def visualize(logdir, outdir, num_agents, num_episodes, checkpoint=None, env_processes=True):
"""Recover checkpoint and render videos from it.
Args:
logdir: Logging directory of the trained algorithm.
outdir: Directory to store rendered videos in.
num_agents: Number of environments to simulate in parallel.
num_episodes: Total number of episodes to simulate.
checkpoint: Checkpoint name to load; defaults to most recent.
env_processes: Whether to step environments in separate processes.
"""
config = utility.load_config(logdir)
with config.unlocked:
config.network = functools.partial(utility.define_network, config.network, config)
config.policy_optimizer = getattr(tf.train, config.policy_optimizer)
config.value_optimizer = getattr(tf.train, config.value_optimizer)
with tf.device('/cpu:0'):
batch_env = utility.define_batch_env(lambda: _create_environment(config, outdir), num_agents,
env_processes)
graph = utility.define_simulation_graph(batch_env, config.algorithm, config)
total_steps = num_episodes * config.max_length
loop = _define_loop(graph, total_steps)
saver = utility.define_saver(exclude=(r'.*_temporary/.*', r'global_step'))
sess_config = tf.ConfigProto(allow_soft_placement=True)
sess_config.gpu_options.allow_growth = True
with tf.Session(config=sess_config) as sess:
utility.initialize_variables(sess, saver, config.logdir, checkpoint, resume=True)
for unused_score in loop.run(sess, saver, total_steps):
pass
batch_env.close()
def train(config, env_processes):
"""Training and evaluation entry point yielding scores.
Resolves some configuration attributes, creates environments, graph, and
training loop. By default, assigns all operations to the CPU.
Args:
config: Object providing configurations via attributes.
env_processes: Whether to step environments in separate processes.
Yields:
Evaluation scores.
"""
tf.reset_default_graph()
with config.unlocked:
config.network = functools.partial(
utility.define_network, config.network, config)
config.policy_optimizer = getattr(tf.train, config.policy_optimizer)
config.value_optimizer = getattr(tf.train, config.value_optimizer)
if config.update_every % config.num_agents:
tf.logging.warn('Number of agents should divide episodes per update.')
with tf.device('/cpu:0'):
batch_env = utility.define_batch_env(
lambda: _create_environment(config),
config.num_agents, env_processes)
graph = utility.define_simulation_graph(
batch_env, config.algorithm, config)
loop = _define_loop(
graph, config.logdir,
config.update_every * config.max_length,
config.eval_episodes * config.max_length)
total_steps = int(
config.steps / config.update_every *
(config.update_every + config.eval_episodes))
# Exclude episode related variables since the Python state of environments is
# not checkpointed and thus new episodes start after resuming.
saver = utility.define_saver(exclude=(r'.*_temporary/.*',))
sess_config = tf.ConfigProto(allow_soft_placement=True)
sess_config.gpu_options.allow_growth = True
with tf.Session(config=sess_config) as sess:
utility.initialize_variables(sess, saver, config.logdir)
for score in loop.run(sess, saver, total_steps):
yield score
batch_env.close()
def train(config, env_processes):
"""Training and evaluation entry point yielding scores.
Resolves some configuration attributes, creates environments, graph, and
training loop. By default, assigns all operations to the CPU.
Args:
config: Object providing configurations via attributes.
env_processes: Whether to step environments in separate processes.
Yields:
Evaluation scores.
"""
tf.reset_default_graph()
with config.unlocked:
config.network = functools.partial(utility.define_network,
config.network, config)
config.policy_optimizer = getattr(tf.train, config.policy_optimizer)
config.value_optimizer = getattr(tf.train, config.value_optimizer)
if config.update_every % config.num_agents:
tf.logging.warn(
'Number of algorithms should divide episodes per update.')
with tf.device('/cpu:0'):
batch_env = utility.define_batch_env(
lambda: _create_environment(config), config.num_agents,
env_processes)
graph = utility.define_simulation_graph(batch_env, config.algorithm,
config)
loop = _define_loop(graph, config.logdir,
config.update_every * config.max_length,
config.eval_episodes * config.max_length)
total_steps = int(config.steps / config.update_every *
(config.update_every + config.eval_episodes))
# Exclude episode related variables since the Python state of environments is
# not checkpointed and thus new episodes start after resuming.
saver = utility.define_saver(exclude=(r'.*_temporary/.*', ))
sess_config = tf.ConfigProto(allow_soft_placement=True)
sess_config.gpu_options.allow_growth = True
with tf.Session(config=sess_config) as sess:
utility.initialize_variables(sess, saver, config.logdir)
for score in loop.run(sess, saver, total_steps):
yield score
batch_env.close()
def visualize(logdir,
outdir,
num_agents,
num_episodes,
checkpoint=None,
env_processes=True):
"""Recover checkpoint and render videos from it.
Args:
logdir: Logging directory of the trained algorithm.
outdir: Directory to store rendered videos in.
num_agents: Number of environments to simulate in parallel.
num_episodes: Total number of episodes to simulate.
checkpoint: Checkpoint name to load; defaults to most recent.
env_processes: Whether to step environments in separate processes.
"""
config = utility.load_config(logdir)
with config.unlocked:
config.network = functools.partial(utility.define_network,
config.network, config)
config.policy_optimizer = getattr(tf.train, config.policy_optimizer)
config.value_optimizer = getattr(tf.train, config.value_optimizer)
with tf.device('/cpu:0'):
batch_env = utility.define_batch_env(
lambda: _create_environment(config, outdir), num_agents,
env_processes)
graph = utility.define_simulation_graph(batch_env, config.algorithm,
config)
total_steps = num_episodes * config.max_length
loop = _define_loop(graph, total_steps)
saver = utility.define_saver(exclude=(r'.*_temporary/.*', r'global_step'))
sess_config = tf.ConfigProto(allow_soft_placement=True)
sess_config.gpu_options.allow_growth = True
with tf.Session(config=sess_config) as sess:
utility.initialize_variables(sess,
saver,
config.logdir,
checkpoint,
resume=True)
for unused_score in loop.run(sess, saver, total_steps):
pass
batch_env.close()