def collect_metrics(buffer_size):
"""Utilitiy to create metrics often used during data collection."""
metrics = [
py_metrics.NumberOfEpisodes(),
py_metrics.EnvironmentSteps(),
py_metrics.AverageReturnMetric(buffer_size=buffer_size),
py_metrics.AverageEpisodeLengthMetric(buffer_size=buffer_size),
]
return metrics
def collect(summary_dir: Text,
environment_name: Text,
collect_policy: py_tf_eager_policy.PyTFEagerPolicyBase,
replay_buffer_server_address: Text,
variable_container_server_address: Text,
suite_load_fn: Callable[
[Text], py_environment.PyEnvironment] = suite_mujoco.load,
initial_collect_steps: int = 10000,
max_train_steps: int = 2000000) -> None:
"""Collects experience using a policy updated after every episode."""
# Create the environment. For now support only single environment collection.
collect_env = suite_load_fn(environment_name)
# Create the variable container.
train_step = train_utils.create_train_step()
variables = {
reverb_variable_container.POLICY_KEY: 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.update(variables)
# Create the replay buffer observer.
rb_observer = reverb_utils.ReverbAddTrajectoryObserver(
reverb.Client(replay_buffer_server_address),
table_name=reverb_replay_buffer.DEFAULT_TABLE,
sequence_length=2,
stride_length=1)
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()
env_step_metric = py_metrics.EnvironmentSteps()
collect_actor = actor.Actor(
collect_env,
collect_policy,
train_step,
steps_per_run=1,
metrics=actor.collect_metrics(10),
summary_dir=summary_dir,
observers=[rb_observer, env_step_metric])
# Run the experience collection loop.
while train_step.numpy() < max_train_steps:
logging.info('Collecting with policy at step: %d', train_step.numpy())
collect_actor.run()
variable_container.update(variables)
def _create_collect_actor(
collect_env: YGOEnvironment, collect_policy: PyTFEagerPolicy,
train_step, rb_observer: ReverbAddTrajectoryObserver) -> actor.Actor:
initial_collect_actor = actor.Actor(
collect_env,
random_py_policy.RandomPyPolicy(collect_env.time_step_spec(),
collect_env.action_spec()),
train_step,
episodes_per_run=_initial_collect_episodes,
observers=[rb_observer])
initial_collect_actor.run()
return actor.Actor(collect_env,
collect_policy,
train_step,
episodes_per_run=1,
metrics=actor.collect_metrics(10),
summary_dir=os.path.join(tempdir, learner.TRAIN_DIR),
observers=[rb_observer,
py_metrics.EnvironmentSteps()])
def testRefereneMetricsNotInObservers(self):
rb_port = portpicker.pick_unused_port(portserver_address='localhost')
env, agent, train_step, _, rb_observer = (
self._build_components(rb_port))
temp_dir = self.create_tempdir().full_path
tf_collect_policy = agent.collect_policy
collect_policy = py_tf_eager_policy.PyTFEagerPolicy(
tf_collect_policy, use_tf_function=True)
metrics = metrics = actor.collect_metrics(buffer_size=1)
step_metric = py_metrics.EnvironmentSteps()
test_actor = actor.Actor(env,
collect_policy,
train_step,
steps_per_run=1,
metrics=metrics,
reference_metrics=[step_metric],
summary_dir=temp_dir,
observers=[rb_observer])
self.assertNotIn(step_metric, test_actor._observers)
def testSaveRestore(self):
metrics = [
py_metrics.AverageReturnMetric(),
py_metrics.AverageEpisodeLengthMetric(),
py_metrics.EnvironmentSteps(),
py_metrics.NumberOfEpisodes()
]
for metric in metrics:
metric(trajectory.boundary((), (), (), 0., 1.))
metric(trajectory.mid((), (), (), 1., 1.))
metric(trajectory.mid((), (), (), 2., 1.))
metric(trajectory.last((), (), (), 3., 0.))
checkpoint = tf.train.Checkpoint(**{m.name: m for m in metrics})
prefix = self.get_temp_dir() + '/ckpt'
save_path = checkpoint.save(prefix)
for metric in metrics:
metric.reset()
self.assertEqual(0, metric.result())
checkpoint.restore(save_path).assert_consumed()
for metric in metrics:
self.assertGreater(metric.result(), 0)
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()
def __init__(
self,
root_dir,
env_name,
num_iterations=200,
max_episode_frames=108000, # ALE frames
terminal_on_life_loss=False,
conv_layer_params=((32, (8, 8), 4), (64, (4, 4), 2), (64, (3, 3),
1)),
fc_layer_params=(512, ),
# Params for collect
initial_collect_steps=80000, # ALE frames
epsilon_greedy=0.01,
epsilon_decay_period=1000000, # ALE frames
replay_buffer_capacity=1000000,
# Params for train
train_steps_per_iteration=1000000, # ALE frames
update_period=16, # ALE frames
target_update_tau=1.0,
target_update_period=32000, # ALE frames
batch_size=32,
learning_rate=2.5e-4,
n_step_update=2,
gamma=0.99,
reward_scale_factor=1.0,
gradient_clipping=None,
# Params for eval
do_eval=True,
eval_steps_per_iteration=500000, # ALE frames
eval_epsilon_greedy=0.001,
# Params for checkpoints, summaries, and logging
log_interval=1000,
summary_interval=1000,
summaries_flush_secs=10,
debug_summaries=True,
summarize_grads_and_vars=True,
eval_metrics_callback=None):
"""A simple Atari train and eval for DQN.
Args:
root_dir: Directory to write log files to.
env_name: Fully-qualified name of the Atari environment (i.e. Pong-v0).
num_iterations: Number of train/eval iterations to run.
max_episode_frames: Maximum length of a single episode, in ALE frames.
terminal_on_life_loss: Whether to simulate an episode termination when a
life is lost.
conv_layer_params: Params for convolutional layers of QNetwork.
fc_layer_params: Params for fully connected layers of QNetwork.
initial_collect_steps: Number of frames to ALE frames to process before
beginning to train. Since this is in ALE frames, there will be
initial_collect_steps/4 items in the replay buffer when training starts.
epsilon_greedy: Final epsilon value to decay to for training.
epsilon_decay_period: Period over which to decay epsilon, from 1.0 to
epsilon_greedy (defined above).
replay_buffer_capacity: Maximum number of items to store in the replay
buffer.
train_steps_per_iteration: Number of ALE frames to run through for each
iteration of training.
update_period: Run a train operation every update_period ALE frames.
target_update_tau: Coeffecient for soft target network updates (1.0 ==
hard updates).
target_update_period: Period, in ALE frames, to copy the live network to
the target network.
batch_size: Number of frames to include in each training batch.
learning_rate: RMS optimizer learning rate.
n_step_update: The number of steps to consider when computing TD error and
TD loss. Applies standard single-step updates when set to 1.
gamma: Discount for future rewards.
reward_scale_factor: Scaling factor for rewards.
gradient_clipping: Norm length to clip gradients.
do_eval: If True, run an eval every iteration. If False, skip eval.
eval_steps_per_iteration: Number of ALE frames to run through for each
iteration of evaluation.
eval_epsilon_greedy: Epsilon value to use for the evaluation policy (0 ==
totally greedy policy).
log_interval: Log stats to the terminal every log_interval training
steps.
summary_interval: Write TF summaries every summary_interval training
steps.
summaries_flush_secs: Flush summaries to disk every summaries_flush_secs
seconds.
debug_summaries: If True, write additional summaries for debugging (see
dqn_agent for which summaries are written).
summarize_grads_and_vars: Include gradients in summaries.
eval_metrics_callback: A callback function that takes (metric_dict,
global_step) as parameters. Called after every eval with the results of
the evaluation.
"""
self._update_period = update_period / ATARI_FRAME_SKIP
self._train_steps_per_iteration = (train_steps_per_iteration /
ATARI_FRAME_SKIP)
self._do_eval = do_eval
self._eval_steps_per_iteration = eval_steps_per_iteration / ATARI_FRAME_SKIP
self._eval_epsilon_greedy = eval_epsilon_greedy
self._initial_collect_steps = initial_collect_steps / ATARI_FRAME_SKIP
self._summary_interval = summary_interval
self._num_iterations = num_iterations
self._log_interval = log_interval
self._eval_metrics_callback = eval_metrics_callback
with gin.unlock_config():
gin.bind_parameter(('tf_agents.environments.atari_preprocessing.'
'AtariPreprocessing.terminal_on_life_loss'),
terminal_on_life_loss)
root_dir = os.path.expanduser(root_dir)
train_dir = os.path.join(root_dir, 'train')
eval_dir = os.path.join(root_dir, 'eval')
train_summary_writer = tf.compat.v2.summary.create_file_writer(
train_dir, flush_millis=summaries_flush_secs * 1000)
train_summary_writer.set_as_default()
self._train_summary_writer = train_summary_writer
self._eval_summary_writer = None
if self._do_eval:
self._eval_summary_writer = tf.compat.v2.summary.create_file_writer(
eval_dir, flush_millis=summaries_flush_secs * 1000)
self._eval_metrics = [
py_metrics.AverageReturnMetric(name='PhaseAverageReturn',
buffer_size=np.inf),
py_metrics.AverageEpisodeLengthMetric(
name='PhaseAverageEpisodeLength', buffer_size=np.inf),
]
self._global_step = tf.compat.v1.train.get_or_create_global_step()
with tf.compat.v2.summary.record_if(lambda: tf.math.equal(
self._global_step % self._summary_interval, 0)):
self._env = suite_atari.load(
env_name,
max_episode_steps=max_episode_frames / ATARI_FRAME_SKIP,
gym_env_wrappers=suite_atari.
DEFAULT_ATARI_GYM_WRAPPERS_WITH_STACKING)
self._env = batched_py_environment.BatchedPyEnvironment(
[self._env])
observation_spec = tensor_spec.from_spec(
self._env.observation_spec())
time_step_spec = ts.time_step_spec(observation_spec)
action_spec = tensor_spec.from_spec(self._env.action_spec())
with tf.device('/cpu:0'):
epsilon = tf.compat.v1.train.polynomial_decay(
1.0,
self._global_step,
epsilon_decay_period / ATARI_FRAME_SKIP /
self._update_period,
end_learning_rate=epsilon_greedy)
with tf.device('/gpu:0'):
optimizer = tf.compat.v1.train.RMSPropOptimizer(
learning_rate=learning_rate,
decay=0.95,
momentum=0.0,
epsilon=0.00001,
centered=True)
categorical_q_net = AtariCategoricalQNetwork(
observation_spec,
action_spec,
conv_layer_params=conv_layer_params,
fc_layer_params=fc_layer_params)
agent = categorical_dqn_agent.CategoricalDqnAgent(
time_step_spec,
action_spec,
categorical_q_network=categorical_q_net,
optimizer=optimizer,
epsilon_greedy=epsilon,
n_step_update=n_step_update,
target_update_tau=target_update_tau,
target_update_period=(target_update_period /
ATARI_FRAME_SKIP /
self._update_period),
gamma=gamma,
reward_scale_factor=reward_scale_factor,
gradient_clipping=gradient_clipping,
debug_summaries=debug_summaries,
summarize_grads_and_vars=summarize_grads_and_vars,
train_step_counter=self._global_step)
self._collect_policy = py_tf_policy.PyTFPolicy(
agent.collect_policy)
if self._do_eval:
self._eval_policy = py_tf_policy.PyTFPolicy(
epsilon_greedy_policy.EpsilonGreedyPolicy(
policy=agent.policy,
epsilon=self._eval_epsilon_greedy))
py_observation_spec = self._env.observation_spec()
py_time_step_spec = ts.time_step_spec(py_observation_spec)
py_action_spec = policy_step.PolicyStep(
self._env.action_spec())
data_spec = trajectory.from_transition(py_time_step_spec,
py_action_spec,
py_time_step_spec)
self._replay_buffer = py_hashed_replay_buffer.PyHashedReplayBuffer(
data_spec=data_spec, capacity=replay_buffer_capacity)
with tf.device('/cpu:0'):
ds = self._replay_buffer.as_dataset(
sample_batch_size=batch_size, num_steps=n_step_update + 1)
ds = ds.prefetch(4)
ds = ds.apply(
tf.data.experimental.prefetch_to_device('/gpu:0'))
with tf.device('/gpu:0'):
self._ds_itr = tf.compat.v1.data.make_one_shot_iterator(ds)
experience = self._ds_itr.get_next()
self._train_op = agent.train(experience)
self._env_steps_metric = py_metrics.EnvironmentSteps()
self._step_metrics = [
py_metrics.NumberOfEpisodes(),
self._env_steps_metric,
]
self._train_metrics = self._step_metrics + [
py_metrics.AverageReturnMetric(buffer_size=10),
py_metrics.AverageEpisodeLengthMetric(buffer_size=10),
]
# The _train_phase_metrics average over an entire train iteration,
# rather than the rolling average of the last 10 episodes.
self._train_phase_metrics = [
py_metrics.AverageReturnMetric(name='PhaseAverageReturn',
buffer_size=np.inf),
py_metrics.AverageEpisodeLengthMetric(
name='PhaseAverageEpisodeLength', buffer_size=np.inf),
]
self._iteration_metric = py_metrics.CounterMetric(
name='Iteration')
# Summaries written from python should run every time they are
# generated.
with tf.compat.v2.summary.record_if(True):
self._steps_per_second_ph = tf.compat.v1.placeholder(
tf.float32, shape=(), name='steps_per_sec_ph')
self._steps_per_second_summary = tf.compat.v2.summary.scalar(
name='global_steps_per_sec',
data=self._steps_per_second_ph,
step=self._global_step)
for metric in self._train_metrics:
metric.tf_summaries(train_step=self._global_step,
step_metrics=self._step_metrics)
for metric in self._train_phase_metrics:
metric.tf_summaries(
train_step=self._global_step,
step_metrics=(self._iteration_metric, ))
self._iteration_metric.tf_summaries(
train_step=self._global_step)
if self._do_eval:
with self._eval_summary_writer.as_default():
for metric in self._eval_metrics:
metric.tf_summaries(
train_step=self._global_step,
step_metrics=(self._iteration_metric, ))
self._train_checkpointer = common.Checkpointer(
ckpt_dir=train_dir,
agent=agent,
global_step=self._global_step,
optimizer=optimizer,
metrics=metric_utils.MetricsGroup(
self._train_metrics + self._train_phase_metrics +
[self._iteration_metric], 'train_metrics'))
self._policy_checkpointer = common.Checkpointer(
ckpt_dir=os.path.join(train_dir, 'policy'),
policy=agent.policy,
global_step=self._global_step)
self._rb_checkpointer = common.Checkpointer(
ckpt_dir=os.path.join(train_dir, 'replay_buffer'),
max_to_keep=1,
replay_buffer=self._replay_buffer)
self._init_agent_op = agent.initialize()
def train_eval(
root_dir,
env_name='HalfCheetah-v2',
# Training params
num_iterations=1600,
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=0.5,
value_clipping=None,
# Replay params
reverb_port=None,
replay_capacity=10000,
# Others
policy_save_interval=5000,
summary_interval=1000,
eval_interval=10000,
eval_episodes=100,
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))
# TODO(b/172267869): Remove this conversion once TensorNormalizer stops
# converting float64 inputs to float32.
observation_tensor_spec = tf.TensorSpec(
dtype=tf.float32, shape=observation_tensor_spec.shape)
train_step = train_utils.create_train_step()
actor_net_builder = ppo_actor_network.PPOActorNetwork()
actor_net = actor_net_builder.create_sequential_actor_net(
actor_fc_layers, action_tensor_spec)
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.keras.optimizers.Adam(
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_samples=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)
eval_greedy_policy = py_tf_eager_policy.PyTFEagerPolicy(
agent.policy, use_tf_function=True)
if eval_interval:
logging.info('Intial evaluation.')
eval_actor = actor.Actor(
eval_env,
eval_greedy_policy,
train_step,
metrics=actor.eval_metrics(eval_episodes),
reference_metrics=[collect_env_step_metric],
summary_dir=os.path.join(root_dir, 'eval'),
episodes_per_run=eval_episodes)
eval_actor.run_and_log()
logging.info('Training on %s', env_name)
last_eval_step = 0
for i in range(num_iterations):
collect_actor.run()
rb_observer.flush()
agent_learner.run()
reverb_replay_train.clear()
reverb_replay_normalization.clear()
current_iteration.assign_add(1)
# Eval only if `eval_interval` has been set. Then, eval if the current train
# step is equal or greater than the `last_eval_step` + `eval_interval` or if
# this is the last iteration. This logic exists because agent_learner.run()
# does not return after every train step.
if (eval_interval and
(agent_learner.train_step_numpy >= eval_interval + last_eval_step
or i == num_iterations - 1)):
logging.info('Evaluating.')
eval_actor.run_and_log()
last_eval_step = agent_learner.train_step_numpy
rb_observer.close()
reverb_server.stop()
def train_eval(
root_dir,
env_name='CartPole-v0',
# Training params
initial_collect_steps=1000,
num_iterations=100000,
fc_layer_params=(100, ),
# Agent params
epsilon_greedy=0.1,
batch_size=64,
learning_rate=1e-3,
n_step_update=1,
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)
time_step_tensor_spec = tensor_spec.from_spec(collect_env.time_step_spec())
action_tensor_spec = tensor_spec.from_spec(collect_env.action_spec())
train_step = train_utils.create_train_step()
num_actions = action_tensor_spec.maximum - action_tensor_spec.minimum + 1
# Define a helper function to create Dense layers configured with the right
# activation and kernel initializer.
def dense_layer(num_units):
return tf.keras.layers.Dense(
num_units,
activation=tf.keras.activations.relu,
kernel_initializer=tf.keras.initializers.VarianceScaling(
scale=2.0, mode='fan_in', distribution='truncated_normal'))
# QNetwork consists of a sequence of Dense layers followed by a dense layer
# with `num_actions` units to generate one q_value per available action as
# it's output.
dense_layers = [dense_layer(num_units) for num_units in fc_layer_params]
q_values_layer = tf.keras.layers.Dense(
num_actions,
activation=None,
kernel_initializer=tf.keras.initializers.RandomUniform(minval=-0.03,
maxval=0.03),
bias_initializer=tf.keras.initializers.Constant(-0.2))
q_net = sequential.Sequential(dense_layers + [q_values_layer])
agent = dqn_agent.DqnAgent(
time_step_tensor_spec,
action_tensor_spec,
q_network=q_net,
epsilon_greedy=epsilon_greedy,
n_step_update=n_step_update,
target_update_tau=target_update_tau,
target_update_period=target_update_period,
optimizer=tf.keras.optimizers.Adam(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=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(num_parallel_calls=3,
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=1,
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,
# 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.keras.optimizers.Adam(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.ReverbAddTrajectoryObserver(
reverb_replay.py_client,
table_name,
sequence_length=sequence_length,
stride_length=1,
pad_end_of_episodes=True)
def experience_dataset_fn():
return reverb_replay.as_dataset(sample_batch_size=batch_size,
num_steps=sequence_length)
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),
reference_metrics=[env_step_metric],
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),
reference_metrics=[env_step_metric],
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()
print(f" -- ACTORS ({now()}) -- ")
rb_observer = reverb_utils.ReverbAddTrajectoryObserver(reverb_replay.py_client,
table_name,
sequence_length=2,
stride_length=1)
initial_collect_actor = actor.Actor(
collect_env,
random_policy,
train_step,
steps_per_run=HyperParms.initial_collect_steps,
observers=[rb_observer])
initial_collect_actor.run()
env_step_metric = py_metrics.EnvironmentSteps()
collect_actor = actor.Actor(collect_env,
collect_policy,
train_step,
steps_per_run=1,
metrics=actor.collect_metrics(10),
summary_dir=os.path.join(tempdir,
learner.TRAIN_DIR),
observers=[rb_observer, env_step_metric])
eval_actor = actor.Actor(
eval_env,
eval_policy,
train_step,
episodes_per_run=HyperParms.num_eval_episodes,
metrics=actor.eval_metrics(HyperParms.num_eval_episodes),
def collect(task,
root_dir,
replay_buffer_server_address,
variable_container_server_address,
create_env_fn,
initial_collect_steps=10000,
num_iterations=10000000):
"""Collects experience using a policy updated after every episode."""
# Create the environment. For now support only single environment collection.
collect_env = create_env_fn()
# Create the path for the serialized collect policy.
collect_policy_saved_model_path = os.path.join(
root_dir, learner.POLICY_SAVED_MODEL_DIR,
learner.COLLECT_POLICY_SAVED_MODEL_DIR)
saved_model_pb_path = os.path.join(collect_policy_saved_model_path,
'saved_model.pb')
try:
# Wait for the collect policy to be outputed by learner (timeout after 2
# days), then load it.
train_utils.wait_for_file(
saved_model_pb_path, sleep_time_secs=2, num_retries=86400)
collect_policy = py_tf_eager_policy.SavedModelPyTFEagerPolicy(
collect_policy_saved_model_path, load_specs_from_pbtxt=True)
except TimeoutError as e:
# If the collect policy does not become available during the wait time of
# the call `wait_for_file`, that probably means the learner is not running.
logging.error('Could not get the file %s. Exiting.', saved_model_pb_path)
raise e
# Create the variable container.
train_step = train_utils.create_train_step()
variables = {
reverb_variable_container.POLICY_KEY: 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.update(variables)
# Create the replay buffer observer.
rb_observer = reverb_utils.ReverbAddTrajectoryObserver(
reverb.Client(replay_buffer_server_address),
table_name=reverb_replay_buffer.DEFAULT_TABLE,
sequence_length=2,
stride_length=1)
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()
env_step_metric = py_metrics.EnvironmentSteps()
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, str(task)),
observers=[rb_observer, env_step_metric])
# Run the experience collection loop.
for _ in range(num_iterations):
logging.info('Collecting with policy at step: %d', train_step.numpy())
collect_actor.run()
variable_container.update(variables)
def train_eval(
root_dir,
dataset_path,
env_name,
# Training params
tpu=False,
use_gpu=False,
num_gradient_updates=1000000,
actor_fc_layers=(256, 256),
critic_joint_fc_layers=(256, 256, 256),
# Agent params
batch_size=256,
bc_steps=0,
actor_learning_rate=3e-5,
critic_learning_rate=3e-4,
alpha_learning_rate=3e-4,
reward_scale_factor=1.0,
cql_alpha_learning_rate=3e-4,
cql_alpha=5.0,
cql_tau=10.0,
num_cql_samples=10,
reward_noise_variance=0.0,
include_critic_entropy_term=False,
use_lagrange_cql_alpha=True,
log_cql_alpha_clipping=None,
softmax_temperature=1.0,
# Data params
reward_shift=0.0,
action_clipping=None,
use_trajectories=False,
data_shuffle_buffer_size_per_record=1,
data_shuffle_buffer_size=100,
data_num_shards=1,
data_block_length=10,
data_parallel_reads=None,
data_parallel_calls=10,
data_prefetch=10,
data_cycle_length=10,
# Others
policy_save_interval=10000,
eval_interval=10000,
summary_interval=1000,
learner_iterations_per_call=1,
eval_episodes=10,
debug_summaries=False,
summarize_grads_and_vars=False,
seed=None):
"""Trains and evaluates CQL-SAC."""
logging.info('Training CQL-SAC on: %s', env_name)
tf.random.set_seed(seed)
np.random.seed(seed)
# Load environment.
env = load_d4rl(env_name)
tf_env = tf_py_environment.TFPyEnvironment(env)
strategy = strategy_utils.get_strategy(tpu, use_gpu)
if not dataset_path.endswith('.tfrecord'):
dataset_path = os.path.join(dataset_path, env_name,
'%s*.tfrecord' % env_name)
logging.info('Loading dataset from %s', dataset_path)
dataset_paths = tf.io.gfile.glob(dataset_path)
# Create dataset.
with strategy.scope():
dataset = create_tf_record_dataset(
dataset_paths,
batch_size,
shuffle_buffer_size_per_record=data_shuffle_buffer_size_per_record,
shuffle_buffer_size=data_shuffle_buffer_size,
num_shards=data_num_shards,
cycle_length=data_cycle_length,
block_length=data_block_length,
num_parallel_reads=data_parallel_reads,
num_parallel_calls=data_parallel_calls,
num_prefetch=data_prefetch,
strategy=strategy,
reward_shift=reward_shift,
action_clipping=action_clipping,
use_trajectories=use_trajectories)
# Create agent.
time_step_spec = tf_env.time_step_spec()
observation_spec = time_step_spec.observation
action_spec = tf_env.action_spec()
with strategy.scope():
train_step = train_utils.create_train_step()
actor_net = actor_distribution_network.ActorDistributionNetwork(
observation_spec,
action_spec,
fc_layer_params=actor_fc_layers,
continuous_projection_net=tanh_normal_projection_network.
TanhNormalProjectionNetwork)
critic_net = critic_network.CriticNetwork(
(observation_spec, action_spec),
joint_fc_layer_params=critic_joint_fc_layers,
kernel_initializer='glorot_uniform',
last_kernel_initializer='glorot_uniform')
agent = cql_sac_agent.CqlSacAgent(
time_step_spec,
action_spec,
actor_network=actor_net,
critic_network=critic_net,
actor_optimizer=tf.keras.optimizers.Adam(
learning_rate=actor_learning_rate),
critic_optimizer=tf.keras.optimizers.Adam(
learning_rate=critic_learning_rate),
alpha_optimizer=tf.keras.optimizers.Adam(
learning_rate=alpha_learning_rate),
cql_alpha=cql_alpha,
num_cql_samples=num_cql_samples,
include_critic_entropy_term=include_critic_entropy_term,
use_lagrange_cql_alpha=use_lagrange_cql_alpha,
cql_alpha_learning_rate=cql_alpha_learning_rate,
target_update_tau=5e-3,
target_update_period=1,
random_seed=seed,
cql_tau=cql_tau,
reward_noise_variance=reward_noise_variance,
num_bc_steps=bc_steps,
td_errors_loss_fn=tf.math.squared_difference,
gamma=0.99,
reward_scale_factor=reward_scale_factor,
gradient_clipping=None,
log_cql_alpha_clipping=log_cql_alpha_clipping,
softmax_temperature=softmax_temperature,
debug_summaries=debug_summaries,
summarize_grads_and_vars=summarize_grads_and_vars,
train_step_counter=train_step)
agent.initialize()
# Create learner.
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=100)
]
cql_learner = learner.Learner(root_dir,
train_step,
agent,
experience_dataset_fn=lambda: dataset,
triggers=learning_triggers,
summary_interval=summary_interval,
strategy=strategy)
# Create actor for evaluation.
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(env,
eval_greedy_policy,
train_step,
metrics=actor.eval_metrics(eval_episodes),
summary_dir=os.path.join(root_dir, 'eval'),
episodes_per_run=eval_episodes)
# Run.
dummy_trajectory = trajectory.mid((), (), (), 0., 1.)
num_learner_iterations = int(num_gradient_updates /
learner_iterations_per_call)
for _ in range(num_learner_iterations):
# Mimic collecting environment steps since we loaded a static dataset.
for _ in range(learner_iterations_per_call):
collect_env_step_metric(dummy_trajectory)
cql_learner.run(iterations=learner_iterations_per_call)
if eval_interval and train_step.numpy() % eval_interval == 0:
eval_actor.run_and_log()
def train_eval(
root_dir,
strategy: tf.distribute.Strategy,
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
policy_save_interval=10000,
replay_buffer_save_interval=100000,
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)
_, action_tensor_spec, time_step_tensor_spec = (
spec_utils.get_tensor_specs(collect_env))
actor_net = create_sequential_actor_network(
actor_fc_layers=actor_fc_layers, action_tensor_spec=action_tensor_spec)
critic_net = create_sequential_critic_network(
obs_fc_layer_units=critic_obs_fc_layers,
action_fc_layer_units=critic_action_fc_layers,
joint_fc_layer_units=critic_joint_fc_layers)
with strategy.scope():
train_step = train_utils.create_train_step()
agent = sac_agent.SacAgent(
time_step_tensor_spec,
action_tensor_spec,
actor_network=actor_net,
critic_network=critic_net,
actor_optimizer=tf.keras.optimizers.Adam(
learning_rate=actor_learning_rate),
critic_optimizer=tf.keras.optimizers.Adam(
learning_rate=critic_learning_rate),
alpha_optimizer=tf.keras.optimizers.Adam(
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_checkpoint_dir = os.path.join(root_dir, learner.TRAIN_DIR,
learner.REPLAY_BUFFER_CHECKPOINT_DIR)
reverb_checkpointer = reverb.platform.checkpointers_lib.DefaultCheckpointer(
path=reverb_checkpoint_dir)
reverb_server = reverb.Server([table],
port=reverb_port,
checkpointer=reverb_checkpointer)
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)
def experience_dataset_fn():
return reverb_replay.as_dataset(sample_batch_size=batch_size,
num_steps=2).prefetch(50)
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.ReverbCheckpointTrigger(
train_step,
interval=replay_buffer_save_interval,
reverb_client=reverb_replay.py_client),
# TODO(b/165023684): Add SIGTERM handler to checkpoint before preemption.
triggers.StepPerSecondLogTrigger(train_step, interval=1000),
]
agent_learner = learner.Learner(root_dir,
train_step,
agent,
experience_dataset_fn,
triggers=learning_triggers,
strategy=strategy)
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()
