.floats(min_value=0.5, max_value=1.0)),
**hu.gcs)
def test_boolean_mask_indices(self, x, gc, dc):
op = core.CreateOperator("BooleanMask",
["data", "mask"],
["masked_data", "masked_indices"])
mask = np.random.choice(a=[True, False], size=x.shape[0])
def ref(x, mask):
return (x[mask], np.where(mask)[0])
self.assertReferenceChecks(gc, op, [x, mask], ref)
self.assertDeviceChecks(dc, op, [x, mask], [0])
@staticmethod
def _dtype_conversion(x, dtype, gc, dc):
"""SequenceMask only supports fp16 with CUDA/ROCm."""
if dtype == np.float16:
assume(core.IsGPUDeviceType(gc.device_type))
dc = [d for d in dc if core.IsGPUDeviceType(d.device_type)]
x = x.astype(dtype)
return x, dc
@given(x=hu.tensor(min_dim=2,
max_dim=5,
elements=hu.floats(min_value=0.5, max_value=1.0)),
dtype=st.sampled_from([np.float32, np.float16]),
**hu.gcs)
def test_sequence_mask_with_lengths(self, x, dtype, gc, dc):
x, dc = self._dtype_conversion(x, dtype, gc, dc)
# finite fill value needed for gradient check
fill_val = 1e-3 if dtype == np.float16 else 1e-9
op = core.CreateOperator("SequenceMask",
["data", "lengths"],
["masked_data"],
mode="sequence",
axis=len(x.shape) - 1,
fill_val=fill_val)
elem_dim = x.shape[-1]
leading_dim = 1
for dim in x.shape[:-1]:
leading_dim *= dim
lengths = np.random.randint(0, elem_dim, [leading_dim])\
.astype(np.int32)
def ref(x, lengths):
ref = np.reshape(x, [leading_dim, elem_dim])
for i in range(leading_dim):
for j in range(elem_dim):
if j >= lengths[i]:
ref[i, j] = fill_val
return [ref.reshape(x.shape)]
self.assertReferenceChecks(gc, op, [x, lengths], ref)
self.assertDeviceChecks(dc, op, [x, lengths], [0])
@given(x=hu.tensor(min_dim=2,
max_dim=5,
elements=hu.floats(min_value=0.5, max_value=1.0)),
dtype=st.sampled_from([np.float32, np.float16]),
**hu.gcs)
@settings(deadline=10000)
def test_sequence_mask_with_window(self, x, dtype, gc, dc):
x, dc = self._dtype_conversion(x, dtype, gc, dc)
# finite fill value needed for gradient check
fill_val = 1e-3 if dtype == np.float16 else 1e-9
radius = 2
op = core.CreateOperator("SequenceMask",
["data", "centers"],
["masked_data"],
mode="window",
radius=radius,
axis=len(x.shape) - 1,
fill_val=fill_val)
elem_dim = x.shape[-1]
leading_dim = 1
for dim in x.shape[:-1]:
leading_dim *= dim
centers = np.random.randint(0, elem_dim, [leading_dim])\
.astype(np.int32)
def ref(x, centers):
ref = np.reshape(x, [leading_dim, elem_dim])
for i in range(leading_dim):
for j in range(elem_dim):
if j > centers[i] + radius or j < centers[i] - radius:
ref[i, j] = fill_val
return [ref.reshape(x.shape)]
self.assertReferenceChecks(gc, op, [x, centers], ref)
self.assertDeviceChecks(dc, op, [x, centers], [0])
# Gradient check with np.float16 is found to be flakey, disable for now
# with high threshold (to repro, set threshold to 0.4).
threshold = 1.0 if dtype == np.float16 else 0.005
self.assertGradientChecks(gc, op, [x, centers], 0, [0],
threshold=threshold)
@given(x=hu.tensor(min_dim=2,
max_dim=5,
elements=hu.floats(min_value=0.5, max_value=1.0)),
mode=st.sampled_from(['upper', 'lower', 'upperdiag', 'lowerdiag']),
dtype=st.sampled_from([np.float32, np.float16]),
**hu.gcs)
@settings(deadline=10000)
def test_sequence_mask_triangle(self, x, mode, dtype, gc, dc):
x, dc = self._dtype_conversion(x, dtype, gc, dc)
# finite fill value needed for gradient check
fill_val = 1e-3 if dtype == np.float16 else 1e-9
op = core.CreateOperator("SequenceMask",
["data"],
["masked_data"],
mode=mode,
axis=len(x.shape) - 1,
fill_val=fill_val)
elem_dim = x.shape[-1]
leading_dim = 1
for dim in x.shape[:-1]:
leading_dim *= dim
if mode == 'upper':
def compare(i, j):
return j > i
elif mode == 'lower':
def compare(i, j):
return j < i
elif mode == 'upperdiag':
def compare(i, j):
return j >= i
elif mode == 'lowerdiag':
def compare(i, j):
return j <= i
def ref(x):
ref = np.reshape(x, [leading_dim, elem_dim])
for i in range(leading_dim):
for j in range(elem_dim):
if compare(i, j):
ref[i, j] = fill_val
return [ref.reshape(x.shape)]
self.assertReferenceChecks(gc, op, [x], ref)
self.assertDeviceChecks(dc, op, [x], [0])
# Gradient check with np.float16 is found to be flakey, disable for now
# with high threshold (to repro, set threshold to 0.4).
threshold = 1.0 if dtype == np.float16 else 0.005
stepsize = 0.1 if dtype == np.float16 else 0.05
self.assertGradientChecks(gc, op, [x], 0, [0],
threshold=threshold, stepsize=stepsize)
@given(x=hu.tensor(min_dim=2,
max_dim=5,
elements=hu.floats(min_value=0.5, max_value=1.0)),
dtype=st.sampled_from([np.float32, np.float16]),
**hu.gcs)
@settings(deadline=10000)
def test_sequence_mask_batching_lengths(self, x, dtype, gc, dc):
x, dc = self._dtype_conversion(x, dtype, gc, dc)
# finite fill value needed for gradient check
fill_val = 1e-3 if dtype == np.float16 else 1e-9
# choose _different_ batch and axis dimensions, w/ axis != 0.
axis = 0
batch = 0
while axis == 0 or axis < batch:
inds = np.arange(len(x.shape))
np.random.shuffle(inds)
batch = inds[0]
axis = inds[1]
op = core.CreateOperator("SequenceMask",
["data", "lengths"],
["masked_data"],
mode='sequence',
axis=axis,
fill_val=fill_val,
batch=batch)
before = int(np.prod(x.shape[:batch + 1]))
between = int(np.prod(x.shape[batch + 1:axis]))
after = int(np.prod(x.shape[axis:]))
lengths = np.random.randint(0, after, [between])\
.astype(np.int32)
def ref(z, l):
w = np.reshape(z, [before, between, after])
for b in range(before):
r = w[b, :, :]
for i in range(between):
for j in range(after):
if j >= l[i]:
r[i, j] = fill_val
return [w.reshape(z.shape)]
self.assertReferenceChecks(gc, op, [x, lengths], ref)
self.assertDeviceChecks(dc, op, [x, lengths], [0])
# Gradient check with np.float16 is found to be flakey, disable for now
# with high threshold (to repro, set threshold to 0.4).
threshold = 1.0 if dtype == np.float16 else 0.005
self.assertGradientChecks(gc, op, [x, lengths], 0, [0],
threshold=threshold)
@given(x=hu.tensor(min_dim=4,
max_dim=4,
elements=hu.floats(min_value=0.5, max_value=1.0)),
dtype=st.sampled_from([np.float32, np.float16]),
**hu.gcs)
@settings(deadline=10000)
def test_sequence_mask_batching_window(self, x, dtype, gc, dc):
x, dc = self._dtype_conversion(x, dtype, gc, dc)
# finite fill value needed for gradient check
fill_val = 1e-3 if dtype == np.float16 else 1e-9
radius = 1
# choose _different_ batch and axis dimensions, w/ axis != 0.
axis = 0
batch = 0
while axis == 0 or axis < batch:
inds = np.arange(len(x.shape))
np.random.shuffle(inds)
batch = inds[0]
axis = inds[1]
op = core.CreateOperator("SequenceMask",
["data", "centers"],
["masked_data"],
mode='window',
radius=radius,
axis=axis,
fill_val=fill_val,
batch=batch)
before = int(np.prod(x.shape[:batch + 1]))
between = int(np.prod(x.shape[batch + 1:axis]))
after = int(np.prod(x.shape[axis:]))
centers = np.random.randint(0, after, [between])\
.astype(np.int32)
def ref(z, c):
w = np.reshape(z, [before, between, after])
for b in range(before):
r = w[b, :, :]
for i in range(between):
for j in range(after):
if j > c[i] + radius or j < c[i] - radius:
r[i, j] = fill_val
return [w.reshape(z.shape)]
self.assertReferenceChecks(gc, op, [x, centers], ref)
self.assertDeviceChecks(dc, op, [x, centers], [0])
# Gradient check with np.float16 is found to be flakey, disable for now
# with high threshold (to repro, set threshold to 0.4).
threshold = 1.0 if dtype == np.float16 else 0.005
self.assertGradientChecks(gc, op, [x, centers], 0, [0],
threshold=threshold)
@given(x=hu.tensor(min_dim=3,
max_dim=5,
elements=hu.floats(min_value=0.5, max_value=1.0)),
mode=st.sampled_from(['upper', 'lower', 'upperdiag', 'lowerdiag']),
dtype=st.sampled_from([np.float32, np.float16]),
**hu.gcs)
@settings(deadline=10000)
def test_sequence_mask_batching_triangle(self, x, mode, dtype, gc, dc):
x, dc = self._dtype_conversion(x, dtype, gc, dc)
# finite fill value needed for gradient check
fill_val = 1e-3 if dtype == np.float16 else 1e-9
# choose _different_ batch and axis dimensions, w/ axis != 0.
axis = 0
batch = 0
while axis == 0 or axis < batch:
inds = np.arange(len(x.shape))
np.random.shuffle(inds)
batch = inds[0]
axis = inds[1]
op = core.CreateOperator("SequenceMask",
["data"],
["masked_data"],
mode=mode,
axis=axis,
fill_val=fill_val,
batch=batch)
if mode == 'upper':
def compare(i, j):
return j > i
elif mode == 'lower':
def compare(i, j):
return j < i
elif mode == 'upperdiag':
def compare(i, j):
return j >= i
elif mode == 'lowerdiag':
def compare(i, j):
return j <= i
def ref(z):
before = int(np.prod(z.shape[:batch + 1]))
between = int(np.prod(z.shape[batch + 1:axis]))
after = int(np.prod(z.shape[axis:]))
w = np.reshape(z, [before, between, after])
for b in range(before):
r = w[b, :, :]
for i in range(between):
for j in range(after):
if compare(i, j):
r[i, j] = fill_val
return [w.reshape(z.shape)]
self.assertReferenceChecks(gc, op, [x], ref)
self.assertDeviceChecks(dc, op, [x], [0])
# Gradient check with np.float16 is found to be flakey, disable for now
# with high threshold (to repro, set threshold to 0.4).
threshold = 1.0 if dtype == np.float16 else 0.005
stepsize = 0.1 if dtype == np.float16 else 0.05
self.assertGradientChecks(gc, op, [x], 0, [0],
threshold=threshold, stepsize=stepsize)
@given(x=hu.tensor(min_dim=3,
max_dim=5,
elements=hu.floats(min_value=0.5, max_value=1.0)),
dtype=st.sampled_from([np.float32, np.float16]),
**hu.gcs)
def test_sequence_mask_repeated(self, x, dtype, gc, dc):
x, dc = self._dtype_conversion(x, dtype, gc, dc)
# finite fill value needed for gradient check
fill_val = 1e-3 if dtype == np.float16 else 1e-9
op = core.CreateOperator("SequenceMask",
["data", "lengths"],
["masked_data"],
mode="sequence",
axis=len(x.shape) - 2,
repeat_from_axis=-1,
fill_val=fill_val)
elem_dim = x.shape[-2]
leading_dim = 1
for dim in x.shape[:-2]:
leading_dim *= dim
lengths = np.random.randint(0, elem_dim, [leading_dim])\
.astype(np.int32)
def ref(x, lengths):
ref = np.reshape(x, [leading_dim, elem_dim, -1])
for i in range(leading_dim):
for j in range(elem_dim):
if j >= lengths[i]:
ref[i, j, :] = fill_val
return [ref.reshape(x.shape)]
self.assertReferenceChecks(gc, op, [x, lengths], ref)
self.assertDeviceChecks(dc, op, [x, lengths], [0])