def main():
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('-e', '--env_id', required=True)
parser.add_argument('-p', '--predictor', required=True)
parser.add_argument('-n', '--name', required=True)
parser.add_argument('-s', '--seed', default=1, type=int)
parser.add_argument('-w', '--workers', default=4, type=int)
parser.add_argument('-l', '--n_labels', default=None, type=int)
parser.add_argument('-L', '--pretrain_labels', default=None, type=int)
parser.add_argument('-t', '--num_timesteps', default=5e6, type=int)
parser.add_argument('-a', '--agent', default="parallel_trpo", type=str)
parser.add_argument('-i', '--pretrain_iters', default=10000, type=int)
parser.add_argument('-V', '--no_videos', action="store_true")
parser.add_argument('-x', '--human_labels', default=1000, type=int)
args = parser.parse_args()
print("Setting things up...")
env_id = args.env_id
run_name = "%s/%s-%s" % (env_id, args.name, int(time()))
summary_writer = make_summary_writer(run_name)
env = make_with_torque_removed(env_id)
num_timesteps = int(args.num_timesteps)
experiment_name = slugify(args.name)
if args.predictor == "rl":
predictor = TraditionalRLRewardPredictor(summary_writer)
else:
agent_logger = AgentLogger(summary_writer)
pretrain_labels = args.pretrain_labels if args.pretrain_labels else args.n_labels // 4
if args.n_labels:
label_schedule = LabelAnnealer(agent_logger,
final_timesteps=num_timesteps,
final_labels=args.n_labels,
pretrain_labels=pretrain_labels)
else:
print(
"No label limit given. We will request one label every few seconds."
)
label_schedule = ConstantLabelSchedule(
pretrain_labels=pretrain_labels)
if args.predictor == "synth":
comparison_collector = SyntheticComparisonCollector(
run_name, args.human_labels)
elif args.predictor == "human":
bucket = os.environ.get('RL_TEACHER_GCS_BUCKET')
assert bucket and bucket.startswith(
"gs://"
), "env variable RL_TEACHER_GCS_BUCKET must start with gs://"
comparison_collector = HumanComparisonCollector(
env_id, experiment_name=experiment_name)
else:
raise ValueError("Bad value for --predictor: %s" % args.predictor)
predictor = ComparisonRewardPredictor(
env,
summary_writer,
comparison_collector=comparison_collector,
agent_logger=agent_logger,
label_schedule=label_schedule,
)
print(
"Starting random rollouts to generate pretraining segments. No learning will take place..."
)
pretrain_segments = segments_from_rand_rollout(
env_id,
make_with_torque_removed,
n_desired_segments=pretrain_labels * 2,
clip_length_in_seconds=CLIP_LENGTH,
workers=args.workers)
for i in range(
pretrain_labels): # Turn our random segments into comparisons
comparison_collector.add_segment_pair(
pretrain_segments[i], pretrain_segments[i + pretrain_labels])
# Sleep until the human has labeled most of the pretraining comparisons
while len(comparison_collector.labeled_comparisons) < int(
pretrain_labels * 0.75):
comparison_collector.label_unlabeled_comparisons()
if args.predictor == "synth":
print("%s synthetic labels generated... " %
(len(comparison_collector.labeled_comparisons)))
elif args.predictor == "human":
print(
"%s/%s comparisons labeled. Please add labels w/ the human-feedback-api. Sleeping... "
% (len(comparison_collector.labeled_comparisons),
pretrain_labels))
sleep(5)
# Start the actual training
for i in range(args.pretrain_iters):
predictor.train_predictor() # Train on pretraining labels
if i % 100 == 0:
print("%s/%s predictor pretraining iters... " %
(i, args.pretrain_iters))
# Wrap the predictor to capture videos every so often:
if not args.no_videos:
predictor = SegmentVideoRecorder(predictor,
env,
save_dir=osp.join(
'/tmp/rl_teacher_vids', run_name))
# We use a vanilla agent from openai/baselines that contains a single change that blinds it to the true reward
# The single changed section is in `rl_teacher/agent/trpo/core.py`
print("Starting joint training of predictor and agent")
if args.agent == "parallel_trpo":
train_parallel_trpo(
env_id=env_id,
make_env=make_with_torque_removed,
predictor=predictor,
summary_writer=summary_writer,
workers=args.workers,
runtime=(num_timesteps / 1000),
max_timesteps_per_episode=get_timesteps_per_episode(env),
timesteps_per_batch=8000,
max_kl=0.001,
seed=args.seed,
)
elif args.agent == "pposgd_mpi":
def make_env():
return make_with_torque_removed(env_id)
train_pposgd_mpi(make_env,
num_timesteps=num_timesteps,
seed=args.seed,
predictor=predictor)
else:
raise ValueError("%s is not a valid choice for args.agent" %
args.agent)
def main():
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('-e', '--env_id', required=True)
parser.add_argument('-p', '--predictor', required=True)
parser.add_argument('-n', '--name', required=True)
parser.add_argument('-s', '--seed', default=1, type=int)
parser.add_argument('-w', '--workers', default=4, type=int)
parser.add_argument('-l', '--n_labels', default=None, type=int)
parser.add_argument('-L', '--pretrain_labels', default=None, type=int)
parser.add_argument('-t', '--num_timesteps', default=5e6, type=int)
parser.add_argument('-a', '--agent', default="parallel_trpo", type=str)
parser.add_argument('-i', '--pretrain_iters', default=10000, type=int)
parser.add_argument('-V', '--no_videos', action="store_true")
args = parser.parse_args()
env_id = args.env_id
run_name = "%s/%s-%s" % (env_id, args.name, int(time()))
summary_writer = make_summary_writer(run_name)
num_timesteps = int(args.num_timesteps)
experiment_name = slugify(args.name)
##make torcs envs
# envs = []
# for aidx in range(args.workers):
# agent = AgentTorcs2(aidx, bots=['scr_server'], track='road/g-track-1', text_mode=False, laps=3,
# torcsIdxOffset=0, screen_capture=True)
# # agent = AgentTorcs2(aidx, bots=['scr_server', 'olethros', 'berniw', 'bt', 'damned'], track='road/g-track-1', text_mode=True)
# agent.reset()
# envs.append(agent)
#
#
if args.predictor == "rl":
predictor = TraditionalRLRewardPredictor(summary_writer)
else:
agent_logger = AgentLogger(summary_writer)
if args.predictor == "synth":
comparison_collector = SyntheticComparisonCollector()
elif args.predictor == "human":
# bucket = os.environ.get('RL_TEACHER_GCS_BUCKET')
# assert bucket and bucket.startswith("gs://"), "env variable RL_TEACHER_GCS_BUCKET must start with gs://"
comparison_collector = HumanComparisonCollector( experiment_name=experiment_name)
else:
raise ValueError("Bad value for --predictor: %s" % args.predictor)
pretrain_labels = args.pretrain_labels if args.pretrain_labels else args.n_labels // 4
if args.n_labels:
label_schedule = LabelAnnealer(
agent_logger,
final_timesteps=num_timesteps,
final_labels=args.n_labels,
pretrain_labels=pretrain_labels)
else:
print("No label limit given. We will request one label every few seconds")
label_schedule = ConstantLabelSchedule(pretrain_labels=pretrain_labels)
# logger.info("frames = {}".format('start !!!!!!!!!!!!!'))
print("Starting random rollouts to generate pretraining segments. No learning will take place...")
pretrain_segments = segments_from_rand_rollout( n_desired_segments=pretrain_labels * 2,
clip_length_in_seconds=CLIP_LENGTH, workers=args.workers)
pretrain_segments.sort(key= lambda d :d['maxdistance'])
for i in range(pretrain_labels): # Turn our random segments into comparisons
comparison_collector.add_segment_pair(pretrain_segments[i], pretrain_segments[i + 1])
# Sleep until the human has labeled most of the pretraining comparisons
while len(comparison_collector.labeled_comparisons) < int(pretrain_labels * 0.75):
comparison_collector.label_unlabeled_comparisons()
if args.predictor == "synth":
print("%s synthetic labels generated... " % (len(comparison_collector.labeled_comparisons)))
elif args.predictor == "human":
print("%s/%s comparisons labeled. Please add labels w/ the human-feedback-api. Sleeping... " % (
len(comparison_collector.labeled_comparisons), pretrain_labels))
sleep(5)
# Start the actual training
predictor = ComparisonRewardPredictor(
summary_writer,
comparison_collector=comparison_collector,
agent_logger=agent_logger,
label_schedule=label_schedule,
)
for i in range(args.pretrain_iters):
predictor.train_predictor() # Train on pretraining labels
if i % 100 == 0:
print("%s/%s predictor pretraining iters... " % (i, args.pretrain_iters))
# Wrap the predictor to capture videos every so often:
if not args.no_videos:
predictor = SegmentVideoRecorder(predictor, save_dir=osp.join('/tmp/rl_teacher_vids', run_name))
# We use a vanilla agent from openai/baselines that contains a single change that blinds it to the true reward
# The single changed section is in `rl_teacher/agent/trpo/core.py`
print("Starting joint training of predictor and agent")
if args.agent == "parallel_trpo":
train_parallel_trpo(
predictor=predictor,
summary_writer=summary_writer,
workers=args.workers,
runtime=(num_timesteps / 1000),
max_timesteps_per_episode=10000,
timesteps_per_batch=8000,
max_kl=0.001,
seed=args.seed,
)
elif args.agent == "pposgd_mpi":
pass
# train_pposgd_mpi(num_timesteps=num_timesteps, seed=args.seed, predictor=predictor)
else:
raise ValueError("%s is not a valid choice for args.agent" % args.agent)
def main():
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('-e', '--env_id', default="ShortHopper-v1", type=str)
parser.add_argument('-p', '--predictor', default="human", type=str)
parser.add_argument('-n', '--name', default="human-175-hopper", type=str)
parser.add_argument('-s', '--seed', default=6, type=int)
parser.add_argument('-w', '--workers', default=4, type=int)
parser.add_argument('-l', '--n_labels', default=None, type=int)
parser.add_argument('-L', '--pretrain_labels', default=20, type=int)
parser.add_argument('-t', '--num_timesteps', default=5e6, type=int)
parser.add_argument('-a', '--agent', default="pposgd_mpi", type=str)
parser.add_argument('-i', '--pretrain_iters', default=1, type=int)
parser.add_argument('-V', '--no_videos', action="store_true")
parser.add_argument('--log_path',
help='Directory to save learning curve data.',
default='tmp/openaiTest',
type=str)
args = parser.parse_args()
print("Setting things up...")
env_id = args.env_id
run_name = "%s/%s-%s" % (env_id, args.name, int(time()))
summary_writer = make_summary_writer(run_name)
env = make_with_torque_removed(env_id)
num_timesteps = int(args.num_timesteps)
experiment_name = slugify(args.name)
if args.predictor == "rl":
predictor = TraditionalRLRewardPredictor(summary_writer)
else:
agent_logger = AgentLogger(summary_writer)
pretrain_labels = args.pretrain_labels if args.pretrain_labels else args.n_labels // 4
#online and offline
if args.n_labels:
label_schedule = LabelAnnealer(agent_logger,
final_timesteps=num_timesteps,
final_labels=args.n_labels,
pretrain_labels=pretrain_labels)
else:
print(
"No label limit given. We will request one label every few seconds."
)
label_schedule = ConstantLabelSchedule(
pretrain_labels=pretrain_labels)
if args.predictor == "synth":
comparison_collector = SyntheticComparisonCollector()
elif args.predictor == "human":
bucket = os.environ.get('RL_TEACHER_GCS_BUCKET')
bucket = "gs://rl-teacher-preference"
#assert bucket and bucket.startswith("gs://"), "env variable RL_TEACHER_GCS_BUCKET must start with gs://"
comparison_collector = HumanComparisonCollector(
env_id, experiment_name=experiment_name)
else:
raise ValueError("Bad value for --predictor: %s" % args.predictor)
predictor = ComparisonRewardPredictor(
env,
summary_writer,
comparison_collector=comparison_collector,
agent_logger=agent_logger,
label_schedule=label_schedule,
)
# print("Starting random rollouts to generate pretraining segments. No learning will take place...")
# pretrain_segments = segments_from_rand_rollout(
# env_id, make_with_torque_removed, n_desired_segments=pretrain_labels * 2,
# clip_length_in_seconds=CLIP_LENGTH, workers=args.workers)
# for i in range(pretrain_labels): # Turn our random segments into comparisons
# comparison_collector.add_segment_pair(pretrain_segments[i], pretrain_segments[i + pretrain_labels])
#
# # Sleep until the human has labeled most of the pretraining comparisons
# while len(comparison_collector.labeled_comparisons) < int(pretrain_labels * 0.75):
# comparison_collector.label_unlabeled_comparisons()
# if args.predictor == "synth":
# print("%s synthetic labels generated... " % (len(comparison_collector.labeled_comparisons)))
# elif args.predictor == "human":
# print("%s/%s comparisons labeled. Please add labels w/ the human-feedback-api. Sleeping... " % (
# len(comparison_collector.labeled_comparisons), pretrain_labels))
# sleep(5)
#
# # Start the actual training
#
# for i in range(args.pretrain_iters):
# predictor.train_predictor() # Train on pretraining labels
# if i % 10 == 0:
# print("%s/%s predictor pretraining iters... " % (i, args.pretrain_iters))
#saver = tf.train.Saver(max_to_keep=5)
#save_path = saver.save(sess, "/tmp/GAN/GAN_preference_based_model.ckpt")
#print("Model saved in path: %s" % save_path)
# Wrap the predictor to capture videos every so often:
if not args.no_videos:
predictor = SegmentVideoRecorder(predictor,
env,
save_dir=osp.join(
'/tmp/rl_teacher_vids', run_name))
# We use a vanilla agent from openai/baselines that contains a single change that blinds it to the true reward
# The single changed section is in `rl_teacher/agent/trpo/core.py`
print("Starting joint training of predictor and agent")
if args.agent == "parallel_trpo":
train_parallel_trpo(
env_id=env_id,
make_env=make_with_torque_removed,
predictor=predictor,
summary_writer=summary_writer,
workers=args.workers,
runtime=(num_timesteps / 1000),
max_timesteps_per_episode=get_timesteps_per_episode(env),
timesteps_per_batch=8000,
max_kl=0.001,
seed=args.seed,
)
elif args.agent == "pposgd_mpi":
def make_env():
return make_with_torque_removed(env_id)
try:
from mpi4py import MPI
except ImportError:
MPI = None
def configure_logger(log_path, **kwargs):
if log_path is not None:
logger.configure(log_path)
else:
logger.configure(**kwargs)
if MPI is None or MPI.COMM_WORLD.Get_rank() == 0:
rank = 0
configure_logger(args.log_path)
else:
rank = MPI.COMM_WORLD.Get_rank()
configure_logger(args.log_path, format_strs=[])
train_pposgd_mpi(make_env,
num_timesteps=num_timesteps,
seed=args.seed,
predictor=predictor)
else:
raise ValueError("%s is not a valid choice for args.agent" %
args.agent)
def __init__(self, sess, env, brain_name, trainer_parameters, training, seed,num_timesteps,num_labels,pretrain_labels):
"""
Responsible for collecting experiences and training PPO model.
:param sess: Tensorflow session.
:param env: The UnityEnvironment.
:param trainer_parameters: The parameters for the trainer (dictionary).
:param training: Whether the trainer is set for training.
:param num_timesteps: timesteps collect segment.
:param num_labels: .
"""
self.param_keys = ['batch_size', 'beta', 'buffer_size', 'epsilon', 'gamma', 'hidden_units', 'lambd',
'learning_rate', 'max_steps', 'normalize', 'num_epoch', 'num_layers',
'time_horizon', 'sequence_length', 'summary_freq', 'use_recurrent',
'graph_scope', 'summary_path', 'memory_size', 'use_curiosity', 'curiosity_strength',
'curiosity_enc_size']
for k in self.param_keys:
if k not in trainer_parameters:
raise UnityTrainerException("The hyperparameter {0} could not be found for the PPO trainer of "
"brain {1}.".format(k, brain_name))
super(PPOTrainer, self).__init__(sess, env, brain_name, trainer_parameters, training)
self.use_recurrent = trainer_parameters["use_recurrent"]
self.use_curiosity = bool(trainer_parameters['use_curiosity'])
self.sequence_length = 1
self.step = 0
self.has_updated = False
self.m_size = None
if self.use_recurrent:
self.m_size = trainer_parameters["memory_size"]
self.sequence_length = trainer_parameters["sequence_length"]
if self.m_size == 0:
raise UnityTrainerException("The memory size for brain {0} is 0 even though the trainer uses recurrent."
.format(brain_name))
elif self.m_size % 4 != 0:
raise UnityTrainerException("The memory size for brain {0} is {1} but it must be divisible by 4."
.format(brain_name, self.m_size))
self.variable_scope = trainer_parameters['graph_scope']
with tf.variable_scope(self.variable_scope):
tf.set_random_seed(seed)
self.model = PPOModel(env.brains[brain_name],
lr=float(trainer_parameters['learning_rate']),
h_size=int(trainer_parameters['hidden_units']),
epsilon=float(trainer_parameters['epsilon']),
beta=float(trainer_parameters['beta']),
max_step=float(trainer_parameters['max_steps']),
normalize=trainer_parameters['normalize'],
use_recurrent=trainer_parameters['use_recurrent'],
num_layers=int(trainer_parameters['num_layers']),
m_size=self.m_size,
use_curiosity=bool(trainer_parameters['use_curiosity']),
curiosity_strength=float(trainer_parameters['curiosity_strength']),
curiosity_enc_size=float(trainer_parameters['curiosity_enc_size']))
stats = {'cumulative_reward': [], 'episode_length': [], 'value_estimate': [],
'entropy': [], 'value_loss': [], 'policy_loss': [], 'learning_rate': []}
if self.use_curiosity:
stats['forward_loss'] = []
stats['inverse_loss'] = []
stats['intrinsic_reward'] = []
self.intrinsic_rewards = {}
self.stats = stats
self.training_buffer = Buffer()
self.cumulative_rewards = {}
self.episode_steps = {}
self.is_continuous_action = (env.brains[brain_name].vector_action_space_type == "continuous")
self.is_continuous_observation = (env.brains[brain_name].vector_observation_space_type == "continuous")
self.use_visual_obs = (env.brains[brain_name].number_visual_observations > 0)
self.use_vector_obs = (env.brains[brain_name].vector_observation_space_size > 0)
self.summary_path = trainer_parameters['summary_path']
if not os.path.exists(self.summary_path):
os.makedirs(self.summary_path)
self.summary_writer = tf.summary.FileWriter(self.summary_path)
self.inference_run_list = [self.model.output, self.model.all_probs, self.model.value,
self.model.entropy, self.model.learning_rate]
if self.is_continuous_action:
self.inference_run_list.append(self.model.output_pre)
if self.use_recurrent:
self.inference_run_list.extend([self.model.memory_out])
if (self.is_training and self.is_continuous_observation and
self.use_vector_obs and self.trainer_parameters['normalize']):
self.inference_run_list.extend([self.model.update_mean, self.model.update_variance])
label_schedule = LabelAnnealer(
logger,
final_timesteps=num_timesteps,
final_labels=num_labels,
pretrain_labels=pretrain_labels
)
comparison_collector = HumanComparisonCollector(experiment_name=brain_name)
self.predictor = ComparisonRewardPredictor(
self.brain,
self.summary_writer,
comparison_collector=comparison_collector,
agent_logger=logger,
label_schedule=label_schedule,
clip_length= CLIP_LENGTH)
