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
|
# Copyright 2016 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.
# ==============================================================================
"""Functional tests for AdagradDA operations."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes
from tensorflow.python.ops import embedding_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import resource_variable_ops
from tensorflow.python.ops import variables
from tensorflow.python.platform import test
from tensorflow.python.training import adagrad_da
class AdagradDAOptimizerTest(test.TestCase):
def doTestAdagradDAwithoutRegularizationBasic1(self, use_resource=False):
for dtype in [dtypes.float64, dtypes.float32]:
with self.cached_session() as sess:
global_step = variables.Variable(0, dtype=dtypes.int64)
if use_resource:
var0 = resource_variable_ops.ResourceVariable([0.0, 0.0], dtype=dtype)
var1 = resource_variable_ops.ResourceVariable([0.0, 0.0], dtype=dtype)
else:
var0 = variables.Variable([0.0, 0.0], dtype=dtype)
var1 = variables.Variable([0.0, 0.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.2], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.02], dtype=dtype)
opt = adagrad_da.AdagradDAOptimizer(
3.0,
global_step,
initial_gradient_squared_accumulator_value=0.1,
l1_regularization_strength=0.0,
l2_regularization_strength=0.0)
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]), global_step=global_step)
variables.global_variables_initializer().run()
v0_val, v1_val = sess.run([var0, var1])
self.assertAllClose([0.0, 0.0], v0_val)
self.assertAllClose([0.0, 0.0], v1_val)
# Run a step of AdagradDA
update.run()
v0_val, v1_val = sess.run([var0, var1])
# Let g to be gradient accumulator, gg to be gradient squared
# accumulator, T be the global step, lr is the learning rate, and k the
# initial gradient squared accumulator value.
# w = \dfrac{sign(-g)*lr*|g - l1*T|_{+}}{l2*T*lr + \sqrt{k+gg})}
# For -0.1*3.0*(0.1 - 0)/(0 + sqrt(0.1 + 0.1*0.1)) = -0.904534
# similarly for others.
self.assertAllCloseAccordingToType(
np.array([-0.904534, -1.603567]), v0_val)
self.assertAllCloseAccordingToType(
np.array([-0.094821, -0.189358]), v1_val)
def testAdagradDAWithoutRegularizationBasic1(self):
self.doTestAdagradDAwithoutRegularizationBasic1()
def testResourceAdagradDAWithoutRegularizationBasic1(self):
self.doTestAdagradDAwithoutRegularizationBasic1(use_resource=True)
def testMinimizeSparseResourceVariable(self):
for dtype in [dtypes.float32, dtypes.float64]:
with self.cached_session():
var0 = resource_variable_ops.ResourceVariable([[1.0, 2.0]], dtype=dtype)
global_step = resource_variable_ops.ResourceVariable(
0, dtype=dtypes.int64)
x = constant_op.constant([[4.0], [5.0]], dtype=dtype)
pred = math_ops.matmul(embedding_ops.embedding_lookup([var0], [0]), x)
loss = pred * pred
sgd_op = adagrad_da.AdagradDAOptimizer(
1.0, global_step).minimize(loss)
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllCloseAccordingToType([[1.0, 2.0]], var0.eval())
# Run 1 step of sgd
sgd_op.run()
# Validate updated params
self.assertAllCloseAccordingToType(
[[-1, -1]], var0.eval(), rtol=0.01)
def testAdagradDAwithoutRegularizationBasic2(self):
for dtype in [dtypes.float64, dtypes.float32]:
with self.cached_session() as sess:
global_step = variables.Variable(0, dtype=dtypes.int64)
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([4.0, 3.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.2], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.02], dtype=dtype)
opt = adagrad_da.AdagradDAOptimizer(
3.0,
global_step,
initial_gradient_squared_accumulator_value=0.1,
l1_regularization_strength=0.0,
l2_regularization_strength=0.0)
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]), global_step=global_step)
variables.global_variables_initializer().run()
v0_val, v1_val = sess.run([var0, var1])
self.assertAllCloseAccordingToType([1.0, 2.0], v0_val)
self.assertAllCloseAccordingToType([4.0, 3.0], v1_val)
# Run a step of AdagradDA
update.run()
v0_val, v1_val = sess.run([var0, var1])
self.assertAllCloseAccordingToType(
np.array([-0.904534, -1.603567]), v0_val)
self.assertAllCloseAccordingToType(
np.array([-0.094821, -0.189358]), v1_val)
def testAdagradDAWithL1(self):
for dtype in [dtypes.float64, dtypes.float32]:
with self.cached_session() as sess:
global_step = variables.Variable(0, dtype=dtypes.int64)
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([4.0, 3.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.2], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.02], dtype=dtype)
opt = adagrad_da.AdagradDAOptimizer(
3.0,
global_step,
initial_gradient_squared_accumulator_value=0.1,
l1_regularization_strength=0.001,
l2_regularization_strength=0.0)
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]), global_step=global_step)
variables.global_variables_initializer().run()
v0_val, v1_val = sess.run([var0, var1])
self.assertAllCloseAccordingToType([1.0, 2.0], v0_val)
self.assertAllCloseAccordingToType([4.0, 3.0], v1_val)
# Run a step of AdagradDA
update.run()
v0_val, v1_val = sess.run([var0, var1])
self.assertAllCloseAccordingToType(
np.array([-0.895489, -1.59555]), v0_val)
self.assertAllCloseAccordingToType(
np.array([-0.085339, -0.17989]), v1_val)
def testAdagradDAWithL1_L2(self):
for dtype in [dtypes.float64, dtypes.float32]:
with self.cached_session() as sess:
global_step = variables.Variable(0, dtype=dtypes.int64)
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([4.0, 3.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.2], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.02], dtype=dtype)
opt = adagrad_da.AdagradDAOptimizer(
3.0,
global_step,
initial_gradient_squared_accumulator_value=0.1,
l1_regularization_strength=0.001,
l2_regularization_strength=2.0)
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]), global_step=global_step)
variables.global_variables_initializer().run()
v0_val, v1_val = sess.run([var0, var1])
self.assertAllCloseAccordingToType([1.0, 2.0], v0_val)
self.assertAllCloseAccordingToType([4.0, 3.0], v1_val)
# Run a step of AdagradDA
update.run()
v0_val, v1_val = sess.run([var0, var1])
self.assertAllCloseAccordingToType(
np.array([-0.046907, -0.093659]), v0_val)
self.assertAllCloseAccordingToType(
np.array([-0.004275, -0.009023]), v1_val)
if __name__ == "__main__":
test.main()
|
