def tearDown(self):
logger.remove_all()
if tf.get_default_session() is self.sess:
self.sess.__exit__(None, None, None)
self.sess.close()
# These del are crucial to prevent ENOMEM in the CI
# b/c TensorFlow does not release memory explicitly
del self.graph
del self.sess
gc.collect()
def run_garage(env, seed, log_dir):
"""
Create garage model and training.
Replace the trpo with the algorithm you want to run.
:param env: Environment of the task.
:param seed: Random seed for the trial.
:param log_dir: Log dir path.
:return:import baselines.common.tf_util as U
"""
deterministic.set_seed(seed)
with LocalRunner() as runner:
env = TfEnv(normalize(env))
policy = GaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
)
baseline = GaussianMLPBaseline(
env_spec=env.spec,
regressor_args=dict(
hidden_sizes=(32, 32),
use_trust_region=True,
),
)
algo = TRPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
gae_lambda=0.98,
max_kl_step=0.01,
policy_ent_coeff=0.0,
plot=False,
)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, "progress.csv")
garage_logger.add_output(CsvOutput(tabular_log_file))
garage_logger.add_output(StdOutput())
garage_logger.add_output(TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(n_epochs=976, batch_size=1024)
garage_logger.remove_all()
return tabular_log_file
def run_garage(env, seed, log_dir):
'''
Create garage model and training.
Replace the ddpg with the algorithm you want to run.
:param env: Environment of the task.
:param seed: Random seed for the trial.
:param log_dir: Log dir path.
:return:
'''
deterministic.set_seed(seed)
with LocalRunner() as runner:
env = TfEnv(env)
# Set up params for ddpg
action_noise = OUStrategy(env.spec, sigma=params['sigma'])
policy = ContinuousMLPPolicy(
env_spec=env.spec,
hidden_sizes=params['policy_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=params['qf_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu)
replay_buffer = SimpleReplayBuffer(
env_spec=env.spec,
size_in_transitions=params['replay_buffer_size'],
time_horizon=params['n_rollout_steps'])
ddpg = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
policy_lr=params['policy_lr'],
qf_lr=params['qf_lr'],
target_update_tau=params['tau'],
n_train_steps=params['n_train_steps'],
discount=params['discount'],
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
policy_optimizer=tf.train.AdamOptimizer,
qf_optimizer=tf.train.AdamOptimizer)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
tensorboard_log_dir = osp.join(log_dir)
garage_logger.add_output(StdOutput())
garage_logger.add_output(CsvOutput(tabular_log_file))
garage_logger.add_output(TensorBoardOutput(tensorboard_log_dir))
runner.setup(ddpg, env)
runner.train(n_epochs=params['n_epochs'],
n_epoch_cycles=params['n_epoch_cycles'],
batch_size=params['n_rollout_steps'])
garage_logger.remove_all()
return tabular_log_file
def run_garage(env, seed, log_dir):
'''
Create garage model and training.
Replace the ppo with the algorithm you want to run.
:param env: Environment of the task.
:param seed: Random seed for the trial.
:param log_dir: Log dir path.
:return:
'''
deterministic.set_seed(seed)
with LocalRunner() as runner:
env = TfEnv(normalize(env))
policy = GaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
)
baseline = GaussianMLPBaseline(
env_spec=env.spec,
regressor_args=dict(
hidden_sizes=(64, 64),
use_trust_region=False,
optimizer=FirstOrderOptimizer,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
tf_optimizer_args=dict(learning_rate=1e-3),
),
),
)
algo = PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
policy_ent_coeff=0.0,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
tf_optimizer_args=dict(learning_rate=1e-3),
),
plot=False,
)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
garage_logger.add_output(StdOutput())
garage_logger.add_output(CsvOutput(tabular_log_file))
garage_logger.add_output(TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(n_epochs=488, batch_size=2048)
garage_logger.remove_all()
return tabular_log_file
def run_garage(env, seed, log_dir):
"""
Create garage model and training.
Replace the ppo with the algorithm you want to run.
:param env: Environment of the task.
:param seed: Random seed for the trial.
:param log_dir: Log dir path.
:return:
"""
deterministic.set_seed(seed)
env.reset()
with LocalRunner() as runner:
env = TfEnv(env)
action_noise = OUStrategy(env.spec, sigma=params['sigma'])
policy = ContinuousMLPPolicy(
env_spec=env.spec,
hidden_sizes=params['policy_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh,
input_include_goal=True,
)
qf = ContinuousMLPQFunction(
env_spec=env.spec,
hidden_sizes=params['qf_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
input_include_goal=True,
)
replay_buffer = HerReplayBuffer(
env_spec=env.spec,
size_in_transitions=params['replay_buffer_size'],
time_horizon=params['n_rollout_steps'],
replay_k=0.4,
reward_fun=env.compute_reward,
)
algo = DDPG(
env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
policy_lr=params['policy_lr'],
qf_lr=params['qf_lr'],
plot=False,
target_update_tau=params['tau'],
n_epochs=params['n_epochs'],
n_epoch_cycles=params['n_epoch_cycles'],
n_train_steps=params['n_train_steps'],
discount=params['discount'],
exploration_strategy=action_noise,
policy_optimizer=tf.train.AdamOptimizer,
qf_optimizer=tf.train.AdamOptimizer,
buffer_batch_size=256,
input_include_goal=True,
)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
logger.add_output(StdOutput())
logger.add_output(CsvOutput(tabular_log_file))
logger.add_output(TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(n_epochs=params['n_epochs'],
n_epoch_cycles=params['n_epoch_cycles'],
batch_size=params['n_rollout_steps'])
logger.remove_all()
return tabular_log_file
def run_experiment(argv):
default_log_dir = config.GARAGE_LOG_DIR
now = datetime.datetime.now(dateutil.tz.tzlocal())
# avoid name clashes when running distributed jobs
rand_id = str(uuid.uuid4())[:5]
timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z')
default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id)
parser = argparse.ArgumentParser()
parser.add_argument(
'--n_parallel',
type=int,
default=1,
help=('Number of parallel workers to perform rollouts. '
"0 => don't start any workers"))
parser.add_argument(
'--exp_name',
type=str,
default=default_exp_name,
help='Name of the experiment.')
parser.add_argument(
'--log_dir',
type=str,
default=None,
help='Path to save the log and iteration snapshot.')
parser.add_argument(
'--snapshot_mode',
type=str,
default='all',
help='Mode to save the snapshot. Can be either "all" '
'(all iterations will be saved), "last" (only '
'the last iteration will be saved), "gap" (every'
'`snapshot_gap` iterations are saved), or "none" '
'(do not save snapshots)')
parser.add_argument(
'--snapshot_gap',
type=int,
default=1,
help='Gap between snapshot iterations.')
parser.add_argument(
'--tabular_log_file',
type=str,
default='progress.csv',
help='Name of the tabular log file (in csv).')
parser.add_argument(
'--text_log_file',
type=str,
default='debug.log',
help='Name of the text log file (in pure text).')
parser.add_argument(
'--tensorboard_step_key',
type=str,
default=None,
help='Name of the step key in tensorboard_summary.')
parser.add_argument(
'--params_log_file',
type=str,
default='params.json',
help='Name of the parameter log file (in json).')
parser.add_argument(
'--variant_log_file',
type=str,
default='variant.json',
help='Name of the variant log file (in json).')
parser.add_argument(
'--resume_from_dir',
type=str,
default=None,
help='Directory of the pickle file to resume experiment from.')
parser.add_argument(
'--resume_epoch',
type=str,
default=None,
help='Index of iteration to restore from. '
'Can be "first", "last" or a number. '
'Not applicable when snapshot_mode="last"')
parser.add_argument(
'--plot',
type=ast.literal_eval,
default=False,
help='Whether to plot the iteration results')
parser.add_argument(
'--log_tabular_only',
type=ast.literal_eval,
default=False,
help='Print only the tabular log information (in a horizontal format)')
parser.add_argument('--seed', type=int, help='Random seed for numpy')
parser.add_argument(
'--args_data', type=str, help='Pickled data for objects')
parser.add_argument(
'--variant_data',
type=str,
help='Pickled data for variant configuration')
parser.add_argument(
'--use_cloudpickle', type=ast.literal_eval, default=False)
args = parser.parse_args(argv[1:])
if args.seed is not None:
deterministic.set_seed(args.seed)
# SIGINT is blocked for all processes created in parallel_sampler to avoid
# the creation of sleeping and zombie processes.
#
# If the user interrupts run_experiment, there's a chance some processes
# won't die due to a dead lock condition where one of the children in the
# parallel sampler exits without releasing a lock once after it catches
# SIGINT.
#
# Later the parent tries to acquire the same lock to proceed with his
# cleanup, but it remains sleeping waiting for the lock to be released.
# In the meantime, all the process in parallel sampler remain in the zombie
# state since the parent cannot proceed with their clean up.
with mask_signals([signal.SIGINT]):
if args.n_parallel > 0:
parallel_sampler.initialize(n_parallel=args.n_parallel)
if args.seed is not None:
parallel_sampler.set_seed(args.seed)
if not args.plot:
garage.plotter.Plotter.disable()
garage.tf.plotter.Plotter.disable()
if args.log_dir is None:
if args.resume_from_dir is None:
log_dir = osp.join(default_log_dir, args.exp_name)
else:
log_dir = args.resume_from_dir
else:
log_dir = args.log_dir
tabular_log_file = osp.join(log_dir, args.tabular_log_file)
text_log_file = osp.join(log_dir, args.text_log_file)
params_log_file = osp.join(log_dir, args.params_log_file)
if args.variant_data is not None:
variant_data = pickle.loads(base64.b64decode(args.variant_data))
variant_log_file = osp.join(log_dir, args.variant_log_file)
dump_variant(variant_log_file, variant_data)
else:
variant_data = None
if not args.use_cloudpickle:
log_parameters(params_log_file, args)
logger.add_output(TextOutput(text_log_file))
logger.add_output(CsvOutput(tabular_log_file))
logger.add_output(TensorBoardOutput(log_dir))
logger.add_output(StdOutput())
prev_snapshot_dir = snapshotter.snapshot_dir
prev_mode = snapshotter.snapshot_mode
snapshotter.snapshot_dir = log_dir
snapshotter.snapshot_mode = args.snapshot_mode
snapshotter.snapshot_gap = args.snapshot_gap
logger.push_prefix('[%s] ' % args.exp_name)
if args.resume_from_dir is not None:
with LocalRunner() as runner:
runner.restore(args.resume_from_dir, from_epoch=args.resume_epoch)
runner.resume()
else:
# read from stdin
if args.use_cloudpickle:
import cloudpickle
method_call = cloudpickle.loads(base64.b64decode(args.args_data))
try:
method_call(variant_data)
except BaseException:
children = garage.plotter.Plotter.get_plotters()
children += garage.tf.plotter.Plotter.get_plotters()
if args.n_parallel > 0:
children += [parallel_sampler]
child_proc_shutdown(children)
raise
else:
data = pickle.loads(base64.b64decode(args.args_data))
maybe_iter = concretize(data)
if is_iterable(maybe_iter):
for _ in maybe_iter:
pass
snapshotter.snapshot_mode = prev_mode
snapshotter.snapshot_dir = prev_snapshot_dir
logger.remove_all()
logger.pop_prefix()
def tearDown(self):
logger.remove_all()
def tearDownClass(cls):
snapshotter.snapshot_dir = cls.prev_log_dir
snapshotter.snapshot_mode = cls.prev_mode
logger.remove_all()
cls.log_dir.cleanup()