def _create_arm_policy_and_observations(
self
) -> Tuple[greedy_multi_objective_policy.GreedyMultiObjectiveNeuralPolicy,
Dict[Text, tf.Tensor]]:
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(2, 3, 4)
time_step_spec = ts.time_step_spec(obs_spec)
action_spec = tensor_spec.BoundedTensorSpec((), tf.int32, 0, 3)
objective_networks = [
global_and_arm_feature_network.
create_feed_forward_common_tower_network(obs_spec, (4, 3), (3, 4),
(4, 2)) for _ in range(3)
]
policy = greedy_multi_objective_policy.GreedyMultiObjectiveNeuralPolicy(
time_step_spec,
action_spec,
self._scalarizer,
objective_networks,
accepts_per_arm_features=True,
emit_policy_info=('predicted_rewards_mean', ))
action_feature = tf.cast(tf.reshape(tf.random.shuffle(tf.range(24)),
shape=[2, 4, 3]),
dtype=tf.float32)
observations = {
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [2, 1]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY:
action_feature
}
return policy, observations
def testComputeLossWithArmFeatures(self):
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(
global_dim=2, per_arm_dim=3, num_actions=3)
time_step_spec = ts.time_step_spec(obs_spec)
constraint_net = (global_and_arm_feature_network.
create_feed_forward_common_tower_network(
obs_spec,
global_layers=(4, ),
arm_layers=(4, ),
common_layers=(4, )))
neural_constraint = constraints.NeuralConstraint(
time_step_spec,
self._action_spec,
constraint_network=constraint_net)
observations = {
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY:
tf.cast(tf.reshape(tf.range(18), shape=[2, 3, 3]),
dtype=tf.float32)
}
actions = tf.constant([0, 1], dtype=tf.int32)
rewards = tf.constant([0.5, 3.0], dtype=tf.float32)
init_op = neural_constraint.initialize()
if not tf.executing_eagerly():
with self.cached_session() as sess:
common.initialize_uninitialized_variables(sess)
self.assertIsNone(sess.run(init_op))
loss = neural_constraint.compute_loss(observations, actions, rewards)
self.assertGreater(self.evaluate(loss), 0.0)
def testTrainPerArmAgent(self):
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(2, 3, 3)
time_step_spec = ts.time_step_spec(obs_spec)
reward_net = (global_and_arm_feature_network.
create_feed_forward_common_tower_network(
obs_spec, (4, 3), (3, 4), (4, 2)))
optimizer = tf.compat.v1.train.GradientDescentOptimizer(
learning_rate=0.1)
agent = neural_epsilon_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec,
self._action_spec,
reward_network=reward_net,
optimizer=optimizer,
epsilon=0.1,
accepts_per_arm_features=True)
observations = {
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY:
tf.cast(tf.reshape(tf.range(18), shape=[2, 3, 3]),
dtype=tf.float32)
}
time_steps = ts.restart(observations, batch_size=2)
policy = agent.policy
action_step = policy.action(time_steps)
self.evaluate(tf.compat.v1.initialize_all_variables())
actions = self.evaluate(action_step.action)
self.assertAllEqual(actions.shape, (2, ))
def testTrainPerArmAgent(self):
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(2, 3, 4)
time_step_spec = ts.time_step_spec(obs_spec)
reward_net = (
global_and_arm_feature_network.create_feed_forward_per_arm_network(
obs_spec, (4, 3), (3, 4), (4, 2)))
optimizer = tf.compat.v1.train.GradientDescentOptimizer(
learning_rate=0.1)
agent = greedy_agent.GreedyRewardPredictionAgent(
time_step_spec,
self._action_spec,
reward_network=reward_net,
accepts_per_arm_features=True,
optimizer=optimizer)
observations = {
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY:
tf.cast(tf.reshape(tf.range(24), shape=[2, 4, 3]),
dtype=tf.float32)
}
actions = np.array([0, 3], 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 = policy_step.PolicyStep(
action=tf.convert_to_tensor(actions),
info=policy_utilities.PerArmPolicyInfo(
chosen_arm_features=np.array([[1, 2, 3], [3, 2, 1]],
dtype=np.float32)))
experience = _get_experience(initial_step, action_step, final_step)
agent.train(experience, None)
self.evaluate(tf.compat.v1.initialize_all_variables())
def testPerArmObservation(self, batch_size, actions_from_reward_layer):
global_obs_dim = 7
arm_obs_dim = 3
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(
global_obs_dim, arm_obs_dim, self._num_actions)
time_step_spec = ts.time_step_spec(obs_spec)
dummy_net = arm_network.create_feed_forward_common_tower_network(
obs_spec,
global_layers=(3, 4, 5),
arm_layers=(3, 2),
common_layers=(4, 3),
output_dim=self._encoding_dim)
reward_layer = get_per_arm_reward_layer(
encoding_dim=self._encoding_dim)
policy = neural_linucb_policy.NeuralLinUCBPolicy(
dummy_net,
self._encoding_dim,
reward_layer,
actions_from_reward_layer=tf.constant(actions_from_reward_layer,
dtype=tf.bool),
cov_matrix=self._a[0:1],
data_vector=self._b[0:1],
num_samples=self._num_samples_per_arm[0:1],
epsilon_greedy=0.0,
time_step_spec=time_step_spec,
accepts_per_arm_features=True,
emit_policy_info=('predicted_rewards_mean', ))
current_time_step = self._per_arm_time_step_batch(
batch_size=batch_size,
global_obs_dim=global_obs_dim,
arm_obs_dim=arm_obs_dim)
action_step = policy.action(current_time_step)
self.assertEqual(action_step.action.dtype, tf.int32)
self.evaluate(tf.compat.v1.global_variables_initializer())
action_fn = common.function_in_tf1()(policy.action)
action_step = action_fn(current_time_step)
input_observation = current_time_step.observation
encoded_observation, _ = dummy_net(input_observation)
if actions_from_reward_layer:
predicted_rewards_from_reward_layer = reward_layer(
encoded_observation)
predicted_rewards_expected = self.evaluate(
predicted_rewards_from_reward_layer).reshape(
(-1, self._num_actions))
else:
observation_numpy = self.evaluate(encoded_observation)
predicted_rewards_expected = (
self._get_predicted_rewards_from_per_arm_linucb(
observation_numpy, batch_size))
p_info = self.evaluate(action_step.info)
self.assertEqual(p_info.predicted_rewards_mean.dtype, np.float32)
self.assertAllClose(p_info.predicted_rewards_mean,
predicted_rewards_expected)
def testTrainPerArmAgentVariableActions(self):
num_actions = 5
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(
2, 3, num_actions, add_num_actions_feature=True)
time_step_spec = time_step.time_step_spec(obs_spec)
action_spec = tensor_spec.BoundedTensorSpec(
dtype=tf.int32, shape=(), minimum=0, maximum=num_actions - 1)
encoding_dim = 10
encoder = (
global_and_arm_feature_network.create_feed_forward_common_tower_network(
obs_spec, (4, 3), (3, 4), (4, 2), encoding_dim))
agent = neural_linucb_agent.NeuralLinUCBAgent(
time_step_spec=time_step_spec,
action_spec=action_spec,
encoding_network=encoder,
encoding_network_num_train_steps=10,
encoding_dim=encoding_dim,
accepts_per_arm_features=True,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=0.001))
observations = {
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY:
tf.cast(
tf.reshape(tf.range(30), shape=[2, 5, 3]), dtype=tf.float32),
bandit_spec_utils.NUM_ACTIONS_FEATURE_KEY:
tf.constant([3, 4], dtype=tf.int32)
}
actions = np.array([0, 3], dtype=np.int32)
rewards = np.array([0.5, 3.0], dtype=np.float32)
initial_step = time_step.TimeStep(
tf.constant(
time_step.StepType.FIRST,
dtype=tf.int32,
shape=[2],
name='step_type'),
tf.constant(0.0, dtype=tf.float32, shape=[2], name='reward'),
tf.constant(1.0, dtype=tf.float32, shape=[2], name='discount'),
observations)
final_step = time_step.TimeStep(
tf.constant(
time_step.StepType.LAST,
dtype=tf.int32,
shape=[2],
name='step_type'),
tf.constant(rewards, dtype=tf.float32, name='reward'),
tf.constant(1.0, dtype=tf.float32, shape=[2], name='discount'),
observations)
action_step = policy_step.PolicyStep(
action=tf.convert_to_tensor(actions),
info=policy_utilities.PerArmPolicyInfo(
chosen_arm_features=np.array([[1, 2, 3], [3, 2, 1]],
dtype=np.float32)))
experience = _get_experience(initial_step, action_step, final_step)
loss_info, _ = agent.train(experience, None)
self.evaluate(tf.compat.v1.initialize_all_variables())
loss_value = self.evaluate(loss_info)
self.assertGreater(loss_value, 0.0)
def testLinearAgentUpdatePerArmFeatures(self,
batch_size,
context_dim,
exploration_policy,
dtype,
use_eigendecomp=False):
"""Check that the agent updates for specified actions and rewards."""
# Construct a `Trajectory` for the given action, observation, reward.
num_actions = 5
global_context_dim = context_dim
arm_context_dim = 3
initial_step, final_step = (
_get_initial_and_final_steps_with_per_arm_features(
batch_size, global_context_dim, num_actions, arm_context_dim))
action = np.random.randint(num_actions, size=batch_size, dtype=np.int32)
action_step = policy_step.PolicyStep(
action=tf.convert_to_tensor(action),
info=policy_utilities.PerArmPolicyInfo(
chosen_arm_features=np.arange(
batch_size * arm_context_dim, dtype=np.float32).reshape(
[batch_size, arm_context_dim])))
experience = _get_experience(initial_step, action_step, final_step)
# Construct an agent and perform the update.
observation_spec = bandit_spec_utils.create_per_arm_observation_spec(
context_dim, arm_context_dim, num_actions)
time_step_spec = time_step.time_step_spec(observation_spec)
action_spec = tensor_spec.BoundedTensorSpec(
dtype=tf.int32, shape=(), minimum=0, maximum=num_actions - 1)
agent = linear_agent.LinearBanditAgent(
exploration_policy=exploration_policy,
time_step_spec=time_step_spec,
action_spec=action_spec,
use_eigendecomp=use_eigendecomp,
accepts_per_arm_features=True,
dtype=dtype)
self.evaluate(agent.initialize())
loss_info = agent.train(experience)
self.evaluate(loss_info)
final_a = self.evaluate(agent.cov_matrix)
final_b = self.evaluate(agent.data_vector)
# Compute the expected updated estimates.
global_observation = experience.observation[
bandit_spec_utils.GLOBAL_FEATURE_KEY]
arm_observation = experience.policy_info.chosen_arm_features
overall_observation = tf.squeeze(
tf.concat([global_observation, arm_observation], axis=-1), axis=1)
rewards = tf.squeeze(experience.reward, axis=1)
expected_a_new = tf.matmul(
overall_observation, overall_observation, transpose_a=True)
expected_b_new = bandit_utils.sum_reward_weighted_observations(
rewards, overall_observation)
self.assertAllClose(expected_a_new, final_a[0])
self.assertAllClose(expected_b_new, final_b[0])
def testCreateFeedForwardDotProductNetwork(self, batch_size, feature_dim,
num_actions):
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(
7, feature_dim, num_actions)
net = gafn.create_feed_forward_dot_product_network(
obs_spec, (4, 3, 4), (6, 5, 4))
input_nest = tensor_spec.sample_spec_nest(obs_spec,
outer_dims=(batch_size, ))
output, _ = self.evaluate(net(input_nest))
self.assertAllEqual(output.shape, (batch_size, num_actions))
def testCreateFeedForwardCommonTowerNetwork(self, batch_size, feature_dim,
num_actions):
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(
7, feature_dim, num_actions)
net = gafn.create_feed_forward_common_tower_network(
obs_spec, (4, 3, 2), (6, 5, 4), (7, 6, 5))
input_nest = tensor_spec.sample_spec_nest(obs_spec,
outer_dims=(batch_size, ))
output, _ = net(input_nest)
self.evaluate(tf.compat.v1.global_variables_initializer())
output = self.evaluate(output)
self.assertAllEqual(output.shape, (batch_size, num_actions))
def testNoneTimeStepSpecForPerArmFeaturesRaisesError(self):
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(2, 3, 4)
action_spec = tensor_spec.BoundedTensorSpec((), tf.int32, 0, 3)
objective_networks = [
global_and_arm_feature_network.
create_feed_forward_common_tower_network(obs_spec, (4, 3), (3, 4),
(4, 2)) for _ in range(3)
]
with self.assertRaisesRegexp(
ValueError,
'time_step_spec should not be None for per-arm-features policies'
):
greedy_multi_objective_policy.GreedyMultiObjectiveNeuralPolicy(
None,
action_spec,
self._scalarizer,
objective_networks,
accepts_per_arm_features=True,
emit_policy_info=('predicted_rewards_mean', ))
def testPerArmRewardsVariableNumActions(self):
tf.compat.v1.set_random_seed(3000)
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(
2, 3, 4, add_num_actions_feature=True)
time_step_spec = ts.time_step_spec(obs_spec)
action_spec = tensor_spec.BoundedTensorSpec((), tf.int32, 0, 3)
reward_network = (
global_and_arm_feature_network.create_feed_forward_common_tower_network(
obs_spec, (4, 3), (3, 4), (4, 2)))
policy = greedy_reward_policy.GreedyRewardPredictionPolicy(
time_step_spec,
action_spec,
reward_network=reward_network,
accepts_per_arm_features=True,
emit_policy_info=('predicted_rewards_mean',))
action_feature = tf.cast(
tf.reshape(tf.random.shuffle(tf.range(24)), shape=[2, 4, 3]),
dtype=tf.float32)
observations = {
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY:
action_feature,
bandit_spec_utils.NUM_ACTIONS_FEATURE_KEY:
tf.constant([2, 3], dtype=tf.int32)
}
time_step = ts.restart(observations, batch_size=2)
action_step = policy.action(time_step, seed=1)
self.assertEqual(action_step.action.shape.as_list(), [2])
self.assertEqual(action_step.action.dtype, tf.int32)
# Initialize all variables
self.evaluate(tf.compat.v1.global_variables_initializer())
action, p_info, first_arm_features = self.evaluate([
action_step.action, action_step.info,
observations[bandit_spec_utils.PER_ARM_FEATURE_KEY][0]
])
self.assertAllEqual(action.shape, [2])
self.assertAllEqual(p_info.predicted_rewards_mean.shape, [2, 4])
self.assertAllEqual(p_info.chosen_arm_features.shape, [2, 3])
first_action = action[0]
self.assertAllEqual(p_info.chosen_arm_features[0],
first_arm_features[first_action])
def testTrainPerArmAgent(self):
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(
2, 3, 4, add_num_actions_feature=True)
time_step_spec = ts.time_step_spec(observation_spec=obs_spec,
reward_spec=tensor_spec.TensorSpec(
[3], tf.float32))
action_spec = tensor_spec.BoundedTensorSpec((), tf.int32, 0, 3)
networks_and_loss_fns = [
(global_and_arm_feature_network.
create_feed_forward_common_tower_network(obs_spec, (4, 3), (3, 4),
(4, 2)),
tf.compat.v1.losses.mean_squared_error) for _ in range(3)
]
optimizer = tf.compat.v1.train.GradientDescentOptimizer(
learning_rate=0.01)
agent = greedy_multi_objective_agent.GreedyMultiObjectiveNeuralAgent(
time_step_spec,
action_spec,
self._scalarizer,
objective_network_and_loss_fn_sequence=networks_and_loss_fns,
accepts_per_arm_features=True,
optimizer=optimizer)
observations = {
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY:
tf.cast(tf.reshape(tf.range(24), shape=[2, 4, 3]),
dtype=tf.float32),
bandit_spec_utils.NUM_ACTIONS_FEATURE_KEY:
tf.ones([2], dtype=tf.int32)
}
actions = np.array([0, 3], dtype=np.int32)
objectives = np.array([[1, 2, 3], [6, 5, 4]], dtype=np.float32)
initial_step, final_step = _get_initial_and_final_steps(
observations, objectives)
action_step = policy_step.PolicyStep(
action=tf.convert_to_tensor(actions),
info=policy_utilities.PerArmPolicyInfo(
chosen_arm_features=np.array([[1, 2, 3], [3, 2, 1]],
dtype=np.float32)))
experience = _get_experience(initial_step, action_step, final_step)
agent.train(experience, None)
self.evaluate(tf.compat.v1.initialize_all_variables())
def testComputeMaskFromMultipleSourcesMask(self):
observation_spec = bandit_spec_utils.create_per_arm_observation_spec(
4, 5, 6)
time_step_spec = ts.time_step_spec(observation_spec)
action_spec = tensor_spec.BoundedTensorSpec((), tf.int32, 0, 5)
constraint_net = (
global_and_arm_feature_network.create_feed_forward_common_tower_network(
observation_spec, (3, 4), (4, 3), (2, 3)))
neural_constraint = constraints.NeuralConstraint(
time_step_spec,
action_spec,
constraint_network=constraint_net)
original_mask = [[1, 1, 1, 1, 0, 0], [1, 1, 1, 0, 0, 0]]
observations = ({
'global': tf.constant([[1, 2, 3, 4], [5, 6, 7, 8]], dtype=tf.float32),
'per_arm': tf.reshape(tf.range(60, dtype=tf.float32), shape=[2, 6, 5]),
}, original_mask)
mask = constraints.construct_mask_from_multiple_sources(
observations, lambda x: (x[0], x[1]), [neural_constraint], 6)
self.assertAllGreaterEqual(original_mask - mask, 0)
def testTrainPerArmAgentWithMask(self):
num_actions = 4
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(
2, 3, num_actions)
mask_obs_spec = (obs_spec,
tensor_spec.BoundedTensorSpec(shape=[num_actions],
minimum=0,
maximum=1,
dtype=tf.float32))
time_step_spec = ts.time_step_spec(mask_obs_spec)
reward_net = (global_and_arm_feature_network.
create_feed_forward_common_tower_network(
obs_spec, (4, 3), (3, 4), (4, 2)))
optimizer = tf.compat.v1.train.GradientDescentOptimizer(
learning_rate=0.1)
agent = greedy_agent.GreedyRewardPredictionAgent(
time_step_spec,
self._action_spec,
reward_network=reward_net,
observation_and_action_constraint_splitter=lambda x: [x[0], x[1]],
accepts_per_arm_features=True,
optimizer=optimizer)
observations = ({
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY:
tf.cast(tf.reshape(tf.range(24), shape=[2, 4, 3]),
dtype=tf.float32)
}, tf.ones([2, num_actions]))
actions = np.array([0, 3], dtype=np.int32)
rewards = np.array([0.5, 3.0], dtype=np.float32)
initial_step, final_step = _get_initial_and_final_steps_with_action_mask(
observations, rewards)
action_step = policy_step.PolicyStep(
action=tf.convert_to_tensor(actions),
info=policy_utilities.PerArmPolicyInfo(
chosen_arm_features=np.array([[1, 2, 3], [3, 2, 1]],
dtype=np.float32)))
experience = _get_experience(initial_step, action_step, final_step)
agent.train(experience, None)
self.evaluate(tf.compat.v1.initialize_all_variables())
def setUp(self):
super(LinearBanditPolicyTest, self).setUp()
self._obs_dim = 2
self._num_actions = 5
self._obs_spec = tensor_spec.TensorSpec([self._obs_dim], tf.float32)
self._obs_spec_with_mask = (tensor_spec.TensorSpec([self._obs_dim],
tf.float32),
tensor_spec.TensorSpec([self._num_actions],
tf.int32))
self._per_arm_obs_spec = bandit_spec_utils.create_per_arm_observation_spec(
self._obs_dim, 4, self._num_actions, add_num_actions_feature=True)
self._time_step_spec = ts.time_step_spec(self._obs_spec)
self._time_step_spec_with_mask = ts.time_step_spec(self._obs_spec_with_mask)
self._per_arm_time_step_spec = ts.time_step_spec(self._per_arm_obs_spec)
self._alpha = 1.0
self._action_spec = tensor_spec.BoundedTensorSpec(
shape=(),
dtype=tf.int32,
minimum=0,
maximum=self._num_actions - 1,
name='action')
def testComputeMaskFromMultipleSourcesNumActionsFeature(self):
observation_spec = bandit_spec_utils.create_per_arm_observation_spec(
4, 5, 6, add_num_actions_feature=True)
time_step_spec = ts.time_step_spec(observation_spec)
action_spec = tensor_spec.BoundedTensorSpec((), tf.int32, 0, 5)
constraint_net = (
global_and_arm_feature_network.create_feed_forward_common_tower_network(
observation_spec, (3, 4), (4, 3), (2, 3)))
neural_constraint = constraints.NeuralConstraint(
time_step_spec,
action_spec,
constraint_network=constraint_net)
observations = {
'global': tf.constant([[1, 2, 3, 4], [5, 6, 7, 8]], dtype=tf.float32),
'per_arm': tf.reshape(tf.range(60, dtype=tf.float32), shape=[2, 6, 5]),
'num_actions': tf.constant([4, 3], dtype=tf.int32)
}
mask = constraints.construct_mask_from_multiple_sources(
observations, None, [neural_constraint], 6)
implied_mask = [[1, 1, 1, 1, 0, 0], [1, 1, 1, 0, 0, 0]]
self.assertAllGreaterEqual(implied_mask - mask, 0)
def testPerArmRewards(self):
tf.compat.v1.set_random_seed(3000)
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(2, 3, 4)
time_step_spec = ts.time_step_spec(obs_spec)
action_spec = tensor_spec.BoundedTensorSpec((), tf.int32, 0, 3)
reward_network = (
global_and_arm_feature_network.create_feed_forward_common_tower_network(
obs_spec, (4, 3), (3, 4), (4, 2)))
policy = greedy_reward_policy.GreedyRewardPredictionPolicy(
time_step_spec,
action_spec,
reward_network=reward_network,
accepts_per_arm_features=True,
emit_policy_info=('predicted_rewards_mean',))
action_feature = tf.cast(
tf.reshape(tf.random.shuffle(tf.range(24)), shape=[2, 4, 3]),
dtype=tf.float32)
observations = {
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY: action_feature
}
time_step = ts.restart(observations, batch_size=2)
action_step = policy.action(time_step, seed=1)
self.assertEqual(action_step.action.shape.as_list(), [2])
self.assertEqual(action_step.action.dtype, tf.int32)
# Initialize all variables
self.evaluate(tf.compat.v1.global_variables_initializer())
action, p_info, first_arm_features = self.evaluate([
action_step.action, action_step.info,
observations[bandit_spec_utils.PER_ARM_FEATURE_KEY][0]
])
self.assertAllEqual(action.shape, [2])
self.assertAllEqual(p_info.predicted_rewards_mean.shape, [2, 4])
self.assertAllEqual(p_info.chosen_arm_features.shape, [2, 3])
first_action = action[0]
self.assertAllEqual(p_info.chosen_arm_features[0],
first_arm_features[first_action])
# Check that zeroing out some of the actions does not affect the predicted
# rewards for unchanged actions. This is to make sure that action feature
# padding does not influence the behavior.
if not tf.executing_eagerly():
# The below comparison will only work in tf2 because of the random per-arm
# observations get re-drawn in tf1.
return
padded_action_feature = tf.concat(
[action_feature[:, 0:1, :],
tf.zeros(shape=[2, 3, 3], dtype=tf.float32)],
axis=1)
observations = {
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY: padded_action_feature
}
time_step = ts.restart(observations, batch_size=2)
padded_action_step = policy.action(time_step, seed=1)
padded_p_info = self.evaluate(padded_action_step.info)
self.assertAllEqual(p_info.predicted_rewards_mean[:, 0],
padded_p_info.predicted_rewards_mean[:, 0])
