def testConjugateGradientMultipleRHSPlaceholders(self, n, rhs):
# Test the case where a_mat and b_mat are placeholders and they have unknown
# dimension values.
if tf.executing_eagerly():
return
x_obs = tf.constant(np.random.rand(n, 2),
dtype=tf.float32,
shape=[n, 2])
a_mat = tf.eye(n) + tf.matmul(x_obs, tf.linalg.matrix_transpose(x_obs))
a_mat_ph = tf.compat.v1.placeholder(tf.float32, shape=(None, None))
a_mat_value = self.evaluate(a_mat)
x_exact = tf.constant(np.random.rand(n, rhs),
dtype=tf.float32,
shape=[n, rhs])
b_mat = tf.matmul(a_mat, x_exact)
b_mat_ph = tf.compat.v1.placeholder(tf.float32, shape=(None, None))
b_mat_value = self.evaluate(b_mat)
x_exact_numpy = self.evaluate(x_exact)
with self.cached_session() as sess:
x_approx = linalg.conjugate_gradient(a_mat_ph, b_mat_ph)
x_approx_value = sess.run(x_approx,
feed_dict={
a_mat_ph: a_mat_value,
b_mat_ph: b_mat_value
})
self.assertAllClose(x_exact_numpy,
x_approx_value,
rtol=1e-4,
atol=1e-4)
def testConjugateGradientBasic(self, n, rhs):
x_obs = tf.constant(np.random.rand(n, 2),
dtype=tf.float32,
shape=[n, 2])
a_mat = tf.eye(n) + tf.matmul(x_obs, tf.linalg.matrix_transpose(x_obs))
x_exact = tf.constant(np.random.rand(n),
dtype=tf.float32,
shape=[n, 1])
b = tf.matmul(a_mat, x_exact)
x_approx = self.evaluate(linalg.conjugate_gradient(a_mat, b))
x_exact_numpy = self.evaluate(x_exact)
self.assertAllClose(x_exact_numpy, x_approx, rtol=1e-4, atol=1e-4)
def theta(self):
"""Returns the matrix of per-arm feature weights.
The returned matrix has shape (num_actions, context_dim).
It's equivalent to a stacking of theta vectors from the paper.
"""
thetas = []
for k in range(self._num_models):
thetas.append(
tf.squeeze(linalg.conjugate_gradient(
self._cov_matrix_list[k] + self._tikhonov_weight *
tf.eye(self._overall_context_dim, dtype=self._dtype),
tf.expand_dims(self._data_vector_list[k], axis=-1)),
axis=-1))
return tf.stack(thetas, axis=0)
def _get_actions_from_linucb(
self, encoded_observation: types.Float, mask: Optional[types.Tensor]
) -> Tuple[types.Int, types.Float, types.Float]:
encoded_observation = tf.cast(encoded_observation, dtype=self._dtype)
p_values = []
est_rewards = []
for k in range(self._num_actions):
encoded_observation_for_arm = self._get_encoded_observation_for_arm(
encoded_observation, k)
model_index = policy_utilities.get_model_index(
k, self._accepts_per_arm_features)
a_inv_x = linalg.conjugate_gradient(
self._cov_matrix[model_index] +
tf.eye(self._encoding_dim, dtype=self._dtype),
tf.linalg.matrix_transpose(encoded_observation_for_arm))
mean_reward_est = tf.einsum('j,jk->k',
self._data_vector[model_index],
a_inv_x)
est_rewards.append(mean_reward_est)
ci = tf.reshape(
tf.linalg.tensor_diag_part(
tf.matmul(encoded_observation_for_arm, a_inv_x)), [-1, 1])
p_values.append(
tf.reshape(mean_reward_est, [-1, 1]) +
self._alpha * tf.sqrt(ci))
stacked_p_values = tf.squeeze(tf.stack(p_values, axis=-1), axis=[1])
if mask is None:
chosen_actions = tf.argmax(stacked_p_values,
axis=-1,
output_type=tf.int32)
else:
chosen_actions = policy_utilities.masked_argmax(
stacked_p_values, mask, output_type=tf.int32)
est_mean_reward = tf.cast(tf.stack(est_rewards, axis=-1), tf.float32)
return chosen_actions, est_mean_reward, tf.cast(
stacked_p_values, tf.float32)
def _distribution(self, time_step, policy_state):
observation = time_step.observation
if self.observation_and_action_constraint_splitter is not None:
observation, _ = self.observation_and_action_constraint_splitter(
observation)
observation = tf.nest.map_structure(
lambda o: tf.cast(o, dtype=self._dtype), observation)
global_observation, arm_observations = self._split_observation(
observation)
if self._add_bias:
# The bias is added via a constant 1 feature.
global_observation = tf.concat([
global_observation,
tf.ones([tf.shape(global_observation)[0], 1],
dtype=self._dtype)
],
axis=1)
# Check the shape of the observation matrix. The observations can be
# batched.
if not global_observation.shape.is_compatible_with(
[None, self._global_context_dim]):
raise ValueError(
'Global observation shape is expected to be {}. Got {}.'.
format([None, self._global_context_dim],
global_observation.shape.as_list()))
global_observation = tf.reshape(global_observation,
[-1, self._global_context_dim])
est_rewards = []
confidence_intervals = []
for k in range(self._num_actions):
current_observation = self._get_current_observation(
global_observation, arm_observations, k)
model_index = policy_utilities.get_model_index(
k, self._accepts_per_arm_features)
if self._use_eigendecomp:
q_t_b = tf.matmul(
self._eig_matrix[model_index],
tf.linalg.matrix_transpose(current_observation),
transpose_a=True)
lambda_inv = tf.divide(
tf.ones_like(self._eig_vals[model_index]),
self._eig_vals[model_index] + self._tikhonov_weight)
a_inv_x = tf.matmul(self._eig_matrix[model_index],
tf.einsum('j,jk->jk', lambda_inv, q_t_b))
else:
a_inv_x = linalg.conjugate_gradient(
self._cov_matrix[model_index] + self._tikhonov_weight *
tf.eye(self._overall_context_dim, dtype=self._dtype),
tf.linalg.matrix_transpose(current_observation))
est_mean_reward = tf.einsum('j,jk->k',
self._data_vector[model_index],
a_inv_x)
est_rewards.append(est_mean_reward)
ci = tf.reshape(
tf.linalg.tensor_diag_part(
tf.matmul(current_observation, a_inv_x)), [-1, 1])
confidence_intervals.append(ci)
if self._exploration_strategy == ExplorationStrategy.optimistic:
optimistic_estimates = [
tf.reshape(mean_reward, [-1, 1]) +
self._alpha * tf.sqrt(confidence)
for mean_reward, confidence in zip(est_rewards,
confidence_intervals)
]
# Keeping the batch dimension during the squeeze, even if batch_size == 1.
rewards_for_argmax = tf.squeeze(tf.stack(optimistic_estimates,
axis=-1),
axis=[1])
elif self._exploration_strategy == ExplorationStrategy.sampling:
mu_sampler = tfd.Normal(
loc=tf.stack(est_rewards, axis=-1),
scale=self._alpha * tf.sqrt(
tf.squeeze(tf.stack(confidence_intervals, axis=-1),
axis=1)))
rewards_for_argmax = mu_sampler.sample()
else:
raise ValueError('Exploraton strategy %s not implemented.' %
self._exploration_strategy)
mask = constraints.construct_mask_from_multiple_sources(
time_step.observation,
self._observation_and_action_constraint_splitter, (),
self._num_actions)
if mask is not None:
chosen_actions = policy_utilities.masked_argmax(
rewards_for_argmax,
mask,
output_type=tf.nest.flatten(self._action_spec)[0].dtype)
else:
chosen_actions = tf.argmax(rewards_for_argmax,
axis=-1,
output_type=tf.nest.flatten(
self._action_spec)[0].dtype)
action_distributions = tfp.distributions.Deterministic(
loc=chosen_actions)
policy_info = policy_utilities.populate_policy_info(
arm_observations, chosen_actions, rewards_for_argmax,
tf.stack(est_rewards, axis=-1), self._emit_policy_info,
self._accepts_per_arm_features)
return policy_step.PolicyStep(action_distributions, policy_state,
policy_info)