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
|
# 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.
# ==============================================================================
"""Tests for Keras optimizers."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import gc
import weakref
import numpy as np
from tensorflow.python import keras
from tensorflow.python.eager import context
from tensorflow.python.framework import ops
from tensorflow.python.framework import test_util
from tensorflow.python.keras import testing_utils
from tensorflow.python.platform import test
from tensorflow.python.training.adam import AdamOptimizer
def _get_model(input_dim, num_hidden, output_dim):
model = keras.models.Sequential()
model.add(keras.layers.Dense(num_hidden,
activation='relu',
input_shape=(input_dim,)))
model.add(keras.layers.Dense(output_dim, activation='softmax'))
return model
def _test_optimizer(optimizer, target=0.75):
np.random.seed(1337)
(x_train, y_train), _ = testing_utils.get_test_data(train_samples=1000,
test_samples=200,
input_shape=(10,),
num_classes=2)
y_train = keras.utils.to_categorical(y_train)
model = _get_model(x_train.shape[1], 20, y_train.shape[1])
model.compile(loss='categorical_crossentropy',
optimizer=optimizer,
metrics=['accuracy'])
np.testing.assert_equal(keras.backend.get_value(model.optimizer.iterations),
0)
history = model.fit(x_train, y_train, epochs=2, batch_size=16, verbose=0)
np.testing.assert_equal(keras.backend.get_value(model.optimizer.iterations),
126) # 63 steps per epoch
assert history.history['acc'][-1] >= target
config = keras.optimizers.serialize(optimizer)
optim = keras.optimizers.deserialize(config)
new_config = keras.optimizers.serialize(optim)
new_config['class_name'] = new_config['class_name'].lower()
assert config == new_config
# Test constraints.
model = keras.models.Sequential()
dense = keras.layers.Dense(10,
input_shape=(x_train.shape[1],),
kernel_constraint=lambda x: 0. * x + 1.,
bias_constraint=lambda x: 0. * x + 2.,
activation='relu')
model.add(dense)
model.add(keras.layers.Dense(y_train.shape[1], activation='softmax'))
model.compile(loss='categorical_crossentropy',
optimizer=optimizer,
metrics=['accuracy'])
np.testing.assert_equal(keras.backend.get_value(model.optimizer.iterations),
126) # Using same optimizer from before
model.train_on_batch(x_train[:10], y_train[:10])
np.testing.assert_equal(keras.backend.get_value(model.optimizer.iterations),
127)
kernel, bias = dense.get_weights()
np.testing.assert_allclose(kernel, 1., atol=1e-3)
np.testing.assert_allclose(bias, 2., atol=1e-3)
class KerasOptimizersTest(test.TestCase):
def test_sgd(self):
with self.cached_session():
_test_optimizer(keras.optimizers.SGD(lr=0.01,
momentum=0.9,
nesterov=True))
def test_rmsprop(self):
with self.cached_session():
_test_optimizer(keras.optimizers.RMSprop())
_test_optimizer(keras.optimizers.RMSprop(decay=1e-3))
def test_adagrad(self):
with self.cached_session():
_test_optimizer(keras.optimizers.Adagrad())
_test_optimizer(keras.optimizers.Adagrad(decay=1e-3))
def test_adadelta(self):
with self.cached_session():
_test_optimizer(keras.optimizers.Adadelta(), target=0.6)
# Accuracy seems dependent on the initialization. Even adding tf.Print
# nodes in the graph seemed to affect the initialization seed, and hence
# the accuracy.
_test_optimizer(keras.optimizers.Adadelta(decay=1e-3), target=0.4)
def test_adam(self):
with self.cached_session():
_test_optimizer(keras.optimizers.Adam())
_test_optimizer(keras.optimizers.Adam(decay=1e-3))
_test_optimizer(keras.optimizers.Adam(amsgrad=True))
def test_adamax(self):
with self.cached_session():
_test_optimizer(keras.optimizers.Adamax())
_test_optimizer(keras.optimizers.Adamax(decay=1e-3))
def test_nadam(self):
with self.cached_session():
_test_optimizer(keras.optimizers.Nadam())
def test_clipnorm(self):
with self.cached_session():
_test_optimizer(keras.optimizers.SGD(lr=0.01,
momentum=0.9,
clipnorm=0.5))
def test_clipvalue(self):
with self.cached_session():
_test_optimizer(keras.optimizers.SGD(lr=0.01,
momentum=0.9,
clipvalue=0.5))
def test_tfoptimizer(self):
optimizer = keras.optimizers.TFOptimizer(AdamOptimizer(0.01))
model = keras.models.Sequential()
model.add(keras.layers.Dense(
2, input_shape=(3,), kernel_constraint=keras.constraints.MaxNorm(1)))
# This is possible
model.compile(loss='mean_squared_error', optimizer=optimizer)
keras.backend.track_tf_optimizer(optimizer)
model.fit(np.random.random((5, 3)),
np.random.random((5, 2)),
epochs=1,
batch_size=5,
verbose=0)
# not supported
with self.assertRaises(NotImplementedError):
_ = optimizer.weights
with self.assertRaises(NotImplementedError):
optimizer.get_config()
with self.assertRaises(NotImplementedError):
optimizer.from_config(None)
def test_optimizer_garbage_collection(self):
graph = ops.Graph()
with graph.as_default():
optimizer = keras.optimizers.TFOptimizer(AdamOptimizer(0.01))
keras.backend.track_tf_optimizer(optimizer)
optimizer_weak = weakref.ref(optimizer)
graph_weak = weakref.ref(graph)
del graph, optimizer
gc.collect()
# Check that the weak references are dead now.
self.assertIs(graph_weak(), None)
self.assertIs(optimizer_weak(), None)
@test_util.run_in_graph_and_eager_modes
def test_tfoptimizer_iterations(self):
with self.cached_session():
optimizer = keras.optimizers.TFOptimizer(AdamOptimizer(0.01))
model = keras.models.Sequential()
model.add(keras.layers.Dense(
2, input_shape=(3,), kernel_constraint=keras.constraints.MaxNorm(1)))
model.compile(loss='mean_squared_error', optimizer=optimizer)
keras.backend.track_tf_optimizer(optimizer)
self.assertEqual(keras.backend.get_value(model.optimizer.iterations), 0)
model.fit(np.random.random((55, 3)),
np.random.random((55, 2)),
epochs=1,
batch_size=5,
verbose=0)
self.assertEqual(keras.backend.get_value(model.optimizer.iterations), 11)
if not context.executing_eagerly():
# TODO(kathywu): investigate why training with an array input and
# setting the argument steps_per_epoch does not work in eager mode.
model.fit(np.random.random((20, 3)),
np.random.random((20, 2)),
steps_per_epoch=8,
verbose=0)
self.assertEqual(
keras.backend.get_value(model.optimizer.iterations), 19)
def test_negative_clipvalue_or_clipnorm(self):
with self.assertRaises(ValueError):
_ = keras.optimizers.SGD(lr=0.01, clipvalue=-0.5)
with self.assertRaises(ValueError):
_ = keras.optimizers.Adam(clipnorm=-2.0)
if __name__ == '__main__':
test.main()
|
