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
|
import functools
import logging
import hypothesis
from hypothesis import given, settings, HealthCheck
import hypothesis.strategies as st
import numpy as np
from caffe2.python import core
import caffe2.python.hypothesis_test_util as hu
import caffe2.python.serialized_test.serialized_test_util as serial
logger = logging.getLogger(__name__)
def ref_wngrad(param_in, seq_b_in, grad, lr, epsilon,
output_effective_lr=False,
output_effective_lr_and_update=False):
# helper functions for wngrad operator test
seq_b_out = seq_b_in + 1.0 / (seq_b_in + epsilon) * np.sum(grad * grad)
effective_lr = lr / (seq_b_in + epsilon)
grad_adj = effective_lr * grad
param_out = param_in + grad_adj
if output_effective_lr_and_update:
return (param_out.astype(np.float32), seq_b_out.astype(np.float32),
effective_lr.astype(np.float32),
grad_adj.astype(np.float32))
elif output_effective_lr:
return (param_out.astype(np.float32), seq_b_out.astype(np.float32),
effective_lr.astype(np.float32))
return (param_out.astype(np.float32), seq_b_out.astype(np.float32))
def wngrad_sparse_test_helper(parent_test, inputs, seq_b, lr, epsilon,
engine, gc, dc):
# helper functions for wngrad operator test
param, grad = inputs
seq_b = np.array([seq_b, ], dtype=np.float32)
lr = np.array([lr], dtype=np.float32)
# Create an indexing array containing values that are lists of indices,
# which index into grad
indices = np.random.choice(np.arange(grad.shape[0]),
size=np.random.randint(grad.shape[0]), replace=False)
# Sparsify grad
grad = grad[indices]
op = core.CreateOperator(
"SparseWngrad",
["param", "seq_b", "indices", "grad", "lr"],
["param", "seq_b"],
epsilon=epsilon,
engine=engine,
device_option=gc)
def ref_sparse(param, seq_b, indices, grad, lr):
param_out = np.copy(param)
seq_b_out = np.copy(seq_b)
seq_b_out = seq_b + 1.0 / seq_b * np.sum(grad * grad)
for i, index in enumerate(indices):
param_out[index] = param[index] + lr / (seq_b + epsilon) * grad[i]
return (param_out, seq_b_out)
logger.info('test_sparse_adagrad with full precision embedding')
seq_b_i = seq_b.astype(np.float32)
param_i = param.astype(np.float32)
parent_test.assertReferenceChecks(
gc, op, [param_i, seq_b_i, indices, grad, lr],
ref_sparse
)
class TestWngrad(serial.SerializedTestCase):
@given(inputs=hu.tensors(n=2),
seq_b=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
lr=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
epsilon=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
**hu.gcs_cpu_only)
@settings(deadline=10000)
def test_wngrad_dense_base(self, inputs, seq_b, lr, epsilon, gc, dc):
param, grad = inputs
seq_b = np.array([seq_b, ], dtype=np.float32)
lr = np.array([lr], dtype=np.float32)
op = core.CreateOperator(
"Wngrad",
["param", "seq_b", "grad", "lr"],
["param", "seq_b"],
epsilon=epsilon,
device_option=gc,
)
self.assertReferenceChecks(
gc, op,
[param, seq_b, grad, lr],
functools.partial(ref_wngrad, epsilon=epsilon))
@given(inputs=hu.tensors(n=2),
seq_b=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
lr=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
epsilon=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
**hu.gcs_cpu_only)
@settings(deadline=10000)
def test_wngrad_dense_output_effective_lr(self, inputs, seq_b,
lr, epsilon, gc, dc):
param, grad = inputs
seq_b = np.array([seq_b, ], dtype=np.float32)
lr = np.array([lr], dtype=np.float32)
op = core.CreateOperator(
"Wngrad",
["param", "seq_b", "grad", "lr"],
["param", "seq_b", "effective_lr"],
epsilon=epsilon,
device_option=gc,
)
self.assertReferenceChecks(
gc, op,
[param, seq_b, grad, lr],
functools.partial(ref_wngrad, epsilon=epsilon,
output_effective_lr=True))
@given(inputs=hu.tensors(n=2),
seq_b=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
lr=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
epsilon=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
**hu.gcs_cpu_only)
@settings(deadline=10000)
def test_wngrad_dense_output_effective_lr_and_update(
self, inputs, seq_b, lr, epsilon, gc, dc):
param, grad = inputs
seq_b = np.abs(np.array([seq_b, ], dtype=np.float32))
lr = np.array([lr], dtype=np.float32)
op = core.CreateOperator(
"Wngrad",
["param", "seq_b", "grad", "lr"],
["param", "seq_b", "effective_lr", "update"],
epsilon=epsilon,
device_option=gc,
)
self.assertReferenceChecks(
gc, op,
[param, seq_b, grad, lr],
functools.partial(ref_wngrad, epsilon=epsilon,
output_effective_lr_and_update=True))
# Suppress filter_too_much health check.
# Likely caused by `assume` call falling through too often.
@settings(suppress_health_check=[HealthCheck.filter_too_much], deadline=10000)
@given(inputs=hu.tensors(n=2),
seq_b=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
lr=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
epsilon=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
**hu.gcs_cpu_only)
def test_sparse_wngrad(self, inputs, seq_b, lr, epsilon, gc, dc):
return wngrad_sparse_test_helper(self, inputs, seq_b, lr, epsilon,
None, gc, dc)
@given(inputs=hu.tensors(n=1),
lr=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
seq_b=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
epsilon=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
**hu.gcs_cpu_only)
@settings(deadline=10000)
def test_sparse_wngrad_empty(self, inputs, seq_b, lr, epsilon, gc, dc):
param = inputs[0]
seq_b = np.array([seq_b, ], dtype=np.float32)
lr = np.array([lr], dtype=np.float32)
grad = np.empty(shape=(0,) + param.shape[1:], dtype=np.float32)
indices = np.empty(shape=(0,), dtype=np.int64)
hypothesis.note('indices.shape: %s' % str(indices.shape))
op = core.CreateOperator(
"SparseWngrad",
["param", "seq_b", "indices", "grad", "lr"],
["param", "seq_b"],
epsilon=epsilon,
device_option=gc)
def ref_sparse(param, seq_b, indices, grad, lr):
param_out = np.copy(param)
seq_b_out = np.copy(seq_b)
return (param_out, seq_b_out)
print('test_sparse_adagrad_empty with full precision embedding')
seq_b_i = seq_b.astype(np.float32)
param_i = param.astype(np.float32)
self.assertReferenceChecks(
gc, op, [param_i, seq_b_i, indices, grad, lr], ref_sparse
)
|
