def train_eval(
root_dir,
env_name='CartPole-v0',
num_iterations=100000,
train_sequence_length=1,
# Params for QNetwork
fc_layer_params=(100, ),
# Params for QRnnNetwork
input_fc_layer_params=(50, ),
lstm_size=(20, ),
output_fc_layer_params=(20, ),
# Params for collect
initial_collect_steps=1000,
collect_steps_per_iteration=1,
epsilon_greedy=0.1,
replay_buffer_capacity=100000,
# Params for target update
target_update_tau=0.05,
target_update_period=5,
# Params for train
train_steps_per_iteration=1,
batch_size=64,
learning_rate=1e-3,
gamma=0.99,
reward_scale_factor=1.0,
gradient_clipping=None,
use_tf_functions=True,
# Params for eval
num_eval_episodes=10,
eval_interval=1000,
# Params for checkpoints
train_checkpoint_interval=10000,
policy_checkpoint_interval=5000,
rb_checkpoint_interval=20000,
# Params for summaries and logging
log_interval=1000,
summary_interval=1000,
summaries_flush_secs=10,
debug_summaries=False,
summarize_grads_and_vars=False,
eval_metrics_callback=None):
"""A simple train and eval for DQN."""
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()
eval_summary_writer = tf.compat.v2.summary.create_file_writer(
eval_dir, flush_millis=summaries_flush_secs * 1000)
eval_metrics = [
tf_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
tf_metrics.AverageEpisodeLengthMetric(buffer_size=num_eval_episodes)
]
global_step = tf.compat.v1.train.get_or_create_global_step()
with tf.compat.v2.summary.record_if(
lambda: tf.math.equal(global_step % summary_interval, 0)):
tf_env = tf_py_environment.TFPyEnvironment(suite_gym.load(env_name))
eval_tf_env = tf_py_environment.TFPyEnvironment(
suite_gym.load(env_name))
if train_sequence_length > 1:
q_net = q_rnn_network.QRnnNetwork(
tf_env.observation_spec(),
tf_env.action_spec(),
input_fc_layer_params=input_fc_layer_params,
lstm_size=lstm_size,
output_fc_layer_params=output_fc_layer_params)
else:
q_net = q_network.QNetwork(tf_env.observation_spec(),
tf_env.action_spec(),
fc_layer_params=fc_layer_params)
# TODO(b/127301657): Decay epsilon based on global step, cf. cl/188907839
tf_agent = dqn_agent.DqnAgent(
tf_env.time_step_spec(),
tf_env.action_spec(),
q_network=q_net,
epsilon_greedy=epsilon_greedy,
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=dqn_agent.element_wise_squared_loss,
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=global_step)
tf_agent.initialize()
train_metrics = [
tf_metrics.NumberOfEpisodes(),
tf_metrics.EnvironmentSteps(),
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
eval_policy = tf_agent.policy
collect_policy = tf_agent.collect_policy
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
data_spec=tf_agent.collect_data_spec,
batch_size=tf_env.batch_size,
max_length=replay_buffer_capacity)
collect_driver = dynamic_step_driver.DynamicStepDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_steps=collect_steps_per_iteration)
train_checkpointer = common.Checkpointer(
ckpt_dir=train_dir,
agent=tf_agent,
global_step=global_step,
metrics=metric_utils.MetricsGroup(train_metrics, 'train_metrics'))
policy_checkpointer = common.Checkpointer(ckpt_dir=os.path.join(
train_dir, 'policy'),
policy=eval_policy,
global_step=global_step)
rb_checkpointer = common.Checkpointer(ckpt_dir=os.path.join(
train_dir, 'replay_buffer'),
max_to_keep=1,
replay_buffer=replay_buffer)
train_checkpointer.initialize_or_restore()
rb_checkpointer.initialize_or_restore()
if use_tf_functions:
# To speed up collect use common.function.
collect_driver.run = common.function(collect_driver.run)
tf_agent.train = common.function(tf_agent.train)
initial_collect_policy = random_tf_policy.RandomTFPolicy(
tf_env.time_step_spec(), tf_env.action_spec())
# Collect initial replay data.
logging.info(
'Initializing replay buffer by collecting experience for %d steps with '
'a random policy.', initial_collect_steps)
dynamic_step_driver.DynamicStepDriver(
tf_env,
initial_collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_steps=initial_collect_steps).run()
results = metric_utils.eager_compute(
eval_metrics,
eval_tf_env,
eval_policy,
num_episodes=num_eval_episodes,
summary_writer=eval_summary_writer,
summary_prefix='Metrics',
)
if eval_metrics_callback is not None:
eval_metrics_callback(results, global_step.numpy())
metric_utils.log_metrics(eval_metrics)
time_step = None
policy_state = collect_policy.get_initial_state(tf_env.batch_size)
timed_at_step = global_step.numpy()
time_acc = 0
# Dataset generates trajectories with shape [Bx2x...]
dataset = replay_buffer.as_dataset(num_parallel_calls=3,
sample_batch_size=batch_size,
num_steps=train_sequence_length +
1).prefetch(3)
iterator = iter(dataset)
for _ in range(num_iterations):
start_time = time.time()
time_step, policy_state = collect_driver.run(
time_step=time_step,
policy_state=policy_state,
)
for _ in range(train_steps_per_iteration):
experience, _ = next(iterator)
train_loss = tf_agent.train(experience)
time_acc += time.time() - start_time
if global_step.numpy() % log_interval == 0:
logging.info('step = %d, loss = %f', global_step.numpy(),
train_loss.loss)
steps_per_sec = (global_step.numpy() -
timed_at_step) / time_acc
logging.info('%.3f steps/sec', steps_per_sec)
tf.contrib.summary.scalar(name='global_steps/sec',
tensor=steps_per_sec)
timed_at_step = global_step.numpy()
time_acc = 0
for train_metric in train_metrics:
train_metric.tf_summaries(step_metrics=train_metrics[:2])
if global_step.numpy() % train_checkpoint_interval == 0:
train_checkpointer.save(global_step=global_step.numpy())
if global_step.numpy() % policy_checkpoint_interval == 0:
policy_checkpointer.save(global_step=global_step.numpy())
if global_step.numpy() % rb_checkpoint_interval == 0:
rb_checkpointer.save(global_step=global_step.numpy())
if global_step.numpy() % eval_interval == 0:
results = metric_utils.eager_compute(
eval_metrics,
eval_tf_env,
eval_policy,
num_episodes=num_eval_episodes,
summary_writer=eval_summary_writer,
summary_prefix='Metrics',
)
if eval_metrics_callback is not None:
eval_metrics_callback(results, global_step.numpy())
metric_utils.log_metrics(eval_metrics)
return train_loss
def train_eval(
root_dir,
env_name='HalfCheetah-v2',
num_iterations=2000000,
actor_fc_layers=(400, 300),
critic_obs_fc_layers=(400, ),
critic_action_fc_layers=None,
critic_joint_fc_layers=(300, ),
# Params for collect
initial_collect_steps=1000,
collect_steps_per_iteration=1,
replay_buffer_capacity=100000,
ou_stddev=0.2,
ou_damping=0.15,
# Params for target update
target_update_tau=0.05,
target_update_period=5,
# Params for train
train_steps_per_iteration=1,
batch_size=64,
actor_learning_rate=1e-4,
critic_learning_rate=1e-3,
dqda_clipping=None,
td_errors_loss_fn=tf.compat.v1.losses.huber_loss,
gamma=0.995,
reward_scale_factor=1.0,
gradient_clipping=None,
use_tf_functions=True,
# Params for eval
num_eval_episodes=10,
eval_interval=10000,
# Params for checkpoints, summaries, and logging
log_interval=1000,
summary_interval=1000,
summaries_flush_secs=10,
debug_summaries=False,
summarize_grads_and_vars=False,
eval_metrics_callback=None):
"""A simple train and eval for TD3."""
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()
eval_summary_writer = tf.compat.v2.summary.create_file_writer(
eval_dir, flush_millis=summaries_flush_secs * 1000)
eval_metrics = [
tf_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
tf_metrics.AverageEpisodeLengthMetric(buffer_size=num_eval_episodes)
]
global_step = tf.compat.v1.train.get_or_create_global_step()
with tf.compat.v2.summary.record_if(
lambda: tf.math.equal(global_step % summary_interval, 0)):
tf_env = tf_py_environment.TFPyEnvironment(suite_mujoco.load(env_name))
eval_tf_env = tf_py_environment.TFPyEnvironment(
suite_mujoco.load(env_name))
actor_net = actor_network.ActorNetwork(
tf_env.time_step_spec().observation,
tf_env.action_spec(),
fc_layer_params=actor_fc_layers,
)
critic_net_input_specs = (tf_env.time_step_spec().observation,
tf_env.action_spec())
critic_net = critic_network.CriticNetwork(
critic_net_input_specs,
observation_fc_layer_params=critic_obs_fc_layers,
action_fc_layer_params=critic_action_fc_layers,
joint_fc_layer_params=critic_joint_fc_layers,
)
tf_agent = td3_agent.Td3Agent(
tf_env.time_step_spec(),
tf_env.action_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),
ou_stddev=ou_stddev,
ou_damping=ou_damping,
target_update_tau=target_update_tau,
target_update_period=target_update_period,
dqda_clipping=dqda_clipping,
td_errors_loss_fn=td_errors_loss_fn,
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=global_step,
)
tf_agent.initialize()
train_metrics = [
tf_metrics.NumberOfEpisodes(),
tf_metrics.EnvironmentSteps(),
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
eval_policy = tf_agent.policy
collect_policy = tf_agent.collect_policy
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
tf_agent.collect_data_spec,
batch_size=tf_env.batch_size,
max_length=replay_buffer_capacity)
initial_collect_driver = dynamic_step_driver.DynamicStepDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch],
num_steps=initial_collect_steps)
collect_driver = dynamic_step_driver.DynamicStepDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_steps=collect_steps_per_iteration)
if use_tf_functions:
initial_collect_driver.run = common.function(
initial_collect_driver.run)
collect_driver.run = common.function(collect_driver.run)
tf_agent.train = common.function(tf_agent.train)
# Collect initial replay data.
logging.info(
'Initializing replay buffer by collecting experience for %d steps with '
'a random policy.', initial_collect_steps)
initial_collect_driver.run()
results = metric_utils.eager_compute(
eval_metrics,
eval_tf_env,
eval_policy,
num_episodes=num_eval_episodes,
train_step=global_step,
summary_writer=eval_summary_writer,
summary_prefix='Metrics',
)
if eval_metrics_callback is not None:
eval_metrics_callback(results, global_step.numpy())
metric_utils.log_metrics(eval_metrics)
time_step = None
policy_state = collect_policy.get_initial_state(tf_env.batch_size)
timed_at_step = global_step.numpy()
time_acc = 0
# Dataset generates trajectories with shape [Bx2x...]
dataset = replay_buffer.as_dataset(num_parallel_calls=3,
sample_batch_size=batch_size,
num_steps=2).prefetch(3)
iterator = iter(dataset)
for _ in range(num_iterations):
start_time = time.time()
time_step, policy_state = collect_driver.run(
time_step=time_step,
policy_state=policy_state,
)
for _ in range(train_steps_per_iteration):
experience, _ = next(iterator)
train_loss = tf_agent.train(experience)
time_acc += time.time() - start_time
if global_step.numpy() % log_interval == 0:
logging.info('step = %d, loss = %f', global_step.numpy(),
train_loss.loss)
steps_per_sec = (global_step.numpy() -
timed_at_step) / time_acc
logging.info('%.3f steps/sec', steps_per_sec)
tf.compat.v2.summary.scalar(name='global_steps_per_sec',
data=steps_per_sec,
step=global_step)
timed_at_step = global_step.numpy()
time_acc = 0
for train_metric in train_metrics:
train_metric.tf_summaries(train_step=global_step,
step_metrics=train_metrics[:2])
if global_step.numpy() % eval_interval == 0:
results = metric_utils.eager_compute(
eval_metrics,
eval_tf_env,
eval_policy,
num_episodes=num_eval_episodes,
train_step=global_step,
summary_writer=eval_summary_writer,
summary_prefix='Metrics',
)
if eval_metrics_callback is not None:
eval_metrics_callback(results, global_step.numpy())
metric_utils.log_metrics(eval_metrics)
return train_loss
def train_eval(
root_dir,
env_name='HalfCheetah-v2',
env_load_fn=suite_mujoco.load,
random_seed=0,
# TODO(b/127576522): rename to policy_fc_layers.
actor_fc_layers=(200, 100),
value_fc_layers=(200, 100),
use_rnns=False,
# Params for collect
num_environment_steps=10000000,
collect_episodes_per_iteration=30,
num_parallel_environments=30,
replay_buffer_capacity=1001, # Per-environment
# Params for train
num_epochs=25,
learning_rate=1e-4,
# Params for eval
num_eval_episodes=30,
eval_interval=500,
# Params for summaries and logging
log_interval=50,
summary_interval=50,
summaries_flush_secs=1,
use_tf_functions=True,
debug_summaries=False,
summarize_grads_and_vars=False):
"""A simple train and eval for PPO."""
if root_dir is None:
raise AttributeError('train_eval requires a root_dir.')
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()
eval_summary_writer = tf.compat.v2.summary.create_file_writer(
eval_dir, flush_millis=summaries_flush_secs * 1000)
eval_metrics = [
tf_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
tf_metrics.AverageEpisodeLengthMetric(buffer_size=num_eval_episodes)
]
global_step = tf.compat.v1.train.get_or_create_global_step()
with tf.compat.v2.summary.record_if(
lambda: tf.math.equal(global_step % summary_interval, 0)):
tf.compat.v1.set_random_seed(random_seed)
eval_tf_env = tf_py_environment.TFPyEnvironment(env_load_fn(env_name))
tf_env = tf_py_environment.TFPyEnvironment(
parallel_py_environment.ParallelPyEnvironment(
[lambda: env_load_fn(env_name)] * num_parallel_environments))
optimizer = tf.compat.v1.train.AdamOptimizer(
learning_rate=learning_rate)
if use_rnns:
actor_net = actor_distribution_rnn_network.ActorDistributionRnnNetwork(
tf_env.observation_spec(),
tf_env.action_spec(),
input_fc_layer_params=actor_fc_layers,
output_fc_layer_params=None)
value_net = value_rnn_network.ValueRnnNetwork(
tf_env.observation_spec(),
input_fc_layer_params=value_fc_layers,
output_fc_layer_params=None)
else:
actor_net = actor_distribution_network.ActorDistributionNetwork(
tf_env.observation_spec(),
tf_env.action_spec(),
fc_layer_params=actor_fc_layers)
value_net = value_network.ValueNetwork(
tf_env.observation_spec(), fc_layer_params=value_fc_layers)
tf_agent = ppo_agent.PPOAgent(
tf_env.time_step_spec(),
tf_env.action_spec(),
optimizer,
actor_net=actor_net,
value_net=value_net,
num_epochs=num_epochs,
debug_summaries=debug_summaries,
summarize_grads_and_vars=summarize_grads_and_vars,
train_step_counter=global_step)
tf_agent.initialize()
environment_steps_metric = tf_metrics.EnvironmentSteps()
step_metrics = [
tf_metrics.NumberOfEpisodes(),
environment_steps_metric,
]
train_metrics = step_metrics + [
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
eval_policy = tf_agent.policy
collect_policy = tf_agent.collect_policy
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
tf_agent.collect_data_spec,
batch_size=num_parallel_environments,
max_length=replay_buffer_capacity)
collect_driver = dynamic_episode_driver.DynamicEpisodeDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_episodes=collect_episodes_per_iteration)
if use_tf_functions:
# TODO(b/123828980): Enable once the cause for slowdown was identified.
collect_driver.run = common.function(collect_driver.run,
autograph=False)
tf_agent.train = common.function(tf_agent.train, autograph=False)
collect_time = 0
train_time = 0
timed_at_step = global_step.numpy()
while environment_steps_metric.result() < num_environment_steps:
global_step_val = global_step.numpy()
if global_step_val % eval_interval == 0:
metric_utils.eager_compute(
eval_metrics,
eval_tf_env,
eval_policy,
num_episodes=num_eval_episodes,
train_step=global_step,
summary_writer=eval_summary_writer,
summary_prefix='Metrics',
)
start_time = time.time()
collect_driver.run()
collect_time += time.time() - start_time
start_time = time.time()
trajectories = replay_buffer.gather_all()
total_loss, _ = tf_agent.train(experience=trajectories)
replay_buffer.clear()
train_time += time.time() - start_time
for train_metric in train_metrics:
train_metric.tf_summaries(train_step=global_step,
step_metrics=step_metrics)
if global_step_val % log_interval == 0:
logging.info('step = %d, loss = %f', global_step_val,
total_loss)
steps_per_sec = ((global_step_val - timed_at_step) /
(collect_time + train_time))
logging.info('%.3f steps/sec', steps_per_sec)
logging.info('collect_time = {}, train_time = {}'.format(
collect_time, train_time))
with tf.compat.v2.summary.record_if(True):
tf.compat.v2.summary.scalar(name='global_steps_per_sec',
data=steps_per_sec,
step=global_step)
timed_at_step = global_step_val
collect_time = 0
train_time = 0
# One final eval before exiting.
metric_utils.eager_compute(
eval_metrics,
eval_tf_env,
eval_policy,
num_episodes=num_eval_episodes,
train_step=global_step,
summary_writer=eval_summary_writer,
summary_prefix='Metrics',
)
def train_eval(
root_dir,
env_name='CartPole-v0',
num_iterations=1000,
# TODO(kbanoop): rename to policy_fc_layers.
actor_fc_layers=(100, ),
# Params for collect
collect_episodes_per_iteration=2,
replay_buffer_capacity=2000,
# Params for train
learning_rate=1e-3,
gradient_clipping=None,
normalize_returns=True,
# Params for eval
num_eval_episodes=10,
eval_interval=100,
# Params for checkpoints, summaries, and logging
log_interval=100,
summary_interval=100,
summaries_flush_secs=1,
debug_summaries=True,
summarize_grads_and_vars=False,
eval_metrics_callback=None):
"""A simple train and eval for Reinforce."""
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()
eval_summary_writer = tf.compat.v2.summary.create_file_writer(
eval_dir, flush_millis=summaries_flush_secs * 1000)
eval_metrics = [
tf_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
tf_metrics.AverageEpisodeLengthMetric(buffer_size=num_eval_episodes),
]
global_step = tf.compat.v1.train.get_or_create_global_step()
with tf.compat.v2.summary.record_if(
lambda: tf.math.equal(global_step % summary_interval, 0)):
tf_env = tf_py_environment.TFPyEnvironment(suite_gym.load(env_name))
eval_tf_env = tf_py_environment.TFPyEnvironment(
suite_gym.load(env_name))
# TODO(kbanoop): Handle distributions without gin.
actor_net = actor_distribution_network.ActorDistributionNetwork(
tf_env.time_step_spec().observation,
tf_env.action_spec(),
fc_layer_params=actor_fc_layers)
tf_agent = reinforce_agent.ReinforceAgent(
tf_env.time_step_spec(),
tf_env.action_spec(),
actor_network=actor_net,
optimizer=tf.compat.v1.train.AdamOptimizer(
learning_rate=learning_rate),
normalize_returns=normalize_returns,
gradient_clipping=gradient_clipping,
debug_summaries=debug_summaries,
summarize_grads_and_vars=summarize_grads_and_vars,
train_step_counter=global_step)
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
tf_agent.collect_data_spec,
batch_size=tf_env.batch_size,
max_length=replay_buffer_capacity)
tf_agent.initialize()
train_metrics = [
tf_metrics.NumberOfEpisodes(),
tf_metrics.EnvironmentSteps(),
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
eval_policy = tf_agent.policy
collect_policy = tf_agent.collect_policy
collect_driver = dynamic_episode_driver.DynamicEpisodeDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_episodes=collect_episodes_per_iteration)
# Compute evaluation metrics.
metrics = metric_utils.eager_compute(
eval_metrics,
eval_tf_env,
eval_policy,
num_episodes=num_eval_episodes,
summary_writer=eval_summary_writer,
summary_prefix='Metrics',
)
# TODO(sfishman): Move this functionality into eager_compute_summaries
if eval_metrics_callback is not None:
eval_metrics_callback(metrics, global_step.numpy())
time_step = None
policy_state = collect_policy.get_initial_state(tf_env.batch_size)
timed_at_step = global_step.numpy()
time_acc = 0
for _ in range(num_iterations):
start_time = time.time()
time_step, policy_state = collect_driver.run(
time_step=time_step,
policy_state=policy_state,
)
experience = replay_buffer.gather_all()
total_loss = tf_agent.train(experience)
replay_buffer.clear()
time_acc += time.time() - start_time
global_step_val = global_step.numpy()
if global_step_val % log_interval == 0:
logging.info('step = %d, loss = %f', global_step_val,
total_loss.loss)
steps_per_sec = (global_step_val - timed_at_step) / time_acc
logging.info('%.3f steps/sec', steps_per_sec)
tf.contrib.summary.scalar(name='global_steps/sec',
tensor=steps_per_sec)
timed_at_step = global_step_val
time_acc = 0
if global_step_val % eval_interval == 0:
metrics = metric_utils.eager_compute(
eval_metrics,
eval_tf_env,
eval_policy,
num_episodes=num_eval_episodes,
summary_writer=eval_summary_writer,
summary_prefix='Metrics',
)
# TODO(sfishman): Move this functionality into eager_compute_summaries
if eval_metrics_callback is not None:
eval_metrics_callback(metrics, global_step_val)
def train_eval(
root_dir,
env_name='CartPole-v0',
num_iterations=100000,
fc_layer_params=(100,),
# Params for collect
initial_collect_steps=1000,
collect_steps_per_iteration=1,
epsilon_greedy=0.1,
replay_buffer_capacity=100000,
# Params for target update
target_update_tau=0.05,
target_update_period=5,
# Params for train
train_steps_per_iteration=1,
batch_size=64,
learning_rate=1e-3,
gamma=0.99,
reward_scale_factor=1.0,
gradient_clipping=None,
# Params for eval
num_eval_episodes=10,
eval_interval=1000,
# Params for summaries and logging
log_interval=1000,
summary_interval=1000,
summaries_flush_secs=10,
debug_summaries=False,
summarize_grads_and_vars=False,
eval_metrics_callback=None):
"""A simple train and eval for DQN."""
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.contrib.summary.create_file_writer(
train_dir, flush_millis=summaries_flush_secs * 1000)
train_summary_writer.set_as_default()
eval_summary_writer = tf.contrib.summary.create_file_writer(
eval_dir, flush_millis=summaries_flush_secs * 1000)
eval_metrics = [
tf_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
tf_metrics.AverageEpisodeLengthMetric(buffer_size=num_eval_episodes)
]
# TODO(kbanoop): Figure out if it is possible to avoid the with block.
with tf.contrib.summary.record_summaries_every_n_global_steps(
summary_interval):
tf_env = tf_py_environment.TFPyEnvironment(suite_gym.load(env_name))
eval_tf_env = tf_py_environment.TFPyEnvironment(suite_gym.load(env_name))
trajectory_spec = trajectory.from_transition(
time_step=tf_env.time_step_spec(),
action_step=policy_step.PolicyStep(action=tf_env.action_spec()),
next_time_step=tf_env.time_step_spec())
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
data_spec=trajectory_spec,
batch_size=tf_env.batch_size,
max_length=replay_buffer_capacity)
q_net = q_network.QNetwork(
tf_env.time_step_spec().observation,
tf_env.action_spec(),
fc_layer_params=fc_layer_params)
tf_agent = dqn_agent.DqnAgent(
tf_env.time_step_spec(),
tf_env.action_spec(),
q_network=q_net,
# TODO(kbanoop): Decay epsilon based on global step, cf. cl/188907839
epsilon_greedy=epsilon_greedy,
target_update_tau=target_update_tau,
target_update_period=target_update_period,
optimizer=tf.train.AdamOptimizer(learning_rate=learning_rate),
td_errors_loss_fn=dqn_agent.element_wise_squared_loss,
gamma=gamma,
reward_scale_factor=reward_scale_factor,
gradient_clipping=gradient_clipping,
debug_summaries=debug_summaries,
summarize_grads_and_vars=summarize_grads_and_vars)
tf_agent.initialize()
train_metrics = [
tf_metrics.NumberOfEpisodes(),
tf_metrics.EnvironmentSteps(),
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
eval_policy = tf_agent.policy()
collect_policy = tf_agent.collect_policy()
collect_driver = dynamic_step_driver.DynamicStepDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_steps=collect_steps_per_iteration)
global_step = tf.train.get_or_create_global_step()
initial_collect_policy = random_tf_policy.RandomTFPolicy(
tf_env.time_step_spec(), tf_env.action_spec())
# Collect initial replay data.
tf.logging.info(
'Initializing replay buffer by collecting experience for %d steps with '
'a random policy.' % initial_collect_steps)
dynamic_step_driver.DynamicStepDriver(
tf_env,
initial_collect_policy,
observers=[replay_buffer.add_batch],
num_steps=initial_collect_steps).run()
metrics = metric_utils.eager_compute(
eval_metrics,
eval_tf_env,
eval_policy,
num_episodes=num_eval_episodes,
summary_writer=eval_summary_writer,
summary_prefix='Metrics',
)
if eval_metrics_callback is not None:
eval_metrics_callback(metrics, global_step.numpy())
time_step = None
policy_state = ()
timed_at_step = global_step.numpy()
time_acc = 0
# Dataset generates trajectories with shape [Bx2x...]
dataset = replay_buffer.as_dataset(
num_parallel_calls=3,
sample_batch_size=batch_size,
num_steps=2).prefetch(3)
iterator = iter(dataset)
for _ in range(num_iterations):
start_time = time.time()
time_step, policy_state = collect_driver.run(
time_step=time_step,
policy_state=policy_state,
)
for _ in range(train_steps_per_iteration):
experience, _ = next(iterator)
train_loss = tf_agent.train(experience, train_step_counter=global_step)
time_acc += time.time() - start_time
if global_step.numpy() % log_interval == 0:
tf.logging.info('step = %d, loss = %f', global_step.numpy(), train_loss)
steps_per_sec = (global_step.numpy() - timed_at_step) / time_acc
tf.logging.info('%.3f steps/sec' % steps_per_sec)
tf.contrib.summary.scalar(name='global_steps/sec', tensor=steps_per_sec)
timed_at_step = global_step.numpy()
time_acc = 0
for train_metric in train_metrics:
train_metric.tf_summaries(step_metrics=train_metrics[:2])
if global_step.numpy() % eval_interval == 0:
metrics = metric_utils.eager_compute(
eval_metrics,
eval_tf_env,
eval_policy,
num_episodes=num_eval_episodes,
summary_writer=eval_summary_writer,
summary_prefix='Metrics',
)
if eval_metrics_callback is not None:
eval_metrics_callback(metrics, global_step.numpy())
return train_loss