-
Notifications
You must be signed in to change notification settings - Fork 721
/
dropout_thompson_sampling_agent_test.py
230 lines (209 loc) · 8.52 KB
/
dropout_thompson_sampling_agent_test.py
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
# coding=utf-8
# Copyright 2020 The TF-Agents Authors.
#
# 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
#
# https://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 dropout_thompson_sampling_agent.py."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import tensorflow as tf # pylint: disable=g-explicit-tensorflow-version-import
from tf_agents.bandits.agents import dropout_thompson_sampling_agent
from tf_agents.bandits.drivers import driver_utils
from tf_agents.policies import utils as policy_utilities
from tf_agents.specs import tensor_spec
from tf_agents.trajectories import policy_step
from tf_agents.trajectories import time_step as ts
def _get_initial_and_final_steps(observations, rewards):
batch_size = observations.shape[0]
initial_step = ts.TimeStep(
tf.constant(
ts.StepType.FIRST,
dtype=tf.int32,
shape=[batch_size],
name='step_type',
),
tf.constant(0.0, dtype=tf.float32, shape=[batch_size], name='reward'),
tf.constant(1.0, dtype=tf.float32, shape=[batch_size], name='discount'),
tf.constant(observations, dtype=tf.float32, name='observation'),
)
final_step = ts.TimeStep(
tf.constant(
ts.StepType.LAST, dtype=tf.int32, shape=[batch_size], name='step_type'
),
tf.constant(rewards, dtype=tf.float32, name='reward'),
tf.constant(1.0, dtype=tf.float32, shape=[batch_size], name='discount'),
tf.constant(observations + 100.0, dtype=tf.float32, name='observation'),
)
return initial_step, final_step
def _get_initial_and_final_steps_with_action_mask(
batch_size, context_dim, num_actions
):
observation = np.array(range(batch_size * context_dim)).reshape(
[batch_size, context_dim]
)
observation = tf.constant(observation, dtype=tf.float32)
mask = 1 - tf.eye(batch_size, num_columns=num_actions, dtype=tf.int32)
reward = np.random.uniform(0.0, 1.0, [batch_size])
initial_step = ts.TimeStep(
tf.constant(
ts.StepType.FIRST,
dtype=tf.int32,
shape=[batch_size],
name='step_type',
),
tf.constant(0.0, dtype=tf.float32, shape=[batch_size], name='reward'),
tf.constant(1.0, dtype=tf.float32, shape=[batch_size], name='discount'),
(observation, mask),
)
final_step = ts.TimeStep(
tf.constant(
ts.StepType.LAST, dtype=tf.int32, shape=[batch_size], name='step_type'
),
tf.constant(reward, dtype=tf.float32, shape=[batch_size], name='reward'),
tf.constant(1.0, dtype=tf.float32, shape=[batch_size], name='discount'),
(observation + 100.0, mask),
)
return initial_step, final_step
def _get_action_step(action):
return policy_step.PolicyStep(
action=tf.convert_to_tensor(action), info=policy_utilities.PolicyInfo()
)
def _get_experience(initial_step, action_step, final_step):
single_experience = driver_utils.trajectory_for_bandit(
initial_step, action_step, final_step
)
# Adds a 'time' dimension.
return tf.nest.map_structure(
lambda x: tf.expand_dims(tf.convert_to_tensor(x), 1), single_experience
)
class AgentTest(tf.test.TestCase):
def setUp(self):
super(AgentTest, self).setUp()
tf.compat.v1.enable_resource_variables()
self._obs_spec = tensor_spec.TensorSpec([2], tf.float32)
self._time_step_spec = ts.time_step_spec(self._obs_spec)
self._action_spec = tensor_spec.BoundedTensorSpec(
dtype=tf.int32, shape=(), minimum=0, maximum=2
)
def testCreateAgent(self):
agent = dropout_thompson_sampling_agent.DropoutThompsonSamplingAgent(
self._time_step_spec,
self._action_spec,
optimizer=None,
dropout_rate=0.1,
network_layers=(20, 20, 20),
)
self.assertIsNotNone(agent.policy)
def testTrainAgent(self):
optimizer = tf.compat.v1.train.GradientDescentOptimizer(learning_rate=0.1)
agent = dropout_thompson_sampling_agent.DropoutThompsonSamplingAgent(
self._time_step_spec,
self._action_spec,
optimizer=optimizer,
dropout_rate=0.1,
network_layers=(20, 20, 20),
dropout_only_top_layer=False,
)
observations = np.array([[1, 2], [3, 4]], dtype=np.float32)
actions = np.array([0, 1], dtype=np.int32)
rewards = np.array([0.5, 3.0], dtype=np.float32)
initial_step, final_step = _get_initial_and_final_steps(
observations, rewards
)
action_step = _get_action_step(actions)
experience = _get_experience(initial_step, action_step, final_step)
loss_before, _ = agent.train(experience, None)
loss_after, _ = agent.train(experience, None)
self.evaluate(tf.compat.v1.global_variables_initializer())
self.assertAllGreater(self.evaluate(loss_before), 0)
self.assertAllGreater(self.evaluate(loss_after), 0)
def testAgentWithMask(self):
optimizer = tf.compat.v1.train.GradientDescentOptimizer(learning_rate=0.1)
obs_spec = (
tensor_spec.TensorSpec([2], tf.float32),
tensor_spec.TensorSpec([3], tf.int32),
)
agent = dropout_thompson_sampling_agent.DropoutThompsonSamplingAgent(
ts.time_step_spec(obs_spec),
self._action_spec,
optimizer=optimizer,
observation_and_action_constraint_splitter=lambda x: (x[0], x[1]),
dropout_rate=0.1,
network_layers=(20, 20, 20),
dropout_only_top_layer=False,
)
actions = np.array([0, 1], dtype=np.int32)
initial_step, final_step = _get_initial_and_final_steps_with_action_mask(
2, 2, 3
)
action_step = _get_action_step(actions)
experience = _get_experience(initial_step, action_step, final_step)
loss_before, _ = agent.train(experience, None)
loss_after, _ = agent.train(experience, None)
self.evaluate(tf.compat.v1.global_variables_initializer())
self.assertAllGreater(self.evaluate(loss_before), 0)
self.assertAllGreater(self.evaluate(loss_after), 0)
def testTrainAgentHeteroscedastic(self):
optimizer = tf.compat.v1.train.GradientDescentOptimizer(learning_rate=0.1)
agent = dropout_thompson_sampling_agent.DropoutThompsonSamplingAgent(
self._time_step_spec,
self._action_spec,
optimizer=optimizer,
dropout_rate=0.1,
network_layers=(20, 20, 20),
dropout_only_top_layer=False,
heteroscedastic=True,
)
observations = np.array([[1, 2], [3, 4]], dtype=np.float32)
actions = np.array([0, 1], dtype=np.int32)
rewards = np.array([0.5, 3.0], dtype=np.float32)
initial_step, final_step = _get_initial_and_final_steps(
observations, rewards
)
action_step = _get_action_step(actions)
experience = _get_experience(initial_step, action_step, final_step)
loss_before, _ = agent.train(experience, None)
loss_after, _ = agent.train(experience, None)
self.evaluate(tf.compat.v1.global_variables_initializer())
self.assertIsNotNone(self.evaluate(loss_before))
self.assertIsNotNone(self.evaluate(loss_after))
def testAgentWithMaskHeteroscedastic(self):
optimizer = tf.compat.v1.train.GradientDescentOptimizer(learning_rate=0.1)
obs_spec = (
tensor_spec.TensorSpec([2], tf.float32),
tensor_spec.TensorSpec([3], tf.int32),
)
agent = dropout_thompson_sampling_agent.DropoutThompsonSamplingAgent(
ts.time_step_spec(obs_spec),
self._action_spec,
optimizer=optimizer,
observation_and_action_constraint_splitter=lambda x: (x[0], x[1]),
dropout_rate=0.1,
network_layers=(20, 20, 20),
dropout_only_top_layer=False,
heteroscedastic=True,
)
actions = np.array([0, 1], dtype=np.int32)
initial_step, final_step = _get_initial_and_final_steps_with_action_mask(
2, 2, 3
)
action_step = _get_action_step(actions)
experience = _get_experience(initial_step, action_step, final_step)
loss_before, _ = agent.train(experience, None)
loss_after, _ = agent.train(experience, None)
self.evaluate(tf.compat.v1.global_variables_initializer())
self.assertIsNotNone(self.evaluate(loss_before))
self.assertIsNotNone(self.evaluate(loss_after))
if __name__ == '__main__':
tf.test.main()