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
|
# Copyright 2015 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 test for GradientDescent."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from tensorflow.python.eager import backprop
from tensorflow.python.eager import context
from tensorflow.python.eager import function
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
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 resources
from tensorflow.python.ops import variables
from tensorflow.python.platform import test
from tensorflow.python.training import gradient_descent
class GradientDescentOptimizerTest(test.TestCase):
def testBasic(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.test_session():
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([3.0, 4.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.1], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.01], dtype=dtype)
optimizer = gradient_descent.GradientDescentOptimizer(3.0)
sgd_op = optimizer.apply_gradients(
zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllCloseAccordingToType([1.0, 2.0], var0.eval())
self.assertAllCloseAccordingToType([3.0, 4.0], var1.eval())
# Run 1 step of sgd
sgd_op.run()
# Validate updated params
self.assertAllCloseAccordingToType([1.0 - 3.0 * 0.1, 2.0 - 3.0 * 0.1],
var0.eval())
self.assertAllCloseAccordingToType([3.0 - 3.0 * 0.01, 4.0 - 3.0 * 0.01],
var1.eval())
self.assertEqual(0, len(optimizer.variables()))
def testBasicResourceVariable(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.test_session():
var0 = resource_variable_ops.ResourceVariable([1.0, 2.0], dtype=dtype)
var1 = resource_variable_ops.ResourceVariable([3.0, 4.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.1], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.01], dtype=dtype)
sgd_op = gradient_descent.GradientDescentOptimizer(3.0).apply_gradients(
zip([grads0, grads1], [var0, var1]))
# TODO(apassos) calling initialize_resources on all resources here
# doesn't work because the sessions and graph are reused across unit
# tests and this would mean trying to reinitialize variables. Figure out
# a long-term solution for this.
resources.initialize_resources([var0, var1]).run()
# Fetch params to validate initial values
self.assertAllCloseAccordingToType([1.0, 2.0], var0.eval())
self.assertAllCloseAccordingToType([3.0, 4.0], var1.eval())
# Run 1 step of sgd
sgd_op.run()
# Validate updated params
self.assertAllCloseAccordingToType([1.0 - 3.0 * 0.1, 2.0 - 3.0 * 0.1],
var0.eval())
self.assertAllCloseAccordingToType([3.0 - 3.0 * 0.01, 4.0 - 3.0 * 0.01],
var1.eval())
def testBasicCallableParams(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.test_session():
var0 = resource_variable_ops.ResourceVariable([1.0, 2.0], dtype=dtype)
var1 = resource_variable_ops.ResourceVariable([3.0, 4.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.1], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.01], dtype=dtype)
lr = lambda: 3.0
sgd_op = gradient_descent.GradientDescentOptimizer(lr).apply_gradients(
zip([grads0, grads1], [var0, var1]))
# TODO(apassos) calling initialize_resources on all resources here
# doesn't work because the sessions and graph are reused across unit
# tests and this would mean trying to reinitialize variables. Figure out
# a long-term solution for this.
resources.initialize_resources([var0, var1]).run()
# Fetch params to validate initial values
self.assertAllCloseAccordingToType([1.0, 2.0], var0.eval())
self.assertAllCloseAccordingToType([3.0, 4.0], var1.eval())
# Run 1 step of sgd
sgd_op.run()
# Validate updated params
self.assertAllCloseAccordingToType([1.0 - 3.0 * 0.1, 2.0 - 3.0 * 0.1],
var0.eval())
self.assertAllCloseAccordingToType([3.0 - 3.0 * 0.01, 4.0 - 3.0 * 0.01],
var1.eval())
def testMinimizeResourceVariable(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.test_session():
var0 = resource_variable_ops.ResourceVariable([[1.0, 2.0]], dtype=dtype)
var1 = resource_variable_ops.ResourceVariable([3.0], dtype=dtype)
x = constant_op.constant([[4.0], [5.0]], dtype=dtype)
pred = math_ops.matmul(var0, x) + var1
loss = pred * pred
sgd_op = gradient_descent.GradientDescentOptimizer(1.0).minimize(loss)
# TODO(apassos) calling initialize_resources on all resources here
# doesn't work because the sessions and graph are reused across unit
# tests and this would mean trying to reinitialize variables. Figure out
# a long-term solution for this.
resources.initialize_resources([var0, var1]).run()
# Fetch params to validate initial values
self.assertAllCloseAccordingToType([[1.0, 2.0]], var0.eval())
self.assertAllCloseAccordingToType([3.0], var1.eval())
# Run 1 step of sgd
sgd_op.run()
# Validate updated params
np_pred = 1.0 * 4.0 + 2.0 * 5.0 + 3.0
np_grad = 2 * np_pred
self.assertAllCloseAccordingToType(
[[1.0 - np_grad * 4.0, 2.0 - np_grad * 5.0]], var0.eval())
self.assertAllCloseAccordingToType([3.0 - np_grad], var1.eval())
def testMinimizeSparseResourceVariable(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.test_session():
var0 = resource_variable_ops.ResourceVariable([[1.0, 2.0]], dtype=dtype)
var1 = resource_variable_ops.ResourceVariable([3.0], dtype=dtype)
x = constant_op.constant([[4.0], [5.0]], dtype=dtype)
pred = math_ops.matmul(embedding_ops.embedding_lookup([var0], [0]), x)
pred += var1
loss = pred * pred
sgd_op = gradient_descent.GradientDescentOptimizer(1.0).minimize(loss)
# TODO(apassos) calling initialize_resources on all resources here
# doesn't work because the sessions and graph are reused across unit
# tests and this would mean trying to reinitialize variables. Figure out
# a long-term solution for this.
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllCloseAccordingToType([[1.0, 2.0]], var0.eval())
self.assertAllCloseAccordingToType([3.0], var1.eval())
# Run 1 step of sgd
sgd_op.run()
# Validate updated params
np_pred = 1.0 * 4.0 + 2.0 * 5.0 + 3.0
np_grad = 2 * np_pred
self.assertAllCloseAccordingToType(
[[1.0 - np_grad * 4.0, 2.0 - np_grad * 5.0]], var0.eval())
self.assertAllCloseAccordingToType([3.0 - np_grad], var1.eval())
def testTensorLearningRate(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.test_session():
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([3.0, 4.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.1], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.01], dtype=dtype)
lrate = constant_op.constant(3.0)
sgd_op = gradient_descent.GradientDescentOptimizer(
lrate).apply_gradients(zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllCloseAccordingToType([1.0, 2.0], var0.eval())
self.assertAllCloseAccordingToType([3.0, 4.0], var1.eval())
# Run 1 step of sgd
sgd_op.run()
# Validate updated params
self.assertAllCloseAccordingToType([1.0 - 3.0 * 0.1, 2.0 - 3.0 * 0.1],
var0.eval())
self.assertAllCloseAccordingToType([3.0 - 3.0 * 0.01, 4.0 - 3.0 * 0.01],
var1.eval())
def testGradWrtRef(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.test_session():
opt = gradient_descent.GradientDescentOptimizer(3.0)
values = [1.0, 3.0]
vars_ = [variables.Variable([v], dtype=dtype) for v in values]
grads_and_vars = opt.compute_gradients(vars_[0] + vars_[1], vars_)
variables.global_variables_initializer().run()
for grad, _ in grads_and_vars:
self.assertAllCloseAccordingToType([1.0], grad.eval())
def testWithGlobalStep(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.test_session():
global_step = variables.Variable(0, trainable=False)
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([3.0, 4.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.1], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.01], dtype=dtype)
sgd_op = gradient_descent.GradientDescentOptimizer(3.0).apply_gradients(
zip([grads0, grads1], [var0, var1]), global_step=global_step)
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllCloseAccordingToType([1.0, 2.0], var0.eval())
self.assertAllCloseAccordingToType([3.0, 4.0], var1.eval())
# Run 1 step of sgd
sgd_op.run()
# Validate updated params and global_step
self.assertAllCloseAccordingToType([1.0 - 3.0 * 0.1, 2.0 - 3.0 * 0.1],
var0.eval())
self.assertAllCloseAccordingToType([3.0 - 3.0 * 0.01, 4.0 - 3.0 * 0.01],
var1.eval())
self.assertAllCloseAccordingToType(1, global_step.eval())
def testSparseBasic(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.test_session():
var0 = variables.Variable([[1.0], [2.0]], dtype=dtype)
var1 = variables.Variable([[3.0], [4.0]], dtype=dtype)
grads0 = ops.IndexedSlices(
constant_op.constant(
[0.1], shape=[1, 1], dtype=dtype),
constant_op.constant([0]),
constant_op.constant([2, 1]))
grads1 = ops.IndexedSlices(
constant_op.constant(
[0.01], shape=[1, 1], dtype=dtype),
constant_op.constant([1]),
constant_op.constant([2, 1]))
sgd_op = gradient_descent.GradientDescentOptimizer(3.0).apply_gradients(
zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllCloseAccordingToType([[1.0], [2.0]], var0.eval())
self.assertAllCloseAccordingToType([[3.0], [4.0]], var1.eval())
# Run 1 step of sgd
sgd_op.run()
# Validate updated params
self.assertAllCloseAccordingToType([[1.0 - 3.0 * 0.1], [2.0]],
var0.eval())
self.assertAllCloseAccordingToType([[3.0], [4.0 - 3.0 * 0.01]],
var1.eval())
def testCapturingInDefunWhileExecutingEagerly(self):
with context.eager_mode():
optimizer = gradient_descent.GradientDescentOptimizer(1.0)
def step():
v = resource_variable_ops.ResourceVariable(1.0)
with backprop.GradientTape() as tape:
loss = v ** 2
grad = tape.gradient(loss, v)
optimizer.apply_gradients([(grad, v)])
return v.read_value()
compiled_step = function.defun(step)
self.assertEqual(float(step()), -1.0)
self.assertEqual(float(compiled_step()), -1.0)
# This shouldn't fail; in particular, the learning rate tensor should
# be an EagerTensor once again, not a graph Tensor.
self.assertEqual(float(step()), -1.0)
if __name__ == "__main__":
test.main()
|
