def get_cartpole_env_and_specs(): env = suite_gym.load('CartPole-v0') _, action_tensor_spec, time_step_tensor_spec = ( spec_utils.get_tensor_specs(env)) return env, action_tensor_spec, time_step_tensor_spec
def _build_components(self, rb_port): env = suite_gym.load('CartPole-v0') observation_tensor_spec, action_tensor_spec, time_step_tensor_spec = ( spec_utils.get_tensor_specs(env)) train_step = train_utils.create_train_step() q_net = q_network.QNetwork(observation_tensor_spec, action_tensor_spec, fc_layer_params=(100, )) agent = dqn_agent.DqnAgent( time_step_tensor_spec, action_tensor_spec, q_network=q_net, optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=0.001), train_step_counter=train_step) replay_buffer, rb_observer = ( replay_buffer_utils.get_reverb_buffer_and_observer( agent.collect_data_spec, sequence_length=2, replay_capacity=1000, port=rb_port)) return env, agent, train_step, replay_buffer, rb_observer
def test_get_tensor_specs(self): collect_env = suite_gym.load('CartPole-v0') observation_spec, action_spec, time_step_spec = ( spec_utils.get_tensor_specs(collect_env)) self.assertEqual(observation_spec.name, 'observation') self.assertEqual(action_spec.name, 'action') self.assertEqual(time_step_spec.observation.name, 'observation') self.assertEqual(time_step_spec.reward.name, 'reward')
def train_eval( root_dir, env_name='HalfCheetah-v2', # Training params num_iterations=20000, actor_fc_layers=(64, 64), value_fc_layers=(64, 64), learning_rate=3e-4, collect_sequence_length=2048, minibatch_size=64, num_epochs=10, # Agent params importance_ratio_clipping=0.2, lambda_value=0.95, discount_factor=0.99, entropy_regularization=0., value_pred_loss_coef=0.5, use_gae=True, use_td_lambda_return=True, gradient_clipping=None, value_clipping=None, # Replay params reverb_port=None, replay_capacity=10000, # Others policy_save_interval=5000, summary_interval=1000, eval_interval=10000, eval_episodes=30, debug_summaries=False, summarize_grads_and_vars=False): """Trains and evaluates PPO (Importance Ratio Clipping). Args: root_dir: Main directory path where checkpoints, saved_models, and summaries will be written to. env_name: Name for the Mujoco environment to load. num_iterations: The number of iterations to perform collection and training. actor_fc_layers: List of fully_connected parameters for the actor network, where each item is the number of units in the layer. value_fc_layers: : List of fully_connected parameters for the value network, where each item is the number of units in the layer. learning_rate: Learning rate used on the Adam optimizer. collect_sequence_length: Number of steps to take in each collect run. minibatch_size: Number of elements in each mini batch. If `None`, the entire collected sequence will be treated as one batch. num_epochs: Number of iterations to repeat over all collected data per data collection step. (Schulman,2017) sets this to 10 for Mujoco, 15 for Roboschool and 3 for Atari. importance_ratio_clipping: Epsilon in clipped, surrogate PPO objective. For more detail, see explanation at the top of the doc. lambda_value: Lambda parameter for TD-lambda computation. discount_factor: Discount factor for return computation. Default to `0.99` which is the value used for all environments from (Schulman, 2017). entropy_regularization: Coefficient for entropy regularization loss term. Default to `0.0` because no entropy bonus was used in (Schulman, 2017). value_pred_loss_coef: Multiplier for value prediction loss to balance with policy gradient loss. Default to `0.5`, which was used for all environments in the OpenAI baseline implementation. This parameters is irrelevant unless you are sharing part of actor_net and value_net. In that case, you would want to tune this coeeficient, whose value depends on the network architecture of your choice. use_gae: If True (default False), uses generalized advantage estimation for computing per-timestep advantage. Else, just subtracts value predictions from empirical return. use_td_lambda_return: If True (default False), uses td_lambda_return for training value function; here: `td_lambda_return = gae_advantage + value_predictions`. `use_gae` must be set to `True` as well to enable TD -lambda returns. If `use_td_lambda_return` is set to True while `use_gae` is False, the empirical return will be used and a warning will be logged. gradient_clipping: Norm length to clip gradients. value_clipping: Difference between new and old value predictions are clipped to this threshold. Value clipping could be helpful when training very deep networks. Default: no clipping. reverb_port: Port for reverb server, if None, use a randomly chosen unused port. replay_capacity: The maximum number of elements for the replay buffer. Items will be wasted if this is smalled than collect_sequence_length. policy_save_interval: How often, in train_steps, the policy will be saved. summary_interval: How often to write data into Tensorboard. eval_interval: How often to run evaluation, in train_steps. eval_episodes: Number of episodes to evaluate over. debug_summaries: Boolean for whether to gather debug summaries. summarize_grads_and_vars: If true, gradient summaries will be written. """ collect_env = suite_mujoco.load(env_name) eval_env = suite_mujoco.load(env_name) num_environments = 1 observation_tensor_spec, action_tensor_spec, time_step_tensor_spec = ( spec_utils.get_tensor_specs(collect_env)) train_step = train_utils.create_train_step() actor_net = actor_distribution_network.ActorDistributionNetwork( observation_tensor_spec, action_tensor_spec, fc_layer_params=actor_fc_layers, activation_fn=tf.nn.tanh, kernel_initializer=tf.keras.initializers.Orthogonal()) value_net = value_network.ValueNetwork( observation_tensor_spec, fc_layer_params=value_fc_layers, kernel_initializer=tf.keras.initializers.Orthogonal()) current_iteration = tf.Variable(0, dtype=tf.int64) def learning_rate_fn(): # Linearly decay the learning rate. return learning_rate * (1 - current_iteration / num_iterations) agent = ppo_clip_agent.PPOClipAgent( time_step_tensor_spec, action_tensor_spec, optimizer=tf.compat.v1.train.AdamOptimizer( learning_rate=learning_rate_fn, epsilon=1e-5), actor_net=actor_net, value_net=value_net, importance_ratio_clipping=importance_ratio_clipping, lambda_value=lambda_value, discount_factor=discount_factor, entropy_regularization=entropy_regularization, value_pred_loss_coef=value_pred_loss_coef, # This is a legacy argument for the number of times we repeat the data # inside of the train function, incompatible with mini batch learning. # We set the epoch number from the replay buffer and tf.Data instead. num_epochs=1, use_gae=use_gae, use_td_lambda_return=use_td_lambda_return, gradient_clipping=gradient_clipping, value_clipping=value_clipping, # TODO(b/150244758): Default compute_value_and_advantage_in_train to False # after Reverb open source. compute_value_and_advantage_in_train=False, # Skips updating normalizers in the agent, as it's handled in the learner. update_normalizers_in_train=False, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=train_step) agent.initialize() reverb_server = reverb.Server( [ reverb.Table( # Replay buffer storing experience for training. name='training_table', sampler=reverb.selectors.Fifo(), remover=reverb.selectors.Fifo(), rate_limiter=reverb.rate_limiters.MinSize(1), max_size=replay_capacity, max_times_sampled=1, ), reverb. Table( # Replay buffer storing experience for normalization. name='normalization_table', sampler=reverb.selectors.Fifo(), remover=reverb.selectors.Fifo(), rate_limiter=reverb.rate_limiters.MinSize(1), max_size=replay_capacity, max_times_sampled=1, ) ], port=reverb_port) # Create the replay buffer. reverb_replay_train = reverb_replay_buffer.ReverbReplayBuffer( agent.collect_data_spec, sequence_length=collect_sequence_length, table_name='training_table', server_address='localhost:{}'.format(reverb_server.port), # The only collected sequence is used to populate the batches. max_cycle_length=1, rate_limiter_timeout_ms=1000) reverb_replay_normalization = reverb_replay_buffer.ReverbReplayBuffer( agent.collect_data_spec, sequence_length=collect_sequence_length, table_name='normalization_table', server_address='localhost:{}'.format(reverb_server.port), # The only collected sequence is used to populate the batches. max_cycle_length=1, rate_limiter_timeout_ms=1000) rb_observer = reverb_utils.ReverbTrajectorySequenceObserver( reverb_replay_train.py_client, ['training_table', 'normalization_table'], sequence_length=collect_sequence_length, stride_length=collect_sequence_length) saved_model_dir = os.path.join(root_dir, learner.POLICY_SAVED_MODEL_DIR) collect_env_step_metric = py_metrics.EnvironmentSteps() learning_triggers = [ triggers.PolicySavedModelTrigger(saved_model_dir, agent, train_step, interval=policy_save_interval, metadata_metrics={ triggers.ENV_STEP_METADATA_KEY: collect_env_step_metric }), triggers.StepPerSecondLogTrigger(train_step, interval=summary_interval), ] def training_dataset_fn(): return reverb_replay_train.as_dataset( sample_batch_size=num_environments, sequence_preprocess_fn=agent.preprocess_sequence) def normalization_dataset_fn(): return reverb_replay_normalization.as_dataset( sample_batch_size=num_environments, sequence_preprocess_fn=agent.preprocess_sequence) agent_learner = ppo_learner.PPOLearner( root_dir, train_step, agent, experience_dataset_fn=training_dataset_fn, normalization_dataset_fn=normalization_dataset_fn, num_batches=1, num_epochs=num_epochs, minibatch_size=minibatch_size, shuffle_buffer_size=collect_sequence_length, triggers=learning_triggers) tf_collect_policy = agent.collect_policy collect_policy = py_tf_eager_policy.PyTFEagerPolicy(tf_collect_policy, use_tf_function=True) collect_actor = actor.Actor(collect_env, collect_policy, train_step, steps_per_run=collect_sequence_length, observers=[rb_observer], metrics=actor.collect_metrics(buffer_size=10) + [collect_env_step_metric], reference_metrics=[collect_env_step_metric], summary_dir=os.path.join( root_dir, learner.TRAIN_DIR), summary_interval=summary_interval) tf_greedy_policy = agent.policy greedy_policy = py_tf_eager_policy.PyTFEagerPolicy(tf_greedy_policy, use_tf_function=True) if eval_interval: logging.info('Intial evaluation.') eval_actor = actor.Actor(eval_env, greedy_policy, train_step, metrics=actor.eval_metrics(eval_episodes), summary_dir=os.path.join(root_dir, 'eval'), episodes_per_run=eval_episodes) eval_actor.run_and_log() logging.info('Training.') for _ in range(num_iterations): collect_actor.run() # TODO(b/159615593): Update to use observer.flush. # Reset the reverb observer to make sure the data collected is flushed and # written to the RB. rb_observer.reset() agent_learner.run() reverb_replay_train.clear() reverb_replay_normalization.clear() current_iteration.assign_add(1) if eval_interval and agent_learner.train_step_numpy % eval_interval == 0: logging.info('Evaluating.') eval_actor.run_and_log() rb_observer.close() reverb_server.stop()
eval_interval = 1000 # @param {type:"integer"} policy_save_interval = 5000 # @param {type:"integer"} env = get_tf_wrapped_robo_rugby_env() print('Observation Spec:') print(env.time_step_spec().observation) print('Action Spec:') print(env.action_spec()) collect_env = get_tf_wrapped_robo_rugby_env() eval_env = get_tf_wrapped_robo_rugby_env() objStrategy = strategy_utils.get_strategy(tpu=False, use_gpu=True) specObservation, specAction, specTimeStep = ( spec_utils.get_tensor_specs(collect_env)) with objStrategy.scope(): # Critic network trains the Actor network nnCritic = critic_network.CriticNetwork( (specObservation, specAction), observation_fc_layer_params=None, action_fc_layer_params=None, joint_fc_layer_params=HyperParms.critic_joint_fc_layer_params, kernel_initializer='glorot_uniform', last_kernel_initializer='glorot_uniform') with objStrategy.scope(): nnActor = actor_distribution_network.ActorDistributionNetwork( specObservation, specAction,
def train( root_dir, strategy, replay_buffer_server_address, variable_container_server_address, create_agent_fn, create_env_fn, # Training params learning_rate=3e-4, batch_size=256, num_iterations=32000, learner_iterations_per_call=100): """Trains a DQN agent.""" # Get the specs from the environment. logging.info('Training SAC with learning rate: %f', learning_rate) env = create_env_fn() observation_tensor_spec, action_tensor_spec, time_step_tensor_spec = ( spec_utils.get_tensor_specs(env)) # Create the agent. with strategy.scope(): train_step = train_utils.create_train_step() agent = create_agent_fn(train_step, observation_tensor_spec, action_tensor_spec, time_step_tensor_spec, learning_rate) agent.initialize() # Create the policy saver which saves the initial model now, then it # periodically checkpoints the policy weigths. saved_model_dir = os.path.join(root_dir, learner.POLICY_SAVED_MODEL_DIR) save_model_trigger = triggers.PolicySavedModelTrigger( saved_model_dir, agent, train_step, interval=1000) # Create the variable container. variables = { reverb_variable_container.POLICY_KEY: agent.collect_policy.variables(), reverb_variable_container.TRAIN_STEP_KEY: train_step } variable_container = reverb_variable_container.ReverbVariableContainer( variable_container_server_address, table_names=[reverb_variable_container.DEFAULT_TABLE]) variable_container.push(variables) # Create the replay buffer. reverb_replay = reverb_replay_buffer.ReverbReplayBuffer( agent.collect_data_spec, sequence_length=2, table_name=reverb_replay_buffer.DEFAULT_TABLE, server_address=replay_buffer_server_address) # Initialize the dataset. def experience_dataset_fn(): with strategy.scope(): return reverb_replay.as_dataset( sample_batch_size=batch_size, num_steps=2).prefetch(3) # Create the learner. learning_triggers = [ save_model_trigger, triggers.StepPerSecondLogTrigger(train_step, interval=1000) ] sac_learner = learner.Learner( root_dir, train_step, agent, experience_dataset_fn, triggers=learning_triggers, strategy=strategy) # Run the training loop. # TODO(b/162440911) change the loop use train_step to handle preemptions for _ in range(num_iterations): sac_learner.run(iterations=learner_iterations_per_call) variable_container.push(variables)
target_update_period = 1 gamma = 0.99 reward_scale_factor = 1.0 actor_fc_layer_params = (256, 256) critic_joint_fc_layer_params = (256, 256) log_interval = 5000 num_eval_episodes = 20 eval_interval = 10000 policy_save_interval = 5000 collect_env = tf_py_environment.TFPyEnvironment(suite_pybullet.load(env_name)) eval_env = tf_py_environment.TFPyEnvironment(suite_pybullet.load(env_name)) observation_spec, action_spec, time_step_spec = (spec_utils.get_tensor_specs(collect_env)) critic_net = critic_network.CriticNetwork( (observation_spec, action_spec), observation_fc_layer_params=None, action_fc_layer_params=None, joint_fc_layer_params=critic_joint_fc_layer_params, kernel_initializer=tf.keras.initializers.HeNormal(), last_kernel_initializer=tf.keras.initializers.HeNormal() ) actor_net = actor_distribution_network.ActorDistributionNetwork( observation_spec, action_spec, fc_layer_params=actor_fc_layer_params, continuous_projection_net=(tanh_normal_projection_network.TanhNormalProjectionNetwork)
def train_eval( root_dir, env_name='HalfCheetah-v2', # Training params initial_collect_steps=10000, num_iterations=3200000, actor_fc_layers=(256, 256), critic_obs_fc_layers=None, critic_action_fc_layers=None, critic_joint_fc_layers=(256, 256), # Agent params batch_size=256, actor_learning_rate=3e-4, critic_learning_rate=3e-4, alpha_learning_rate=3e-4, gamma=0.99, target_update_tau=0.005, target_update_period=1, reward_scale_factor=0.1, # Replay params reverb_port=None, replay_capacity=1000000, # Others # Defaults to not checkpointing saved policy. If you wish to enable this, # please note the caveat explained in README.md. policy_save_interval=-1, eval_interval=10000, eval_episodes=30, debug_summaries=False, summarize_grads_and_vars=False): """Trains and evaluates SAC.""" logging.info('Training SAC on: %s', env_name) collect_env = suite_mujoco.load(env_name) eval_env = suite_mujoco.load(env_name) observation_tensor_spec, action_tensor_spec, time_step_tensor_spec = ( spec_utils.get_tensor_specs(collect_env)) train_step = train_utils.create_train_step() actor_net = actor_distribution_network.ActorDistributionNetwork( observation_tensor_spec, action_tensor_spec, fc_layer_params=actor_fc_layers, continuous_projection_net=tanh_normal_projection_network. TanhNormalProjectionNetwork) critic_net = critic_network.CriticNetwork( (observation_tensor_spec, action_tensor_spec), observation_fc_layer_params=critic_obs_fc_layers, action_fc_layer_params=critic_action_fc_layers, joint_fc_layer_params=critic_joint_fc_layers, kernel_initializer='glorot_uniform', last_kernel_initializer='glorot_uniform') agent = sac_agent.SacAgent( time_step_tensor_spec, action_tensor_spec, actor_network=actor_net, critic_network=critic_net, actor_optimizer=tf.compat.v1.train.AdamOptimizer( learning_rate=actor_learning_rate), critic_optimizer=tf.compat.v1.train.AdamOptimizer( learning_rate=critic_learning_rate), alpha_optimizer=tf.compat.v1.train.AdamOptimizer( learning_rate=alpha_learning_rate), target_update_tau=target_update_tau, target_update_period=target_update_period, td_errors_loss_fn=tf.math.squared_difference, gamma=gamma, reward_scale_factor=reward_scale_factor, gradient_clipping=None, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=train_step) agent.initialize() table_name = 'uniform_table' table = reverb.Table(table_name, max_size=replay_capacity, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), rate_limiter=reverb.rate_limiters.MinSize(1)) reverb_server = reverb.Server([table], port=reverb_port) reverb_replay = reverb_replay_buffer.ReverbReplayBuffer( agent.collect_data_spec, sequence_length=2, table_name=table_name, local_server=reverb_server) rb_observer = reverb_utils.ReverbAddTrajectoryObserver( reverb_replay.py_client, table_name, sequence_length=2, stride_length=1) dataset = reverb_replay.as_dataset(sample_batch_size=batch_size, num_steps=2).prefetch(50) experience_dataset_fn = lambda: dataset saved_model_dir = os.path.join(root_dir, learner.POLICY_SAVED_MODEL_DIR) env_step_metric = py_metrics.EnvironmentSteps() learning_triggers = [ triggers.PolicySavedModelTrigger( saved_model_dir, agent, train_step, interval=policy_save_interval, metadata_metrics={triggers.ENV_STEP_METADATA_KEY: env_step_metric}), triggers.StepPerSecondLogTrigger(train_step, interval=1000), ] agent_learner = learner.Learner(root_dir, train_step, agent, experience_dataset_fn, triggers=learning_triggers) random_policy = random_py_policy.RandomPyPolicy( collect_env.time_step_spec(), collect_env.action_spec()) initial_collect_actor = actor.Actor(collect_env, random_policy, train_step, steps_per_run=initial_collect_steps, observers=[rb_observer]) logging.info('Doing initial collect.') initial_collect_actor.run() tf_collect_policy = agent.collect_policy collect_policy = py_tf_eager_policy.PyTFEagerPolicy(tf_collect_policy, use_tf_function=True) collect_actor = actor.Actor(collect_env, collect_policy, train_step, steps_per_run=1, metrics=actor.collect_metrics(10), summary_dir=os.path.join( root_dir, learner.TRAIN_DIR), observers=[rb_observer, env_step_metric]) tf_greedy_policy = greedy_policy.GreedyPolicy(agent.policy) eval_greedy_policy = py_tf_eager_policy.PyTFEagerPolicy( tf_greedy_policy, use_tf_function=True) eval_actor = actor.Actor( eval_env, eval_greedy_policy, train_step, episodes_per_run=eval_episodes, metrics=actor.eval_metrics(eval_episodes), summary_dir=os.path.join(root_dir, 'eval'), ) if eval_interval: logging.info('Evaluating.') eval_actor.run_and_log() logging.info('Training.') for _ in range(num_iterations): collect_actor.run() agent_learner.run(iterations=1) if eval_interval and agent_learner.train_step_numpy % eval_interval == 0: logging.info('Evaluating.') eval_actor.run_and_log() rb_observer.close() reverb_server.stop()
policy_save_interval = 5000 #####LOAD ENVIRONMENT ##### env_name = "MinitaurBulletEnv-v0" env = suite_pybullet.load(env_name) ####TWO ENV instantiated. One for Train, One for Eval ###### train_py_env = suite_gym.load(env_name) eval_py_env = suite_gym.load(env_name) #Converts Numpy Arrays to Tensors, so they are compatible with Tensorflow agents and policies train_env = tf_py_environment.TFPyEnvironment(train_py_env) eval_env = tf_py_environment.TFPyEnvironment(eval_py_env) time_step = env.reset() observation_spec, action_spec, time_step_spec = ( spec_utils.get_tensor_specs(train_env)) #######Networks##### #conv_layer_params = [(32,3,3),(32,3,3),(32,3,3)] conv_layer_params = None fc_layer_params = (400, 300) kernel_initializer = tf.keras.initializers.VarianceScaling( scale=1. / 3., mode='fan_in', distribution='uniform') final_layer_initializer = tf.keras.initializers.RandomUniform(minval=-0.0003, maxval=0.0003) actor_net = actor_network.ActorNetwork( observation_spec, action_spec, conv_layer_params=conv_layer_params, fc_layer_params=fc_layer_params,
def train_eval( root_dir, env_name, # Training params train_sequence_length, initial_collect_steps=1000, collect_steps_per_iteration=1, num_iterations=100000, # RNN params. q_network_fn=q_lstm_network, # defaults to q_lstm_network. # Agent params epsilon_greedy=0.1, batch_size=64, learning_rate=1e-3, gamma=0.99, target_update_tau=0.05, target_update_period=5, reward_scale_factor=1.0, # Replay params reverb_port=None, replay_capacity=100000, # Others policy_save_interval=1000, eval_interval=1000, eval_episodes=10): """Trains and evaluates DQN.""" collect_env = suite_gym.load(env_name) eval_env = suite_gym.load(env_name) unused_observation_tensor_spec, action_tensor_spec, time_step_tensor_spec = ( spec_utils.get_tensor_specs(collect_env)) train_step = train_utils.create_train_step() num_actions = action_tensor_spec.maximum - action_tensor_spec.minimum + 1 q_net = q_network_fn(num_actions=num_actions) sequence_length = train_sequence_length + 1 agent = dqn_agent.DqnAgent( time_step_tensor_spec, action_tensor_spec, q_network=q_net, epsilon_greedy=epsilon_greedy, # n-step updates aren't supported with RNNs yet. n_step_update=1, target_update_tau=target_update_tau, target_update_period=target_update_period, optimizer=tf.compat.v1.train.AdamOptimizer( learning_rate=learning_rate), td_errors_loss_fn=common.element_wise_squared_loss, gamma=gamma, reward_scale_factor=reward_scale_factor, train_step_counter=train_step) table_name = 'uniform_table' table = reverb.Table(table_name, max_size=replay_capacity, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), rate_limiter=reverb.rate_limiters.MinSize(1)) reverb_server = reverb.Server([table], port=reverb_port) reverb_replay = reverb_replay_buffer.ReverbReplayBuffer( agent.collect_data_spec, sequence_length=sequence_length, table_name=table_name, local_server=reverb_server) rb_observer = reverb_utils.ReverbTrajectorySequenceObserver( reverb_replay.py_client, table_name, sequence_length=sequence_length, stride_length=1) dataset = reverb_replay.as_dataset(num_parallel_calls=3, sample_batch_size=batch_size, num_steps=sequence_length).prefetch(3) experience_dataset_fn = lambda: dataset saved_model_dir = os.path.join(root_dir, learner.POLICY_SAVED_MODEL_DIR) env_step_metric = py_metrics.EnvironmentSteps() learning_triggers = [ triggers.PolicySavedModelTrigger( saved_model_dir, agent, train_step, interval=policy_save_interval, metadata_metrics={triggers.ENV_STEP_METADATA_KEY: env_step_metric}), triggers.StepPerSecondLogTrigger(train_step, interval=100), ] dqn_learner = learner.Learner(root_dir, train_step, agent, experience_dataset_fn, triggers=learning_triggers) # If we haven't trained yet make sure we collect some random samples first to # fill up the Replay Buffer with some experience. random_policy = random_py_policy.RandomPyPolicy( collect_env.time_step_spec(), collect_env.action_spec()) initial_collect_actor = actor.Actor(collect_env, random_policy, train_step, steps_per_run=initial_collect_steps, observers=[rb_observer]) logging.info('Doing initial collect.') initial_collect_actor.run() tf_collect_policy = agent.collect_policy collect_policy = py_tf_eager_policy.PyTFEagerPolicy(tf_collect_policy, use_tf_function=True) collect_actor = actor.Actor( collect_env, collect_policy, train_step, steps_per_run=collect_steps_per_iteration, observers=[rb_observer, env_step_metric], metrics=actor.collect_metrics(10), summary_dir=os.path.join(root_dir, learner.TRAIN_DIR), ) tf_greedy_policy = agent.policy greedy_policy = py_tf_eager_policy.PyTFEagerPolicy(tf_greedy_policy, use_tf_function=True) eval_actor = actor.Actor( eval_env, greedy_policy, train_step, episodes_per_run=eval_episodes, metrics=actor.eval_metrics(eval_episodes), summary_dir=os.path.join(root_dir, 'eval'), ) if eval_interval: logging.info('Evaluating.') eval_actor.run_and_log() logging.info('Training.') for _ in range(num_iterations): collect_actor.run() dqn_learner.run(iterations=1) if eval_interval and dqn_learner.train_step_numpy % eval_interval == 0: logging.info('Evaluating.') eval_actor.run_and_log() rb_observer.close() reverb_server.stop()
def train_eval( root_dir, env_name='Pong-v0', # Training params update_frequency=4, # Number of collect steps per policy update initial_collect_steps=50000, # 50k collect steps num_iterations=50000000, # 50M collect steps # Taken from Rainbow as it's not specified in Mnih,15. max_episode_frames_collect=50000, # env frames observed by the agent max_episode_frames_eval=108000, # env frames observed by the agent # Agent params epsilon_greedy=0.1, epsilon_decay_period=250000, # 1M collect steps / update_frequency batch_size=32, learning_rate=0.00025, n_step_update=1, gamma=0.99, target_update_tau=1.0, target_update_period=2500, # 10k collect steps / update_frequency reward_scale_factor=1.0, # Replay params reverb_port=None, replay_capacity=1000000, # Others policy_save_interval=250000, eval_interval=1000, eval_episodes=30, debug_summaries=True): """Trains and evaluates DQN.""" collect_env = suite_atari.load( env_name, max_episode_steps=max_episode_frames_collect, gym_env_wrappers=suite_atari.DEFAULT_ATARI_GYM_WRAPPERS_WITH_STACKING) eval_env = suite_atari.load( env_name, max_episode_steps=max_episode_frames_eval, gym_env_wrappers=suite_atari.DEFAULT_ATARI_GYM_WRAPPERS_WITH_STACKING) unused_observation_tensor_spec, action_tensor_spec, time_step_tensor_spec = ( spec_utils.get_tensor_specs(collect_env)) train_step = train_utils.create_train_step() num_actions = action_tensor_spec.maximum - action_tensor_spec.minimum + 1 epsilon = tf.compat.v1.train.polynomial_decay( 1.0, train_step, epsilon_decay_period, end_learning_rate=epsilon_greedy) agent = dqn_agent.DqnAgent( time_step_tensor_spec, action_tensor_spec, q_network=create_q_network(num_actions), epsilon_greedy=epsilon, n_step_update=n_step_update, target_update_tau=target_update_tau, target_update_period=target_update_period, optimizer=tf.compat.v1.train.RMSPropOptimizer( learning_rate=learning_rate, decay=0.95, momentum=0.95, epsilon=0.01, centered=True), td_errors_loss_fn=common.element_wise_huber_loss, gamma=gamma, reward_scale_factor=reward_scale_factor, train_step_counter=train_step, debug_summaries=debug_summaries) table_name = 'uniform_table' table = reverb.Table(table_name, max_size=replay_capacity, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), rate_limiter=reverb.rate_limiters.MinSize(1)) reverb_server = reverb.Server([table], port=reverb_port) reverb_replay = reverb_replay_buffer.ReverbReplayBuffer( agent.collect_data_spec, sequence_length=2, table_name=table_name, local_server=reverb_server) rb_observer = reverb_utils.ReverbAddTrajectoryObserver( reverb_replay.py_client, table_name, sequence_length=2, stride_length=1) dataset = reverb_replay.as_dataset(sample_batch_size=batch_size, num_steps=2).prefetch(3) experience_dataset_fn = lambda: dataset saved_model_dir = os.path.join(root_dir, learner.POLICY_SAVED_MODEL_DIR) env_step_metric = py_metrics.EnvironmentSteps() learning_triggers = [ triggers.PolicySavedModelTrigger( saved_model_dir, agent, train_step, interval=policy_save_interval, metadata_metrics={triggers.ENV_STEP_METADATA_KEY: env_step_metric}), triggers.StepPerSecondLogTrigger(train_step, interval=100), ] dqn_learner = learner.Learner(root_dir, train_step, agent, experience_dataset_fn, triggers=learning_triggers) # If we haven't trained yet make sure we collect some random samples first to # fill up the Replay Buffer with some experience. random_policy = random_py_policy.RandomPyPolicy( collect_env.time_step_spec(), collect_env.action_spec()) initial_collect_actor = actor.Actor(collect_env, random_policy, train_step, steps_per_run=initial_collect_steps, observers=[rb_observer]) logging.info('Doing initial collect.') initial_collect_actor.run() tf_collect_policy = agent.collect_policy collect_policy = py_tf_eager_policy.PyTFEagerPolicy(tf_collect_policy, use_tf_function=True) collect_actor = actor.Actor( collect_env, collect_policy, train_step, steps_per_run=update_frequency, observers=[rb_observer, env_step_metric], metrics=actor.collect_metrics(10), summary_dir=os.path.join(root_dir, learner.TRAIN_DIR), ) tf_greedy_policy = agent.policy greedy_policy = py_tf_eager_policy.PyTFEagerPolicy(tf_greedy_policy, use_tf_function=True) eval_actor = actor.Actor( eval_env, greedy_policy, train_step, episodes_per_run=eval_episodes, metrics=actor.eval_metrics(eval_episodes), summary_dir=os.path.join(root_dir, 'eval'), ) if eval_interval: logging.info('Evaluating.') eval_actor.run_and_log() logging.info('Training.') for _ in range(num_iterations): collect_actor.run() dqn_learner.run(iterations=1) if eval_interval and dqn_learner.train_step_numpy % eval_interval == 0: logging.info('Evaluating.') eval_actor.run_and_log() rb_observer.close() reverb_server.stop()