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
|
import functools
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
class TestAdadelta(serial.SerializedTestCase):
@staticmethod
def ref_adadelta(param_in,
mom_in,
mom_delta_in,
grad, lr,
epsilon,
decay,
using_fp16=False):
param_in_f32 = param_in
mom_in_f32 = mom_in
mom_delta_in_f32 = mom_delta_in
if(using_fp16):
param_in_f32 = param_in.astype(np.float32)
mom_in_f32 = mom_in.astype(np.float32)
mom_delta_in_f32 = mom_delta_in.astype(np.float32)
mom_out = decay * mom_in_f32 + (1.0 - decay) * grad * grad
new_grad = (np.sqrt(mom_delta_in_f32 + epsilon) /
np.sqrt(mom_out + epsilon)) * grad
param_out = param_in_f32 + lr * new_grad
mom_delta_out = decay * mom_delta_in_f32 + (1.0 - decay
) * new_grad * new_grad
if(using_fp16):
return (param_out.astype(np.float16), mom_out.astype(np.float16),
mom_delta_out.astype(np.float16))
else:
return (param_out.astype(np.float32), mom_out.astype(np.float32),
mom_delta_out.astype(np.float32))
@given(inputs=hu.tensors(n=4),
lr=hu.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
epsilon=hu.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
decay=hu.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
**hu.gcs)
@settings(deadline=10000)
def test_adadelta(self, inputs, lr, epsilon, decay, gc, dc):
param, moment, moment_delta, grad = inputs
moment = np.abs(moment)
moment_delta = np.abs(moment_delta)
lr = np.array([lr], dtype=np.float32)
op = core.CreateOperator(
"Adadelta",
["param", "moment", "moment_delta", "grad", "lr"],
["param", "moment", "moment_delta"],
epsilon=epsilon,
decay=decay,
device_option=gc,
)
self.assertReferenceChecks(
gc, op,
[param, moment, moment_delta, grad, lr],
functools.partial(self.ref_adadelta, epsilon=epsilon, decay=decay))
# 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=4),
lr=hu.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
epsilon=hu.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
decay=hu.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
**hu.gcs)
def test_sparse_adadelta(self, inputs, lr, epsilon, decay, gc, dc):
param, moment, moment_delta, grad = inputs
moment = np.abs(moment)
moment_delta = np.abs(moment_delta)
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(
"SparseAdadelta",
["param", "moment", "moment_delta", "indices", "grad", "lr"],
["param", "moment", "moment_delta"],
epsilon=epsilon,
decay=decay,
device_option=gc)
def ref_sparse(param, moment, moment_delta, indices, grad, lr, decay,
ref_using_fp16):
param_out = np.copy(param)
moment_out = np.copy(moment)
moment_delta_out = np.copy(moment_delta)
for i, index in enumerate(indices):
param_out[index], moment_out[index], moment_delta_out[
index] = self.ref_adadelta(param[index], moment[index],
moment_delta[index], grad[i], lr,
epsilon, decay, ref_using_fp16)
return (param_out, moment_out, moment_delta_out)
ref_using_fp16_values = [False]
if gc == hu.gpu_do:
ref_using_fp16_values.append(True)
for ref_using_fp16 in ref_using_fp16_values:
moment_i = None
moment_delta_i = None
param_i = None
if(ref_using_fp16):
moment_i = moment.astype(np.float16)
moment_delta_i = moment_delta.astype(np.float16)
param_i = param.astype(np.float16)
else:
moment_i = moment.astype(np.float32)
moment_delta_i = moment_delta.astype(np.float32)
param_i = param.astype(np.float32)
self.assertReferenceChecks(gc, op, [
param_i, moment_i, moment_delta_i, indices, grad, lr, decay,
ref_using_fp16
], ref_sparse)
@given(inputs=hu.tensors(n=3),
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),
decay=st.floats(min_value=0.01, max_value=0.99,
allow_nan=False, allow_infinity=False),
**hu.gcs)
@settings(deadline=None)
def test_sparse_adadelta_empty(self, inputs, lr, epsilon, decay, gc, dc):
param, moment, moment_delta = inputs
moment = np.abs(moment)
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(
"SparseAdadelta",
["param", "moment", "moment_delta", "indices", "grad", "lr"],
["param", "moment", "moment_delta"],
epsilon=epsilon,
decay=decay,
device_option=gc)
def ref_sparse_empty(param, moment, moment_delta, indices, grad, lr, decay):
param_out = np.copy(param)
moment_out = np.copy(moment)
moment_delta_out = np.copy(moment_delta)
return (param_out, moment_out, moment_delta_out)
ref_using_fp16_values = [False]
if gc == hu.gpu_do:
ref_using_fp16_values.append(True)
for ref_using_fp16 in ref_using_fp16_values:
moment_i = None
moment_delta_i = None
param_i = None
if(ref_using_fp16):
moment_i = moment.astype(np.float16)
moment_delta_i = moment_delta.astype(np.float16)
param_i = param.astype(np.float16)
else:
moment_i = moment.astype(np.float32)
moment_delta_i = moment_delta.astype(np.float32)
param_i = param.astype(np.float32)
self.assertReferenceChecks(
gc,
op,
[param_i, moment_i, moment_delta_i, indices, grad, lr, decay],
ref_sparse_empty
)
|
