add policy_utils

compiler_opt/es/policy_utils.py

@@ -0,0 +1,110 @@
+# coding=utf-8
+# Copyright 2020 Google LLC
+#
+# 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.
+"""Util functions to create and edit a tf_agent policy."""
+
+import gin
+import numpy as np
+import numpy.typing as npt
+import tensorflow as tf
+from typing import Protocol, Sequence
+
+from compiler_opt.rl import policy_saver, registry
+from tf_agents.networks import network
+from tf_agents.policies import actor_policy, greedy_policy, tf_policy
+
+
+class HasModelVariables(Protocol):
+  model_variables: Sequence[tf.Variable]
+
+
+# TODO(abenalaast): Issue #280
+@gin.configurable(module='policy_utils')
+def create_actor_policy(actor_network_ctor: network.DistributionNetwork,
+                        greedy: bool = False) -> tf_policy.TFPolicy:
+  """Creates an actor policy."""
+  problem_config = registry.get_configuration()
+  time_step_spec, action_spec = problem_config.get_signature_spec()
+  layers = tf.nest.map_structure(
+      problem_config.get_preprocessing_layer_creator(),
+      time_step_spec.observation)
+
+  actor_network = actor_network_ctor(
+      input_tensor_spec=time_step_spec.observation,
+      output_tensor_spec=action_spec,
+      preprocessing_layers=layers)
+
+  policy = actor_policy.ActorPolicy(
+      time_step_spec=time_step_spec,
+      action_spec=action_spec,
+      actor_network=actor_network)
+
+  if greedy:
+    policy = greedy_policy.GreedyPolicy(policy)
+
+  return policy
+
+
+def get_vectorized_parameters_from_policy(
+    policy: 'tf_policy.TFPolicy | HasModelVariables'
+) -> npt.NDArray[np.float32]:
+  """Returns a policy's variable values as a single np array."""
+  if isinstance(policy, tf_policy.TFPolicy):
+    variables = policy.variables()
+  elif hasattr(policy, 'model_variables'):
+    variables = policy.model_variables
+  else:
+    raise ValueError(f'Policy must be a TFPolicy or a loaded SavedModel. '
+                     f'Passed policy: {policy}')
+
+  parameters = [var.numpy().flatten() for var in variables]
+  parameters = np.concatenate(parameters, axis=0)
+  return parameters
+
+
+def set_vectorized_parameters_for_policy(
+    policy: 'tf_policy.TFPolicy | HasModelVariables',
+    parameters: npt.NDArray[np.float32]) -> None:
+  """Separates values in parameters into the policy's shapes
+  and sets the policy variables to those values"""
+  if isinstance(policy, tf_policy.TFPolicy):
+    variables = policy.variables()
+  elif hasattr(policy, 'model_variables'):
+    variables = policy.model_variables
+  else:
+    raise ValueError(f'Policy must be a TFPolicy or a loaded SavedModel. '
+                     f'Passed policy: {policy}')
+
+  param_pos = 0
+  for variable in variables:
+    shape = tf.shape(variable).numpy()
+    num_elems = np.prod(shape)
+    param = np.reshape(parameters[param_pos:param_pos + num_elems], shape)
+    variable.assign(param)
+    param_pos += num_elems
+  if param_pos != len(parameters):
+    raise ValueError(
+        f'Parameter dimensions are not matched! Expected {len(parameters)} '
+        f'but only found {param_pos}.')
+
+
+def save_policy(policy: 'tf_policy.TFPolicy | HasModelVariables',
+                parameters: npt.NDArray[np.float32], save_folder: str,
+                policy_name: str) -> None:
+  """Assigns a policy the name policy_name
+  and saves it to the directory of save_folder
+  with the values in parameters."""
+  set_vectorized_parameters_for_policy(policy, parameters)
+  saver = policy_saver.PolicySaver({policy_name: policy})
+  saver.save(save_folder)

compiler_opt/es/policy_utils_test.py

@@ -0,0 +1,247 @@
+# coding=utf-8
+# Copyright 2020 Google LLC
+#
+# 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 policy_utils."""
+
+from absl.testing import absltest
+import numpy as np
+import os
+import tensorflow as tf
+from tf_agents.networks import actor_distribution_network
+from tf_agents.policies import actor_policy, tf_policy
+
+from compiler_opt.es import policy_utils
+from compiler_opt.rl import policy_saver, registry
+from compiler_opt.rl.inlining import config as inlining_config
+from compiler_opt.rl.inlining import InliningConfig
+from compiler_opt.rl.regalloc import config as regalloc_config
+from compiler_opt.rl.regalloc import RegallocEvictionConfig, regalloc_network
+
+
+# TODO(abenalaast): Issue #280
+class ConfigTest(absltest.TestCase):
+
+  def test_inlining_config(self):
+    problem_config = registry.get_configuration(implementation=InliningConfig)
+    time_step_spec, action_spec = problem_config.get_signature_spec()
+    creator = inlining_config.get_observation_processing_layer_creator(
+        quantile_file_dir='compiler_opt/rl/inlining/vocab/',
+        with_sqrt=False,
+        with_z_score_normalization=False)
+    layers = tf.nest.map_structure(creator, time_step_spec.observation)
+
+    actor_network = actor_distribution_network.ActorDistributionNetwork(
+        input_tensor_spec=time_step_spec.observation,
+        output_tensor_spec=action_spec,
+        preprocessing_layers=layers,
+        preprocessing_combiner=tf.keras.layers.Concatenate(),
+        fc_layer_params=(64, 64, 64, 64),
+        dropout_layer_params=None,
+        activation_fn=tf.keras.activations.relu)
+
+    policy = actor_policy.ActorPolicy(
+        time_step_spec=time_step_spec,
+        action_spec=action_spec,
+        actor_network=actor_network)
+
+    self.assertIsNotNone(policy)
+    self.assertIsInstance(
+        policy._actor_network,  # pylint: disable=protected-access
+        actor_distribution_network.ActorDistributionNetwork)
+
+  def test_regalloc_config(self):
+    problem_config = registry.get_configuration(
+        implementation=RegallocEvictionConfig)
+    time_step_spec, action_spec = problem_config.get_signature_spec()
+    creator = regalloc_config.get_observation_processing_layer_creator(
+        quantile_file_dir='compiler_opt/rl/regalloc/vocab',
+        with_sqrt=False,
+        with_z_score_normalization=False)
+    layers = tf.nest.map_structure(creator, time_step_spec.observation)
+
+    actor_network = regalloc_network.RegAllocNetwork(
+        input_tensor_spec=time_step_spec.observation,
+        output_tensor_spec=action_spec,
+        preprocessing_layers=layers,
+        preprocessing_combiner=tf.keras.layers.Concatenate(),
+        fc_layer_params=(64, 64, 64, 64),
+        dropout_layer_params=None,
+        activation_fn=tf.keras.activations.relu)
+
+    policy = actor_policy.ActorPolicy(
+        time_step_spec=time_step_spec,
+        action_spec=action_spec,
+        actor_network=actor_network)
+
+    self.assertIsNotNone(policy)
+    self.assertIsInstance(
+        policy._actor_network,  # pylint: disable=protected-access
+        regalloc_network.RegAllocNetwork)
+
+
+class VectorTest(absltest.TestCase):
+
+  expected_variable_shapes = [(71, 64), (64), (64, 64), (64), (64, 64), (64),
+                              (64, 64), (64), (64, 2), (2)]
+  expected_length_of_a_perturbation = sum(
+      np.prod(shape) for shape in expected_variable_shapes)
+  params = np.arange(expected_length_of_a_perturbation, dtype=np.float32)
+
+  def test_set_vectorized_parameters_for_policy(self):
+    # create a policy
+    problem_config = registry.get_configuration(implementation=InliningConfig)
+    time_step_spec, action_spec = problem_config.get_signature_spec()
+    creator = inlining_config.get_observation_processing_layer_creator(
+        quantile_file_dir='compiler_opt/rl/inlining/vocab/',
+        with_sqrt=False,
+        with_z_score_normalization=False)
+    layers = tf.nest.map_structure(creator, time_step_spec.observation)
+
+    actor_network = actor_distribution_network.ActorDistributionNetwork(
+        input_tensor_spec=time_step_spec.observation,
+        output_tensor_spec=action_spec,
+        preprocessing_layers=layers,
+        preprocessing_combiner=tf.keras.layers.Concatenate(),
+        fc_layer_params=(64, 64, 64, 64),
+        dropout_layer_params=None,
+        activation_fn=tf.keras.activations.relu)
+
+    policy = actor_policy.ActorPolicy(
+        time_step_spec=time_step_spec,
+        action_spec=action_spec,
+        actor_network=actor_network)
+
+    # save the policy
+    saver = policy_saver.PolicySaver({'policy': policy})
+    testing_path = self.create_tempdir()
+    policy_save_path = os.path.join(testing_path, 'temp_output/policy')
+    saver.save(policy_save_path)
+
+    # set the values of the policy variables
+    policy_utils.set_vectorized_parameters_for_policy(policy, VectorTest.params)
+    # iterate through variables and check their shapes and values
+    expected_values = [*VectorTest.params]
+    for i, variable in enumerate(policy.variables()):  # pylint: disable=not-callable
+      self.assertEqual(variable.shape, VectorTest.expected_variable_shapes[i])
+      variable_values = variable.numpy().flatten()
+      np.testing.assert_array_almost_equal(
+          expected_values[:len(variable_values)], variable_values)
+      expected_values = expected_values[len(variable_values):]
+    self.assertEmpty(expected_values)
+
+    # get saved model to test a loaded policy
+    sm = tf.saved_model.load(policy_save_path + '/policy')
+    self.assertNotIsInstance(sm, tf_policy.TFPolicy)
+    policy_utils.set_vectorized_parameters_for_policy(sm, VectorTest.params)
+    expected_values = [*VectorTest.params]
+    for i, variable in enumerate(sm.model_variables):
+      self.assertEqual(variable.shape, VectorTest.expected_variable_shapes[i])
+      variable_values = variable.numpy().flatten()
+      np.testing.assert_array_almost_equal(
+          expected_values[:len(variable_values)], variable_values)
+      expected_values = expected_values[len(variable_values):]
+    self.assertEmpty(expected_values)
+
+  def test_get_vectorized_parameters_from_policy(self):
+    # create a policy
+    problem_config = registry.get_configuration(implementation=InliningConfig)
+    time_step_spec, action_spec = problem_config.get_signature_spec()
+    creator = inlining_config.get_observation_processing_layer_creator(
+        quantile_file_dir='compiler_opt/rl/inlining/vocab/',
+        with_sqrt=False,
+        with_z_score_normalization=False)
+    layers = tf.nest.map_structure(creator, time_step_spec.observation)
+
+    actor_network = actor_distribution_network.ActorDistributionNetwork(
+        input_tensor_spec=time_step_spec.observation,
+        output_tensor_spec=action_spec,
+        preprocessing_layers=layers,
+        preprocessing_combiner=tf.keras.layers.Concatenate(),
+        fc_layer_params=(64, 64, 64, 64),
+        dropout_layer_params=None,
+        activation_fn=tf.keras.activations.relu)
+
+    policy = actor_policy.ActorPolicy(
+        time_step_spec=time_step_spec,
+        action_spec=action_spec,
+        actor_network=actor_network)
+
+    # save the policy
+    saver = policy_saver.PolicySaver({'policy': policy})
+    testing_path = self.create_tempdir()
+    policy_save_path = os.path.join(testing_path, 'temp_output/policy')
+    saver.save(policy_save_path)
+
+    # functionality verified in previous test
+    policy_utils.set_vectorized_parameters_for_policy(policy, VectorTest.params)
+    # vectorize and check if the outcome is the same as the start
+    output = policy_utils.get_vectorized_parameters_from_policy(policy)
+    np.testing.assert_array_almost_equal(output, VectorTest.params)
+
+    # get saved model to test a loaded policy
+    sm = tf.saved_model.load(policy_save_path + '/policy')
+    self.assertNotIsInstance(sm, tf_policy.TFPolicy)
+    policy_utils.set_vectorized_parameters_for_policy(sm, VectorTest.params)
+    # vectorize and check if the outcome is the same as the start
+    output = policy_utils.get_vectorized_parameters_from_policy(sm)
+    np.testing.assert_array_almost_equal(output, VectorTest.params)
+
+  def test_tfpolicy_and_loaded_policy_produce_same_variable_order(self):
+    # create a policy
+    problem_config = registry.get_configuration(implementation=InliningConfig)
+    time_step_spec, action_spec = problem_config.get_signature_spec()
+    creator = inlining_config.get_observation_processing_layer_creator(
+        quantile_file_dir='compiler_opt/rl/inlining/vocab/',
+        with_sqrt=False,
+        with_z_score_normalization=False)
+    layers = tf.nest.map_structure(creator, time_step_spec.observation)
+
+    actor_network = actor_distribution_network.ActorDistributionNetwork(
+        input_tensor_spec=time_step_spec.observation,
+        output_tensor_spec=action_spec,
+        preprocessing_layers=layers,
+        preprocessing_combiner=tf.keras.layers.Concatenate(),
+        fc_layer_params=(64, 64, 64, 64),
+        dropout_layer_params=None,
+        activation_fn=tf.keras.activations.relu)
+
+    policy = actor_policy.ActorPolicy(
+        time_step_spec=time_step_spec,
+        action_spec=action_spec,
+        actor_network=actor_network)
+
+    # save the policy
+    saver = policy_saver.PolicySaver({'policy': policy})
+    testing_path = self.create_tempdir()
+    policy_save_path = os.path.join(testing_path, 'temp_output/policy')
+    saver.save(policy_save_path)
+
+    # set the values of the variables
+    policy_utils.set_vectorized_parameters_for_policy(policy, VectorTest.params)
+    # save the changes
+    saver.save(policy_save_path)
+    # vectorize the tfpolicy
+    tf_params = policy_utils.get_vectorized_parameters_from_policy(policy)
+
+    # get loaded policy
+    sm = tf.saved_model.load(policy_save_path + '/policy')
+    # vectorize the loaded policy
+    loaded_params = policy_utils.get_vectorized_parameters_from_policy(sm)
+
+    # assert that they result in the same order of values
+    np.testing.assert_array_almost_equal(tf_params, loaded_params)
+
+
+if __name__ == '__main__':
+  absltest.main()