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 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 main(argv):
del argv # Unused.
config = utility.load_config(FLAGS.logdir)
policy_layers = config.policy_layers
value_layers = config.value_layers
env = config.env(render=True, log_path=FLAGS.log_path, env_randomizer=None)
network = config.network
with tf.Session() as sess:
agent = simple_ppo_agent.SimplePPOPolicy(sess,
env,
network,
policy_layers=policy_layers,
value_layers=value_layers,
checkpoint=os.path.join(
FLAGS.logdir,
FLAGS.checkpoint))
sum_reward = 0
observation = env.reset()
while True:
action = agent.get_action([observation])
observation, reward, done, _ = env.step(action[0])
# This sleep is to prevent serial communication error on the real robot.
time.sleep(0.002)
sum_reward += reward
if done:
break
tf.logging.info("reward: {}".format(sum_reward))
def main(argv):
del argv # Unused.
config = utility.load_config(LOG_DIR)
policy_layers = config.policy_layers
value_layers = config.value_layers
env = config.env(render=True)
network = config.network
with tf.Session() as sess:
agent = simple_ppo_agent.SimplePPOPolicy(sess,
env,
network,
policy_layers=policy_layers,
value_layers=value_layers,
checkpoint=os.path.join(
LOG_DIR, CHECKPOINT))
sum_reward = 0
observation = env.reset()
while True:
action = agent.get_action([observation])
observation, reward, done, _ = env.step(action[0])
time.sleep(0.002)
sum_reward += reward
if done:
break
tf.logging.info("reward: %s", sum_reward)
def __init__(self, world, id, json_data):
self.tf_scope = 'agent'
self.graph = tf.Graph()
self.sess = tf.Session(graph=self.graph)
super().__init__(world, id, json_data)
self._build_graph(json_data)
self._init_normalizers()
return
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()