def train_eval(
root_dir,
tf_master='',
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
train_checkpoint_interval=100,
policy_checkpoint_interval=50,
rb_checkpoint_interval=200,
log_interval=50,
summary_interval=50,
summaries_flush_secs=1,
debug_summaries=False,
summarize_grads_and_vars=False,
eval_metrics_callback=None):
"""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 = [
batched_py_metric.BatchedPyMetric(
AverageReturnMetric,
metric_args={'buffer_size': num_eval_episodes},
batch_size=num_parallel_environments),
batched_py_metric.BatchedPyMetric(
AverageEpisodeLengthMetric,
metric_args={'buffer_size': num_eval_episodes},
batch_size=num_parallel_environments),
]
eval_summary_writer_flush_op = eval_summary_writer.flush()
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_py_env = parallel_py_environment.ParallelPyEnvironment(
[lambda: env_load_fn(env_name)] * num_parallel_environments)
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)
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
tf_agent.collect_data_spec,
batch_size=num_parallel_environments,
max_length=replay_buffer_capacity)
eval_py_policy = py_tf_policy.PyTFPolicy(tf_agent.policy)
environment_steps_metric = tf_metrics.EnvironmentSteps()
environment_steps_count = environment_steps_metric.result()
step_metrics = [
tf_metrics.NumberOfEpisodes(),
environment_steps_metric,
]
train_metrics = step_metrics + [
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
# Add to replay buffer and other agent specific observers.
replay_buffer_observer = [replay_buffer.add_batch]
collect_policy = tf_agent.collect_policy
collect_op = dynamic_episode_driver.DynamicEpisodeDriver(
tf_env,
collect_policy,
observers=replay_buffer_observer + train_metrics,
num_episodes=collect_episodes_per_iteration).run()
trajectories = replay_buffer.gather_all()
train_op, _ = tf_agent.train(experience=trajectories)
with tf.control_dependencies([train_op]):
clear_replay_op = replay_buffer.clear()
with tf.control_dependencies([clear_replay_op]):
train_op = tf.identity(train_op)
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=tf_agent.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)
for train_metric in train_metrics:
train_metric.tf_summaries(train_step=global_step,
step_metrics=step_metrics)
with eval_summary_writer.as_default(), \
tf.compat.v2.summary.record_if(True):
for eval_metric in eval_metrics:
eval_metric.tf_summaries(step_metrics=step_metrics)
init_agent_op = tf_agent.initialize()
with tf.compat.v1.Session(tf_master) as sess:
# Initialize graph.
train_checkpointer.initialize_or_restore(sess)
rb_checkpointer.initialize_or_restore(sess)
common.initialize_uninitialized_variables(sess)
sess.run(init_agent_op)
sess.run(train_summary_writer.init())
sess.run(eval_summary_writer.init())
collect_time = 0
train_time = 0
timed_at_step = sess.run(global_step)
steps_per_second_ph = tf.compat.v1.placeholder(
tf.float32, shape=(), name='steps_per_sec_ph')
steps_per_second_summary = tf.contrib.summary.scalar(
name='global_steps/sec', tensor=steps_per_second_ph)
while sess.run(environment_steps_count) < num_environment_steps:
global_step_val = sess.run(global_step)
if global_step_val % eval_interval == 0:
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
log=True,
)
sess.run(eval_summary_writer_flush_op)
start_time = time.time()
sess.run(collect_op)
collect_time += time.time() - start_time
start_time = time.time()
total_loss = sess.run(train_op)
train_time += time.time() - start_time
global_step_val = sess.run(global_step)
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)
sess.run(steps_per_second_summary,
feed_dict={steps_per_second_ph: steps_per_sec})
logging.info(
'%s', 'collect_time = {}, train_time = {}'.format(
collect_time, train_time))
timed_at_step = global_step_val
collect_time = 0
train_time = 0
if global_step_val % train_checkpoint_interval == 0:
train_checkpointer.save(global_step=global_step_val)
if global_step_val % policy_checkpoint_interval == 0:
policy_checkpointer.save(global_step=global_step_val)
if global_step_val % rb_checkpoint_interval == 0:
rb_checkpointer.save(global_step=global_step_val)
# One final eval before exiting.
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
log=True,
)
sess.run(eval_summary_writer_flush_op)
def train_eval(
root_dir,
env_name='HalfCheetah-v1',
env_load_fn=suite_mujoco.load,
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,
num_parallel_environments=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.losses.huber_loss,
gamma=0.995,
reward_scale_factor=1.0,
gradient_clipping=None,
# Params for eval
num_eval_episodes=10,
eval_interval=10000,
# Params for checkpoints, summaries, and logging
train_checkpoint_interval=10000,
policy_checkpoint_interval=5000,
rb_checkpoint_interval=20000,
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 DDPG."""
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 = [
py_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
py_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):
if num_parallel_environments > 1:
tf_env = tf_py_environment.TFPyEnvironment(
parallel_py_environment.ParallelPyEnvironment(
[lambda: env_load_fn(env_name)] * num_parallel_environments))
else:
tf_env = tf_py_environment.TFPyEnvironment(env_load_fn(env_name))
eval_py_env = env_load_fn(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 = ddpg_agent.DdpgAgent(
tf_env.time_step_spec(),
tf_env.action_spec(),
actor_network=actor_net,
critic_network=critic_net,
actor_optimizer=tf.train.AdamOptimizer(
learning_rate=actor_learning_rate),
critic_optimizer=tf.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)
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
tf_agent.collect_data_spec(),
batch_size=tf_env.batch_size,
max_length=replay_buffer_capacity)
eval_py_policy = py_tf_policy.PyTFPolicy(tf_agent.policy())
train_metrics = [
tf_metrics.NumberOfEpisodes(),
tf_metrics.EnvironmentSteps(),
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
global_step = tf.train.get_or_create_global_step()
collect_policy = tf_agent.collect_policy()
initial_collect_op = dynamic_step_driver.DynamicStepDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch],
num_steps=initial_collect_steps).run()
collect_op = dynamic_step_driver.DynamicStepDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_steps=collect_steps_per_iteration).run()
# 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 = dataset.make_initializable_iterator()
trajectories, unused_info = iterator.get_next()
train_op = tf_agent.train(
experience=trajectories, train_step_counter=global_step)
train_checkpointer = common_utils.Checkpointer(
ckpt_dir=train_dir,
agent=tf_agent,
global_step=global_step,
metrics=tf.contrib.checkpoint.List(train_metrics))
policy_checkpointer = common_utils.Checkpointer(
ckpt_dir=os.path.join(train_dir, 'policy'),
policy=tf_agent.policy(),
global_step=global_step)
rb_checkpointer = common_utils.Checkpointer(
ckpt_dir=os.path.join(train_dir, 'replay_buffer'),
max_to_keep=1,
replay_buffer=replay_buffer)
for train_metric in train_metrics:
train_metric.tf_summaries(step_metrics=train_metrics[:2])
summary_op = tf.contrib.summary.all_summary_ops()
with eval_summary_writer.as_default(), \
tf.contrib.summary.always_record_summaries():
for eval_metric in eval_metrics:
eval_metric.tf_summaries()
init_agent_op = tf_agent.initialize()
with tf.Session() as sess:
# Initialize the graph.
train_checkpointer.initialize_or_restore(sess)
rb_checkpointer.initialize_or_restore(sess)
sess.run(iterator.initializer)
# TODO(sguada) Remove once Periodically can be saved.
common_utils.initialize_uninitialized_variables(sess)
sess.run(init_agent_op)
tf.contrib.summary.initialize(session=sess)
sess.run(initial_collect_op)
global_step_val = sess.run(global_step)
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
)
collect_call = sess.make_callable(collect_op)
train_step_call = sess.make_callable([train_op, summary_op, global_step])
timed_at_step = sess.run(global_step)
time_acc = 0
steps_per_second_ph = tf.placeholder(
tf.float32, shape=(), name='steps_per_sec_ph')
steps_per_second_summary = tf.contrib.summary.scalar(
name='global_steps/sec', tensor=steps_per_second_ph)
for _ in range(num_iterations):
start_time = time.time()
collect_call()
for _ in range(train_steps_per_iteration):
loss_info_value, _, global_step_val = train_step_call()
time_acc += time.time() - start_time
if global_step_val % log_interval == 0:
tf.logging.info('step = %d, loss = %f', global_step_val,
loss_info_value.loss)
steps_per_sec = (global_step_val - timed_at_step) / time_acc
tf.logging.info('%.3f steps/sec' % steps_per_sec)
sess.run(
steps_per_second_summary,
feed_dict={steps_per_second_ph: steps_per_sec})
timed_at_step = global_step_val
time_acc = 0
if global_step_val % train_checkpoint_interval == 0:
train_checkpointer.save(global_step=global_step_val)
if global_step_val % policy_checkpoint_interval == 0:
policy_checkpointer.save(global_step=global_step_val)
if global_step_val % rb_checkpoint_interval == 0:
rb_checkpointer.save(global_step=global_step_val)
if global_step_val % eval_interval == 0:
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
)
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
train_checkpoint_interval=100,
policy_checkpoint_interval=100,
rb_checkpoint_interval=200,
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 = [
py_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
py_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)):
eval_py_env = suite_gym.load(env_name)
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)
eval_py_policy = py_tf_policy.PyTFPolicy(tf_agent.policy)
train_metrics = [
tf_metrics.NumberOfEpisodes(),
tf_metrics.EnvironmentSteps(),
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
collect_policy = tf_agent.collect_policy
collect_op = dynamic_episode_driver.DynamicEpisodeDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_episodes=collect_episodes_per_iteration).run()
experience = replay_buffer.gather_all()
train_op = tf_agent.train(experience)
clear_rb_op = replay_buffer.clear()
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=tf_agent.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)
for train_metric in train_metrics:
train_metric.tf_summaries(train_step=global_step,
step_metrics=train_metrics[:2])
with eval_summary_writer.as_default(), \
tf.compat.v2.summary.record_if(True):
for eval_metric in eval_metrics:
eval_metric.tf_summaries()
init_agent_op = tf_agent.initialize()
with tf.compat.v1.Session() as sess:
# Initialize the graph.
train_checkpointer.initialize_or_restore(sess)
rb_checkpointer.initialize_or_restore(sess)
# TODO(sguada) Remove once Periodically can be saved.
common.initialize_uninitialized_variables(sess)
sess.run(init_agent_op)
sess.run(train_summary_writer.init())
sess.run(eval_summary_writer.init())
# Compute evaluation metrics.
global_step_call = sess.make_callable(global_step)
global_step_val = global_step_call()
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
)
collect_call = sess.make_callable(collect_op)
train_step_call = sess.make_callable(train_op)
clear_rb_call = sess.make_callable(clear_rb_op)
timed_at_step = global_step_call()
time_acc = 0
steps_per_second_ph = tf.compat.v1.placeholder(
tf.float32, shape=(), name='steps_per_sec_ph')
steps_per_second_summary = tf.contrib.summary.scalar(
name='global_steps/sec', tensor=steps_per_second_ph)
for _ in range(num_iterations):
start_time = time.time()
collect_call()
total_loss = train_step_call()
clear_rb_call()
time_acc += time.time() - start_time
global_step_val = global_step_call()
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)
sess.run(steps_per_second_summary,
feed_dict={steps_per_second_ph: steps_per_sec})
timed_at_step = global_step_val
time_acc = 0
if global_step_val % train_checkpoint_interval == 0:
train_checkpointer.save(global_step=global_step_val)
if global_step_val % policy_checkpoint_interval == 0:
policy_checkpointer.save(global_step=global_step_val)
if global_step_val % rb_checkpoint_interval == 0:
rb_checkpointer.save(global_step=global_step_val)
if global_step_val % eval_interval == 0:
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
)
def train_eval(
root_dir,
env_name='HalfCheetah-v1',
num_iterations=1000000,
actor_fc_layers=(256, 256),
critic_obs_fc_layers=None,
critic_action_fc_layers=None,
critic_joint_fc_layers=(256, 256),
# Params for collect
initial_collect_steps=10000,
collect_steps_per_iteration=1,
replay_buffer_capacity=1000000,
# Params for target update
target_update_tau=0.005,
target_update_period=1,
# Params for train
train_steps_per_iteration=1,
batch_size=256,
actor_learning_rate=3e-4,
critic_learning_rate=3e-4,
alpha_learning_rate=3e-4,
td_errors_loss_fn=tf.compat.v1.losses.mean_squared_error,
gamma=0.99,
reward_scale_factor=1.0,
gradient_clipping=None,
# Params for eval
num_eval_episodes=100,
eval_interval=500000,
# Params for summaries and logging
train_checkpoint_interval=10000,
policy_checkpoint_interval=5000,
rb_checkpoint_interval=50000,
log_interval=10000,
summary_interval=10000,
summaries_flush_secs=10,
debug_summaries=False,
summarize_grads_and_vars=False,
eval_metrics_callback=None):
"""A simple train and eval for SAC."""
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 = [
py_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
py_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)):
# Create the environment.
tf_env = tf_py_environment.TFPyEnvironment(suite_mujoco.load(env_name))
eval_py_env = suite_mujoco.load(env_name)
# Get the data specs from the environment
time_step_spec = tf_env.time_step_spec()
observation_spec = time_step_spec.observation
action_spec = tf_env.action_spec()
actor_net = actor_distribution_network.ActorDistributionNetwork(
observation_spec,
action_spec,
fc_layer_params=actor_fc_layers,
continuous_projection_net=normal_projection_net)
critic_net = critic_network.CriticNetwork(
(observation_spec, action_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)
tf_agent = sac_agent.SacAgent(
time_step_spec,
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),
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=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)
# Make the replay buffer.
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
data_spec=tf_agent.collect_data_spec,
batch_size=1,
max_length=replay_buffer_capacity)
replay_observer = [replay_buffer.add_batch]
eval_py_policy = py_tf_policy.PyTFPolicy(tf_agent.policy)
train_metrics = [
tf_metrics.NumberOfEpisodes(),
tf_metrics.EnvironmentSteps(),
tf_py_metric.TFPyMetric(py_metrics.AverageReturnMetric()),
tf_py_metric.TFPyMetric(py_metrics.AverageEpisodeLengthMetric()),
]
collect_policy = tf_agent.collect_policy
initial_collect_op = dynamic_step_driver.DynamicStepDriver(
tf_env,
collect_policy,
observers=replay_observer + train_metrics,
num_steps=initial_collect_steps).run()
collect_op = dynamic_step_driver.DynamicStepDriver(
tf_env,
collect_policy,
observers=replay_observer + train_metrics,
num_steps=collect_steps_per_iteration).run()
# Prepare replay buffer as dataset with invalid transitions filtered.
def _filter_invalid_transition(trajectories, unused_arg1):
return ~trajectories.is_boundary()[0]
dataset = replay_buffer.as_dataset(
sample_batch_size=5 * batch_size,
num_steps=2).apply(tf.data.experimental.unbatch()).filter(
_filter_invalid_transition).batch(batch_size).prefetch(
batch_size * 5)
dataset_iterator = tf.compat.v1.data.make_initializable_iterator(
dataset)
trajectories, unused_info = dataset_iterator.get_next()
train_op = tf_agent.train(trajectories)
for train_metric in train_metrics:
train_metric.tf_summaries(step_metrics=train_metrics[:2])
with eval_summary_writer.as_default(), \
tf.compat.v2.summary.record_if(True):
for eval_metric in eval_metrics:
eval_metric.tf_summaries()
train_checkpointer = common_utils.Checkpointer(
ckpt_dir=train_dir,
agent=tf_agent,
global_step=global_step,
metrics=metric_utils.MetricsGroup(train_metrics, 'train_metrics'))
policy_checkpointer = common_utils.Checkpointer(
ckpt_dir=os.path.join(train_dir, 'policy'),
policy=tf_agent.policy,
global_step=global_step)
rb_checkpointer = common_utils.Checkpointer(
ckpt_dir=os.path.join(train_dir, 'replay_buffer'),
max_to_keep=1,
replay_buffer=replay_buffer)
with tf.compat.v1.Session() as sess:
# Initialize graph.
train_checkpointer.initialize_or_restore(sess)
policy_checkpointer.initialize_or_restore(sess)
rb_checkpointer.initialize_or_restore(sess)
# Initialize training.
sess.run(dataset_iterator.initializer)
common_utils.initialize_uninitialized_variables(sess)
sess.run(train_summary_writer.init())
sess.run(eval_summary_writer.init())
global_step_val = sess.run(global_step)
if global_step_val == 0:
# Initial eval of randomly initialized policy
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
log=True,
)
# Run initial collect.
logging.info('Global step %d: Running initial collect op.',
global_step_val)
sess.run(initial_collect_op)
# Checkpoint the initial replay buffer contents.
rb_checkpointer.save(global_step=global_step_val)
logging.info('Finished initial collect.')
else:
logging.info('Global step %d: Skipping initial collect op.',
global_step_val)
collect_call = sess.make_callable(collect_op)
train_step_call = sess.make_callable(train_op)
global_step_call = sess.make_callable(global_step)
timed_at_step = global_step_call()
time_acc = 0
steps_per_second_ph = tf.compat.v1.placeholder(
tf.float32, shape=(), name='steps_per_sec_ph')
steps_per_second_summary = tf.contrib.summary.scalar(
name='global_steps/sec', tensor=steps_per_second_ph)
for _ in range(num_iterations):
start_time = time.time()
collect_call()
for _ in range(train_steps_per_iteration):
total_loss = train_step_call()
time_acc += time.time() - start_time
global_step_val = global_step_call()
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)
sess.run(steps_per_second_summary,
feed_dict={steps_per_second_ph: steps_per_sec})
timed_at_step = global_step_val
time_acc = 0
if global_step_val % eval_interval == 0:
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
log=True,
)
if global_step_val % train_checkpoint_interval == 0:
train_checkpointer.save(global_step=global_step_val)
if global_step_val % policy_checkpoint_interval == 0:
policy_checkpointer.save(global_step=global_step_val)
if global_step_val % rb_checkpoint_interval == 0:
rb_checkpointer.save(global_step=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 checkpoints, summaries and logging
train_checkpoint_interval=10000,
policy_checkpoint_interval=5000,
log_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 = [
py_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
py_metrics.AverageEpisodeLengthMetric(buffer_size=num_eval_episodes),
]
# Note this is a python environment.
env = batched_py_environment.BatchedPyEnvironment(
[suite_gym.load(env_name)])
eval_py_env = suite_gym.load(env_name)
# Convert specs to BoundedTensorSpec.
action_spec = tensor_spec.from_spec(env.action_spec())
observation_spec = tensor_spec.from_spec(env.observation_spec())
time_step_spec = ts.time_step_spec(observation_spec)
q_net = q_network.QNetwork(tensor_spec.from_spec(env.observation_spec()),
tensor_spec.from_spec(env.action_spec()),
fc_layer_params=fc_layer_params)
# The agent must be in graph.
global_step = tf.compat.v1.train.get_or_create_global_step()
agent = dqn_agent.DqnAgent(
time_step_spec,
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_collect_policy = agent.collect_policy
collect_policy = py_tf_policy.PyTFPolicy(tf_collect_policy)
greedy_policy = py_tf_policy.PyTFPolicy(agent.policy)
random_policy = random_py_policy.RandomPyPolicy(env.time_step_spec(),
env.action_spec())
# Python replay buffer.
replay_buffer = py_uniform_replay_buffer.PyUniformReplayBuffer(
capacity=replay_buffer_capacity,
data_spec=tensor_spec.to_nest_array_spec(agent.collect_data_spec))
time_step = env.reset()
# Initialize the replay buffer with some transitions. We use the random
# policy to initialize the replay buffer to make sure we get a good
# distribution of actions.
for _ in range(initial_collect_steps):
time_step = collect_step(env, time_step, random_policy, replay_buffer)
# TODO(b/112041045) Use global_step as counter.
train_checkpointer = common.Checkpointer(ckpt_dir=train_dir,
agent=agent,
global_step=global_step)
policy_checkpointer = common.Checkpointer(ckpt_dir=os.path.join(
train_dir, 'policy'),
policy=agent.policy,
global_step=global_step)
ds = replay_buffer.as_dataset(sample_batch_size=batch_size, num_steps=2)
ds = ds.prefetch(4)
itr = tf.compat.v1.data.make_initializable_iterator(ds)
experience = itr.get_next()
train_op = common.function(agent.train)(experience)
with eval_summary_writer.as_default(), \
tf.compat.v2.summary.record_if(True):
for eval_metric in eval_metrics:
eval_metric.tf_summaries()
with tf.compat.v1.Session() as session:
train_checkpointer.initialize_or_restore(session)
common.initialize_uninitialized_variables(session)
session.run(itr.initializer)
# Copy critic network values to the target critic network.
session.run(agent.initialize())
train = session.make_callable(train_op)
global_step_call = session.make_callable(global_step)
session.run(train_summary_writer.init())
session.run(eval_summary_writer.init())
# Compute inital evaluation metrics.
global_step_val = global_step_call()
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
greedy_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
log=True,
callback=eval_metrics_callback,
)
timed_at_step = global_step_val
collect_time = 0
train_time = 0
steps_per_second_ph = tf.compat.v1.placeholder(tf.float32,
shape=(),
name='steps_per_sec_ph')
steps_per_second_summary = tf.contrib.summary.scalar(
name='global_steps/sec', tensor=steps_per_second_ph)
for _ in range(num_iterations):
start_time = time.time()
for _ in range(collect_steps_per_iteration):
time_step = collect_step(env, time_step, collect_policy,
replay_buffer)
collect_time += time.time() - start_time
start_time = time.time()
for _ in range(train_steps_per_iteration):
loss = train()
train_time += time.time() - start_time
global_step_val = global_step_call()
if global_step_val % log_interval == 0:
logging.info('step = %d, loss = %f', global_step_val,
loss.loss)
steps_per_sec = ((global_step_val - timed_at_step) /
(collect_time + train_time))
session.run(steps_per_second_summary,
feed_dict={steps_per_second_ph: steps_per_sec})
logging.info('%.3f steps/sec', steps_per_sec)
logging.info(
'%s', 'collect_time = {}, train_time = {}'.format(
collect_time, train_time))
timed_at_step = global_step_val
collect_time = 0
train_time = 0
if global_step_val % train_checkpoint_interval == 0:
train_checkpointer.save(global_step=global_step_val)
if global_step_val % policy_checkpoint_interval == 0:
policy_checkpointer.save(global_step=global_step_val)
if global_step_val % eval_interval == 0:
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
greedy_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
log=True,
callback=eval_metrics_callback,
)
# Reset timing to avoid counting eval time.
timed_at_step = global_step_val
start_time = time.time()
def train_eval(
root_dir,
env_name='cartpole',
task_name='balance',
observations_whitelist='position',
num_iterations=100000,
actor_fc_layers=(400, 300),
actor_output_fc_layers=(100, ),
actor_lstm_size=(40, ),
critic_obs_fc_layers=(400, ),
critic_action_fc_layers=None,
critic_joint_fc_layers=(300, ),
critic_output_fc_layers=(100, ),
critic_lstm_size=(40, ),
# Params for collect
initial_collect_steps=1,
collect_episodes_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=200,
batch_size=64,
train_sequence_length=10,
actor_learning_rate=1e-4,
critic_learning_rate=1e-3,
dqda_clipping=None,
gamma=0.995,
reward_scale_factor=1.0,
gradient_clipping=None,
# Params for eval
num_eval_episodes=10,
eval_interval=1000,
# Params for checkpoints, summaries, and logging
train_checkpoint_interval=10000,
policy_checkpoint_interval=5000,
rb_checkpoint_interval=10000,
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 DDPG."""
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 = [
py_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
py_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)):
if observations_whitelist is not None:
env_wrappers = [
functools.partial(
wrappers.FlattenObservationsWrapper,
observations_whitelist=[observations_whitelist])
]
else:
env_wrappers = []
environment = suite_dm_control.load(env_name,
task_name,
env_wrappers=env_wrappers)
tf_env = tf_py_environment.TFPyEnvironment(environment)
eval_py_env = suite_dm_control.load(env_name,
task_name,
env_wrappers=env_wrappers)
actor_net = actor_rnn_network.ActorRnnNetwork(
tf_env.time_step_spec().observation,
tf_env.action_spec(),
input_fc_layer_params=actor_fc_layers,
lstm_size=actor_lstm_size,
output_fc_layer_params=actor_output_fc_layers)
critic_net_input_specs = (tf_env.time_step_spec().observation,
tf_env.action_spec())
critic_net = critic_rnn_network.CriticRnnNetwork(
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,
lstm_size=critic_lstm_size,
output_fc_layer_params=critic_output_fc_layers,
)
tf_agent = ddpg_agent.DdpgAgent(
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,
gamma=gamma,
reward_scale_factor=reward_scale_factor,
gradient_clipping=gradient_clipping,
debug_summaries=debug_summaries,
summarize_grads_and_vars=summarize_grads_and_vars)
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
tf_agent.collect_data_spec,
batch_size=tf_env.batch_size,
max_length=replay_buffer_capacity)
eval_py_policy = py_tf_policy.PyTFPolicy(tf_agent.policy)
train_metrics = [
tf_metrics.NumberOfEpisodes(),
tf_metrics.EnvironmentSteps(),
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
collect_policy = tf_agent.collect_policy
initial_collect_op = dynamic_episode_driver.DynamicEpisodeDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_episodes=initial_collect_steps).run()
collect_op = dynamic_episode_driver.DynamicEpisodeDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_episodes=collect_episodes_per_iteration).run()
# Need extra step to generate transitions of train_sequence_length.
# Dataset generates trajectories with shape [BxTx...]
dataset = replay_buffer.as_dataset(num_parallel_calls=3,
sample_batch_size=batch_size,
num_steps=train_sequence_length +
1).prefetch(3)
iterator = tf.compat.v1.data.make_initializable_iterator(dataset)
trajectories, unused_info = iterator.get_next()
train_fn = common.function(tf_agent.train)
train_op = train_fn(experience=trajectories,
train_step_counter=global_step)
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=tf_agent.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)
for train_metric in train_metrics:
train_metric.tf_summaries(train_step=global_step,
step_metrics=train_metrics[:2])
summary_op = tf.contrib.summary.all_summary_ops()
with eval_summary_writer.as_default(), \
tf.compat.v2.summary.record_if(True):
for eval_metric in eval_metrics:
eval_metric.tf_summaries()
init_agent_op = tf_agent.initialize()
with tf.compat.v1.Session() as sess:
# Initialize the graph.
train_checkpointer.initialize_or_restore(sess)
rb_checkpointer.initialize_or_restore(sess)
sess.run(iterator.initializer)
# TODO(b/126239733) Remove once Periodically can be saved.
common.initialize_uninitialized_variables(sess)
sess.run(init_agent_op)
tf.contrib.summary.initialize(session=sess)
sess.run(initial_collect_op)
global_step_val = sess.run(global_step)
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
log=True,
)
collect_call = sess.make_callable(collect_op)
train_step_call = sess.make_callable([train_op, summary_op])
global_step_call = sess.make_callable(global_step)
timed_at_step = global_step_call()
time_acc = 0
steps_per_second_ph = tf.compat.v1.placeholder(
tf.float32, shape=(), name='steps_per_sec_ph')
steps_per_second_summary = tf.contrib.summary.scalar(
name='global_steps/sec', tensor=steps_per_second_ph)
for _ in range(num_iterations):
start_time = time.time()
collect_call()
for _ in range(train_steps_per_iteration):
loss_info_value, _ = train_step_call()
global_step_val = global_step_call()
time_acc += time.time() - start_time
if global_step_val % log_interval == 0:
logging.info('step = %d, loss = %f', global_step_val,
loss_info_value.loss)
steps_per_sec = (global_step_val -
timed_at_step) / time_acc
logging.info('%.3f steps/sec', steps_per_sec)
sess.run(steps_per_second_summary,
feed_dict={steps_per_second_ph: steps_per_sec})
timed_at_step = global_step_val
time_acc = 0
if global_step_val % train_checkpoint_interval == 0:
train_checkpointer.save(global_step=global_step_val)
if global_step_val % policy_checkpoint_interval == 0:
policy_checkpointer.save(global_step=global_step_val)
if global_step_val % rb_checkpoint_interval == 0:
rb_checkpointer.save(global_step=global_step_val)
if global_step_val % eval_interval == 0:
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
log=True)
def train_eval(
root_dir,
env_name='MaskedCartPole-v0',
num_iterations=100000,
input_fc_layer_params=(50, ),
lstm_size=(20, ),
output_fc_layer_params=(20, ),
train_sequence_length=10,
# Params for collect
initial_collect_steps=50,
collect_episodes_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=10,
batch_size=128,
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
train_checkpoint_interval=10000,
policy_checkpoint_interval=5000,
rb_checkpoint_interval=20000,
log_interval=100,
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 = [
py_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
py_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)):
eval_py_env = suite_gym.load(env_name)
tf_env = tf_py_environment.TFPyEnvironment(suite_gym.load(env_name))
q_net = q_rnn_network.QRnnNetwork(
tf_env.time_step_spec().observation,
tf_env.action_spec(),
input_fc_layer_params=input_fc_layer_params,
lstm_size=lstm_size,
output_fc_layer_params=output_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,
optimizer=tf.compat.v1.train.AdamOptimizer(
learning_rate=learning_rate),
epsilon_greedy=epsilon_greedy,
target_update_tau=target_update_tau,
target_update_period=target_update_period,
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)
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
tf_agent.collect_data_spec,
batch_size=tf_env.batch_size,
max_length=replay_buffer_capacity)
eval_py_policy = py_tf_policy.PyTFPolicy(tf_agent.policy)
train_metrics = [
tf_metrics.NumberOfEpisodes(),
tf_metrics.EnvironmentSteps(),
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
initial_collect_policy = random_tf_policy.RandomTFPolicy(
tf_env.time_step_spec(), tf_env.action_spec())
initial_collect_op = dynamic_episode_driver.DynamicEpisodeDriver(
tf_env,
initial_collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_episodes=initial_collect_steps).run()
collect_policy = tf_agent.collect_policy
collect_op = dynamic_episode_driver.DynamicEpisodeDriver(
tf_env,
collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_episodes=collect_episodes_per_iteration).run()
# Need extra step to generate transitions of train_sequence_length.
# Dataset generates trajectories with shape [BxTx...]
dataset = replay_buffer.as_dataset(num_parallel_calls=3,
sample_batch_size=batch_size,
num_steps=train_sequence_length +
1).prefetch(3)
iterator = tf.compat.v1.data.make_initializable_iterator(dataset)
experience, _ = iterator.get_next()
loss_info = common.function(tf_agent.train)(experience=experience)
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=tf_agent.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)
summary_ops = []
for train_metric in train_metrics:
summary_ops.append(
train_metric.tf_summaries(train_step=global_step,
step_metrics=train_metrics[:2]))
with eval_summary_writer.as_default(), \
tf.compat.v2.summary.record_if(True):
for eval_metric in eval_metrics:
eval_metric.tf_summaries(train_step=global_step)
init_agent_op = tf_agent.initialize()
with tf.compat.v1.Session() as sess:
sess.run(train_summary_writer.init())
sess.run(eval_summary_writer.init())
# Initialize the graph.
train_checkpointer.initialize_or_restore(sess)
rb_checkpointer.initialize_or_restore(sess)
sess.run(iterator.initializer)
common.initialize_uninitialized_variables(sess)
sess.run(init_agent_op)
logging.info('Collecting initial experience.')
sess.run(initial_collect_op)
# Compute evaluation metrics.
global_step_val = sess.run(global_step)
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
log=True,
)
collect_call = sess.make_callable(collect_op)
train_step_call = sess.make_callable([loss_info, summary_ops])
global_step_call = sess.make_callable(global_step)
timed_at_step = global_step_call()
time_acc = 0
steps_per_second_ph = tf.compat.v1.placeholder(
tf.float32, shape=(), name='steps_per_sec_ph')
steps_per_second_summary = tf.compat.v2.summary.scalar(
name='global_steps_per_sec',
data=steps_per_second_ph,
step=global_step)
for _ in range(num_iterations):
# Train/collect/eval.
start_time = time.time()
collect_call()
for _ in range(train_steps_per_iteration):
loss_info_value, _ = train_step_call()
time_acc += time.time() - start_time
global_step_val = global_step_call()
if global_step_val % log_interval == 0:
logging.info('step = %d, loss = %f', global_step_val,
loss_info_value.loss)
steps_per_sec = (global_step_val -
timed_at_step) / time_acc
logging.info('%.3f steps/sec', steps_per_sec)
sess.run(steps_per_second_summary,
feed_dict={steps_per_second_ph: steps_per_sec})
timed_at_step = global_step_val
time_acc = 0
if global_step_val % train_checkpoint_interval == 0:
train_checkpointer.save(global_step=global_step_val)
if global_step_val % policy_checkpoint_interval == 0:
policy_checkpointer.save(global_step=global_step_val)
if global_step_val % rb_checkpoint_interval == 0:
rb_checkpointer.save(global_step=global_step_val)
if global_step_val % eval_interval == 0:
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
log=True,
callback=eval_metrics_callback,
)
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 checkpoints, summaries, and logging
train_checkpoint_interval=10000,
policy_checkpoint_interval=5000,
rb_checkpoint_interval=20000,
log_interval=1000,
summary_interval=1000,
summaries_flush_secs=10,
agent_class=dqn_agent.DqnAgent,
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 = [
py_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
py_metrics.AverageEpisodeLengthMetric(buffer_size=num_eval_episodes),
]
with tf.contrib.summary.record_summaries_every_n_global_steps(
summary_interval):
tf_env = tf_py_environment.TFPyEnvironment(suite_gym.load(env_name))
eval_py_env = suite_gym.load(env_name)
q_net = q_network.QNetwork(tf_env.time_step_spec().observation,
tf_env.action_spec(),
fc_layer_params=fc_layer_params)
tf_agent = agent_class(
tf_env.time_step_spec(),
tf_env.action_spec(),
q_network=q_net,
optimizer=tf.train.AdamOptimizer(learning_rate=learning_rate),
# 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,
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)
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
tf_agent.collect_data_spec(),
batch_size=tf_env.batch_size,
max_length=replay_buffer_capacity)
eval_py_policy = py_tf_policy.PyTFPolicy(tf_agent.policy())
train_metrics = [
tf_metrics.NumberOfEpisodes(),
tf_metrics.EnvironmentSteps(),
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
global_step = tf.train.get_or_create_global_step()
replay_observer = [replay_buffer.add_batch]
initial_collect_policy = random_tf_policy.RandomTFPolicy(
tf_env.time_step_spec(), tf_env.action_spec())
initial_collect_op = dynamic_step_driver.DynamicStepDriver(
tf_env,
initial_collect_policy,
observers=replay_observer,
num_steps=initial_collect_steps).run()
collect_policy = tf_agent.collect_policy()
collect_op = dynamic_step_driver.DynamicStepDriver(
tf_env,
collect_policy,
observers=replay_observer + train_metrics,
num_steps=collect_steps_per_iteration).run()
# 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 = dataset.make_initializable_iterator()
trajectories, _ = iterator.get_next()
train_op = tf_agent.train(experience=trajectories,
train_step_counter=global_step)
train_checkpointer = common_utils.Checkpointer(
ckpt_dir=train_dir,
agent=tf_agent,
global_step=global_step,
metrics=tf.contrib.checkpoint.List(train_metrics))
policy_checkpointer = common_utils.Checkpointer(
ckpt_dir=os.path.join(train_dir, 'policy'),
policy=tf_agent.policy(),
global_step=global_step)
rb_checkpointer = common_utils.Checkpointer(
ckpt_dir=os.path.join(train_dir, 'replay_buffer'),
max_to_keep=1,
replay_buffer=replay_buffer)
for train_metric in train_metrics:
train_metric.tf_summaries(step_metrics=train_metrics[:2])
summary_op = tf.contrib.summary.all_summary_ops()
with eval_summary_writer.as_default(), \
tf.contrib.summary.always_record_summaries():
for eval_metric in eval_metrics:
eval_metric.tf_summaries()
init_agent_op = tf_agent.initialize()
with tf.Session() as sess:
# Initialize the graph.
train_checkpointer.initialize_or_restore(sess)
rb_checkpointer.initialize_or_restore(sess)
sess.run(iterator.initializer)
# TODO(sguada) Remove once Periodically can be saved.
common_utils.initialize_uninitialized_variables(sess)
sess.run(init_agent_op)
tf.contrib.summary.initialize(session=sess)
sess.run(initial_collect_op)
global_step_val = sess.run(global_step)
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
)
collect_call = sess.make_callable(collect_op)
train_step_call = sess.make_callable(
[train_op, summary_op, global_step])
timed_at_step = sess.run(global_step)
collect_time = 0
train_time = 0
steps_per_second_ph = tf.placeholder(tf.float32,
shape=(),
name='steps_per_sec_ph')
steps_per_second_summary = tf.contrib.summary.scalar(
name='global_steps/sec', tensor=steps_per_second_ph)
for _ in range(num_iterations):
# Train/collect/eval.
start_time = time.time()
collect_call()
collect_time += time.time() - start_time
start_time = time.time()
for _ in range(train_steps_per_iteration):
loss_info_value, _, global_step_val = train_step_call()
train_time += time.time() - start_time
if global_step_val % log_interval == 0:
tf.logging.info('step = %d, loss = %f', global_step_val,
loss_info_value.loss)
steps_per_sec = ((global_step_val - timed_at_step) /
(collect_time + train_time))
sess.run(steps_per_second_summary,
feed_dict={steps_per_second_ph: steps_per_sec})
tf.logging.info('%.3f steps/sec' % steps_per_sec)
tf.logging.info(
'collect_time = {}, train_time = {}'.format(
collect_time, train_time))
timed_at_step = global_step_val
collect_time = 0
train_time = 0
if global_step_val % train_checkpoint_interval == 0:
train_checkpointer.save(global_step=global_step_val)
if global_step_val % policy_checkpoint_interval == 0:
policy_checkpointer.save(global_step=global_step_val)
if global_step_val % rb_checkpoint_interval == 0:
rb_checkpointer.save(global_step=global_step_val)
if global_step_val % eval_interval == 0:
metric_utils.compute_summaries(
eval_metrics,
eval_py_env,
eval_py_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
)
def train():
summary_interval = 1000
summaries_flush_secs = 10
num_eval_episodes = 5
root_dir = '/tmp/tensorflow/logs/tfenv01'
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)
# maybe py_metrics?
eval_metrics = [
tf_metrics.AverageReturnMetric(buffer_size=num_eval_episodes),
tf_metrics.AverageEpisodeLengthMetric(buffer_size=num_eval_episodes),
]
environment = TradeEnvironment()
# utils.validate_py_environment(environment, episodes=5)
# Environments
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)):
train_env = tf_py_environment.TFPyEnvironment(environment)
eval_env = tf_py_environment.TFPyEnvironment(environment)
num_iterations = 50
fc_layer_params = (512, ) # ~ (17 + 1001) / 2
input_fc_layer_params = (50, )
output_fc_layer_params = (20, )
lstm_size = (30, )
initial_collect_steps = 20
collect_steps_per_iteration = 1
collect_episodes_per_iteration = 1 # the same as above
batch_size = 64
replay_buffer_capacity = 10000
train_sequence_length = 10
gamma = 0.99 # check if 1.0 works as well
target_update_tau = 0.05
target_update_period = 5
epsilon_greedy = 0.1
gradient_clipping = None
reward_scale_factor = 1.0
learning_rate = 1e-2
log_interval = 30
eval_interval = 15
# train_env.observation_spec(),
q_net = q_rnn_network.QRnnNetwork(
train_env.time_step_spec().observation,
train_env.action_spec(),
input_fc_layer_params=input_fc_layer_params,
lstm_size=lstm_size,
output_fc_layer_params=output_fc_layer_params,
)
optimizer = tf.compat.v1.train.AdamOptimizer(
learning_rate=learning_rate)
tf_agent = dqn_agent.DqnAgent(
train_env.time_step_spec(),
train_env.action_spec(),
q_network=q_net,
optimizer=optimizer,
epsilon_greedy=epsilon_greedy,
target_update_tau=target_update_tau,
target_update_period=target_update_period,
td_errors_loss_fn=dqn_agent.element_wise_squared_loss,
gamma=gamma,
reward_scale_factor=reward_scale_factor,
gradient_clipping=gradient_clipping,
debug_summaries=False,
summarize_grads_and_vars=False,
train_step_counter=global_step,
)
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
tf_agent.collect_data_spec,
batch_size=train_env.batch_size,
max_length=replay_buffer_capacity,
)
train_metrics = [
tf_metrics.NumberOfEpisodes(),
tf_metrics.EnvironmentSteps(),
tf_metrics.AverageReturnMetric(),
tf_metrics.AverageEpisodeLengthMetric(),
]
# Policy which does not allow some actions in certain states
q_policy = FilteredQPolicy(
tf_agent._time_step_spec,
tf_agent._action_spec,
q_network=tf_agent._q_network,
)
# Valid policy to pre-fill replay buffer
initial_collect_policy = DummyTradePolicy(
train_env.time_step_spec(),
train_env.action_spec(),
)
print('Initial collecting...')
initial_collect_op = dynamic_episode_driver.DynamicEpisodeDriver(
train_env,
initial_collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_episodes=initial_collect_steps,
).run()
# Main agent's policy; greedy one
policy = greedy_policy.GreedyPolicy(q_policy)
# Policy used for evaluation, the same as above
eval_policy = greedy_policy.GreedyPolicy(q_policy)
tf_agent._policy = policy
collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy(
q_policy, epsilon=tf_agent._epsilon_greedy)
# Patch random policy for epsilon greedy collect policy
filtered_random_tf_policy = FilteredRandomTFPolicy(
time_step_spec=policy.time_step_spec,
action_spec=policy.action_spec,
)
collect_policy._random_policy = filtered_random_tf_policy
tf_agent._collect_policy = collect_policy
collect_op = dynamic_episode_driver.DynamicEpisodeDriver(
train_env,
collect_policy,
observers=[replay_buffer.add_batch] + train_metrics,
num_episodes=collect_episodes_per_iteration,
).run()
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)
experience, _ = next(iterator)
loss_info = common.function(tf_agent.train)(experience=experience)
# Checkpoints
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=tf_agent.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,
)
summary_ops = []
for train_metric in train_metrics:
summary_ops.append(train_metric.tf_summaries(
train_step=global_step,
step_metrics=train_metrics[:2],
))
with eval_summary_writer.as_default(), \
tf.compat.v2.summary.record_if(True):
for eval_metric in eval_metrics:
eval_metric.tf_summaries(train_step=global_step)
init_agent_op = tf_agent.initialize()
with tf.compat.v1.Session() as sess:
# sess.run(train_summary_writer.init())
# sess.run(eval_summary_writer.init())
# Initialize the graph
# tfe.Saver().restore()
# train_checkpointer.initialize_or_restore()
# rb_checkpointer.initialize_or_restore()
# sess.run(iterator.initializer)
common.initialize_uninitialized_variables(sess)
sess.run(init_agent_op)
print('Collecting initial experience...')
sess.run(initial_collect_op)
global_step_val = sess.run(global_step)
metric_utils.compute_summaries(
eval_metrics,
eval_env,
eval_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
callback=eval_metrics_callback,
log=True,
)
collect_call = sess.make_callable(collect_op)
train_step_call = sess.make_callable([loss_info, summary_ops])
global_step_call = sess.make_callable(global_step)
timed_at_step = global_step_call()
time_acc = 0
steps_per_second_ph = tf.compat.v1.placeholder(
tf.float32, shape=(), name='steps_per_sec_ph')
steps_per_second_summary = tf.compat.v2.summary.scalar(
name='global_steps_per_sec',
data=steps_per_second_ph,
step=global_step,
)
# Train
for i in range(num_iterations):
start_time = time.time()
collect_call()
for _ in range(train_steps_per_iteration):
loss_info_value, _ = train_step_call()
time_acc += time.time() - start_time
global_step_val = global_step_call()
if global_step_val % log_inerval == 0:
print('step=%d, loss=%f',
global_step_val, loss_info_value.loss)
steps_per_sec = (global_step_val-timed_at_step) / time_acc
print('%.3f steps/sec', steps_per_sec)
sess.run(
steps_per_second_summary,
feed_dict={steps_per_second_ph: steps_per_sec},
)
timed_at_step = global_step_val
time_acc = 0
# Save checkpoints
if global_step_val % train_checkpoint_interval == 0:
train_checkpointer.save(global_step=global_step_val)
if global_step_val % policy_checkpoint_interval == 0:
policy_checkpointer.save(global_step=global_step_val)
if global_step_val % rb_checkpoint_interval == 0:
rb_checkpointer.save(global_step=global_step_val)
# Evaluate
if global_step_val % eval_interval == 0:
metric_utils.compute_summaries(
eval_metrics,
eval_env,
eval_policy,
num_episodes=num_eval_episodes,
global_step=global_step_val,
log=True,
callback=eval_metrics_callback,
)
print('Done!')
