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
|
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Tests for PowerSign."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import math
import numpy as np
from tensorflow.compiler.tests import xla_test
from tensorflow.contrib.opt.python.training import powersign
from tensorflow.contrib.opt.python.training import sign_decay
from tensorflow.python.framework import constant_op
from tensorflow.python.ops import resource_variable_ops
from tensorflow.python.ops import variables
from tensorflow.python.platform import test
def py_linear_decay_fn(decay_steps):
def linear_decay(step):
step = min(step, decay_steps)
return float(decay_steps - step) / decay_steps
return linear_decay
def powersign_update_numpy(params,
g_t,
m,
lr,
base=math.e,
beta=0.9,
py_sign_decay_fn=None,
t=None):
m_t = beta * m + (1 - beta) * g_t
if py_sign_decay_fn is None:
sign_decayed = 1.0
else:
sign_decayed = py_sign_decay_fn(t-1)
multiplier = base ** (sign_decayed * np.sign(g_t) * np.sign(m_t))
params_t = params - lr * multiplier * g_t
return params_t, m_t
class PowerSignTest(xla_test.XLATestCase):
def _testDense(self,
learning_rate=0.1,
sign_decay_fn=None,
py_sign_decay_fn=None,
base=math.e,
beta=0.9):
for dtype in self.float_types:
with self.cached_session(), self.test_scope():
# Initialize variables for numpy implementation.
m0, m1 = 0.0, 0.0
var0_np = np.array([1.0, 2.0], dtype=dtype)
grads0_np = np.array([0.1, 0.1], dtype=dtype)
var1_np = np.array([3.0, 4.0], dtype=dtype)
grads1_np = np.array([0.01, 0.01], dtype=dtype)
var0 = resource_variable_ops.ResourceVariable(var0_np)
var1 = resource_variable_ops.ResourceVariable(var1_np)
global_step = resource_variable_ops.ResourceVariable(0, trainable=False)
grads0 = constant_op.constant(grads0_np)
grads1 = constant_op.constant(grads1_np)
opt = powersign.PowerSignOptimizer(
learning_rate=learning_rate,
base=base,
beta=beta,
sign_decay_fn=sign_decay_fn,
)
update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]),
global_step=global_step)
neg_update = opt.apply_gradients(zip([-grads0, -grads1], [var0, var1]),
global_step=global_step)
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0], var0.eval())
self.assertAllClose([3.0, 4.0], var1.eval())
# Run 7 steps of powersign
# first 4 steps with positive gradient
# last 3 steps with negative gradient (sign(gm) should be -1)
for t in range(1, 8):
if t < 5:
update.run()
else:
neg_update.run()
var0_np, m0 = powersign_update_numpy(
var0_np,
grads0_np if t < 5 else -grads0_np,
m0,
learning_rate,
base=base,
beta=beta,
py_sign_decay_fn=py_sign_decay_fn,
t=t,
)
var1_np, m1 = powersign_update_numpy(
var1_np,
grads1_np if t < 5 else -grads1_np,
m1,
learning_rate,
base=base,
beta=beta,
py_sign_decay_fn=py_sign_decay_fn,
t=t,
)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np, var0.eval())
self.assertAllCloseAccordingToType(var1_np, var1.eval())
def testDense(self):
decay_steps = 10
sign_decay_fn = sign_decay.get_linear_decay_fn(decay_steps)
py_sign_decay_fn = py_linear_decay_fn(decay_steps)
self._testDense()
self._testDense(learning_rate=0.1, base=10.0, beta=0.8)
self._testDense(
sign_decay_fn=sign_decay_fn, py_sign_decay_fn=py_sign_decay_fn)
if __name__ == '__main__':
test.main()
|
