def run_metarl(env, seed, log_dir):
'''
Create metarl 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 LocalTFRunner(snapshot_config) as runner:
env = TfEnv(normalize(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,
steps_per_epoch=params['steps_per_epoch'],
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)
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(tensorboard_log_dir))
runner.setup(ddpg, env)
runner.train(n_epochs=params['n_epochs'],
batch_size=params['n_rollout_steps'])
dowel_logger.remove_all()
return tabular_log_file
def ppo_cmb(env, seed, log_dir):
"""Create test continuous mlp baseline on ppo.
Args:
env (gym_env): Environment of the task.
seed (int): Random seed for the trial.
log_dir (str): Log dir path.
Returns:
str: training results in csv format.
"""
deterministic.set_seed(seed)
config = tf.compat.v1.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=num_proc,
inter_op_parallelism_threads=num_proc)
sess = tf.compat.v1.Session(config=config)
with LocalTFRunner(snapshot_config, sess=sess,
max_cpus=num_proc) as runner:
env = TfEnv(normalize(env))
policy = GaussianLSTMPolicy(
env_spec=env.spec,
hidden_dim=policy_params['policy_hidden_sizes'],
hidden_nonlinearity=policy_params['hidden_nonlinearity'],
)
baseline = ContinuousMLPBaseline(
env_spec=env.spec,
regressor_args=baseline_params['regressor_args'],
)
algo = PPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=algo_params['max_path_length'],
discount=algo_params['discount'],
gae_lambda=algo_params['gae_lambda'],
lr_clip_range=algo_params['lr_clip_range'],
entropy_method=algo_params['entropy_method'],
policy_ent_coeff=algo_params['policy_ent_coeff'],
optimizer_args=algo_params['optimizer_args'],
center_adv=algo_params['center_adv'],
stop_entropy_gradient=True)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(algo,
env,
sampler_args=dict(n_envs=algo_params['n_envs']))
runner.train(n_epochs=algo_params['n_epochs'],
batch_size=algo_params['n_rollout_steps'])
dowel_logger.remove_all()
return tabular_log_file
def run_garage_pytorch(env, seed, log_dir):
"""Create garage PyTorch VPG model and training.
Args:
env (dict): Environment of the task.
seed (int): Random positive integer for the trial.
log_dir (str): Log dir path.
Returns:
str: Path to output csv file
"""
env = TfEnv(normalize(env))
deterministic.set_seed(seed)
runner = LocalRunner(snapshot_config)
policy = PyTorch_GMP(env.spec,
hidden_sizes=hyper_parameters['hidden_sizes'],
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
value_function = GaussianMLPValueFunction(env_spec=env.spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
policy_optimizer = OptimizerWrapper((torch.optim.Adam, dict(lr=2.5e-4)),
policy,
max_optimization_epochs=10,
minibatch_size=64)
vf_optimizer = OptimizerWrapper((torch.optim.Adam, dict(lr=2.5e-4)),
value_function,
max_optimization_epochs=10,
minibatch_size=64)
algo = PyTorch_VPG(env_spec=env.spec,
policy=policy,
value_function=value_function,
policy_optimizer=policy_optimizer,
vf_optimizer=vf_optimizer,
max_path_length=hyper_parameters['max_path_length'],
discount=hyper_parameters['discount'],
center_adv=hyper_parameters['center_adv'])
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(n_epochs=hyper_parameters['n_epochs'],
batch_size=hyper_parameters['batch_size'])
dowel_logger.remove_all()
return tabular_log_file
def run_metarl(env, seed, log_dir):
'''
Create metarl 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)
runner = LocalRunner(snapshot_config)
# Set up params for ddpg
policy = TanhGaussianMLPPolicy2(env_spec=env.spec,
hidden_sizes=params['policy_hidden_sizes'],
hidden_nonlinearity=nn.ReLU,
output_nonlinearity=None)
qf1 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=params['qf_hidden_sizes'],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=params['qf_hidden_sizes'],
hidden_nonlinearity=F.relu)
replay_buffer = SACReplayBuffer(env_spec=env.spec,
max_size=params['replay_buffer_size'])
sampler_args = {
'agent': policy,
'max_path_length': 1000,
}
sac = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
gradient_steps_per_itr=params['gradient_steps_per_itr'],
replay_buffer=replay_buffer,
buffer_batch_size=params['buffer_batch_size'])
# 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)
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(tensorboard_log_dir))
runner.setup(algo=sac,
env=env,
sampler_cls=SimpleSampler,
sampler_args=sampler_args)
runner.train(n_epochs=params['n_epochs'],
batch_size=params['gradient_steps_per_itr'])
dowel_logger.remove_all()
return tabular_log_file
def _make_context(self, *args, **kwargs):
"""Make a context from the template information and variant args.
Currently, all arguments should be keyword arguments.
Args:
args (list): Should be empty.
kwargs (dict): Keyword arguments for the wrapped function. Will be
logged to `variant.json`
Returns:
ExperimentContext: The created experiment context.
Raises:
ValueError: If args is not empty.
"""
if args:
raise ValueError('garage.experiment currently only supports '
'keyword arguments')
name = self.name
if name is None:
name = self.function.__name__
if self.name_parameters:
name = self._augment_name(name, kwargs)
log_dir = self.log_dir
if log_dir is None:
log_dir = ('{data}/local/{prefix}/{name}'.format(
data=os.path.join(os.getcwd(), 'data'),
prefix=self.prefix,
name=name))
log_dir = _make_sequential_log_dir(log_dir)
tabular_log_file = os.path.join(log_dir, 'progress.csv')
text_log_file = os.path.join(log_dir, 'debug.log')
variant_log_file = os.path.join(log_dir, 'variant.json')
metadata_log_file = os.path.join(log_dir, 'metadata.json')
dump_json(variant_log_file, kwargs)
git_root_path, metadata = get_metadata()
dump_json(metadata_log_file, metadata)
if git_root_path and self.archive_launch_repo:
make_launcher_archive(git_root_path=git_root_path, log_dir=log_dir)
logger.add_output(dowel.TextOutput(text_log_file))
logger.add_output(dowel.CsvOutput(tabular_log_file))
logger.add_output(
dowel.TensorBoardOutput(log_dir, x_axis='TotalEnvSteps'))
logger.add_output(dowel.StdOutput())
logger.push_prefix('[{}] '.format(name))
logger.log('Logging to {}'.format(log_dir))
return ExperimentContext(snapshot_dir=log_dir,
snapshot_mode=self.snapshot_mode,
snapshot_gap=self.snapshot_gap)
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')
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(n_epochs=976, batch_size=1024)
dowel_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.
Args:
env (gym.Env): Environment of the task.
seed (int): Random seed for the trial.
log_dir (str): Log dir path.
Returns:
str: Path to output csv file
"""
deterministic.set_seed(seed)
config = tf.compat.v1.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=12,
inter_op_parallelism_threads=12)
sess = tf.compat.v1.Session(config=config)
with LocalTFRunner(snapshot_config, sess=sess, max_cpus=12) as runner:
env = TfEnv(normalize(env))
policy = CategoricalGRUPolicy(
env_spec=env.spec,
hidden_dim=32,
hidden_nonlinearity=tf.nn.tanh,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
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),
),
)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(algo, env, sampler_args=dict(n_envs=12))
runner.train(n_epochs=488, batch_size=2048)
dowel_logger.remove_all()
return tabular_log_file
def run_garage_tf(env, seed, log_dir):
"""Create garage TensorFlow PPO model and training.
Args:
env (dict): Environment of the task.
seed (int): Random positive integer for the trial.
log_dir (str): Log dir path.
Returns:
str: Path to output csv file
"""
deterministic.set_seed(seed)
with LocalTFRunner(snapshot_config) as runner:
env = TfEnv(normalize(env))
policy = TF_GMP(
env_spec=env.spec,
hidden_sizes=hyper_parameters['hidden_sizes'],
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TF_PPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=hyper_parameters['max_path_length'],
discount=hyper_parameters['discount'],
gae_lambda=hyper_parameters['gae_lambda'],
center_adv=hyper_parameters['center_adv'],
lr_clip_range=hyper_parameters['lr_clip_range'],
optimizer_args=dict(
batch_size=None,
max_epochs=1,
tf_optimizer_args=dict(
learning_rate=hyper_parameters['learning_rate']),
verbose=True)) # yapf: disable
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(n_epochs=hyper_parameters['n_epochs'],
batch_size=hyper_parameters['batch_size'])
dowel_logger.remove_all()
return tabular_log_file
def __enter__(self):
tabular_log_file = os.path.join(self.log_dir, 'progress.csv')
text_log_file = os.path.join(self.log_dir, 'debug.log')
logger.add_output(dowel.TextOutput(text_log_file))
logger.add_output(dowel.CsvOutput(tabular_log_file))
logger.add_output(dowel.TensorBoardOutput(self.log_dir))
logger.add_output(dowel.StdOutput())
logger.push_prefix('[%s] ' % self.exp_name)
return self
def run_metarl(env, seed, log_dir):
"""Create metarl PyTorch MAML model and training.
Args:
env (MetaRLEnv): Environment of the task.
seed (int): Random positive integer for the trial.
log_dir (str): Log dir path.
Returns:
str: Path to output csv file
"""
deterministic.set_seed(seed)
policy = GaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=hyper_parameters['hidden_sizes'],
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = MAMLTRPO(env=env,
policy=policy,
baseline=baseline,
max_path_length=hyper_parameters['max_path_length'],
discount=hyper_parameters['discount'],
gae_lambda=hyper_parameters['gae_lambda'],
meta_batch_size=hyper_parameters['meta_batch_size'],
inner_lr=hyper_parameters['inner_lr'],
max_kl_step=hyper_parameters['max_kl'],
num_grad_updates=hyper_parameters['num_grad_update'])
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
snapshot_config = SnapshotConfig(snapshot_dir=log_dir,
snapshot_mode='all',
snapshot_gap=1)
runner = LocalRunner(snapshot_config=snapshot_config)
runner.setup(algo, env, sampler_args=dict(n_envs=5))
runner.train(n_epochs=hyper_parameters['n_epochs'],
batch_size=(hyper_parameters['fast_batch_size'] *
hyper_parameters['max_path_length']))
dowel_logger.remove_all()
return tabular_log_file
def _make_context(cls, options, **kwargs):
"""Make a context from the template information and variant args.
Currently, all arguments should be keyword arguments.
Args:
options (dict): Options to `wrap_experiment` itself. See the
function documentation for details.
kwargs (dict): Keyword arguments for the wrapped function. Will be
logged to `variant.json`
Returns:
ExperimentContext: The created experiment context.
"""
name = options['name']
if name is None:
name = options['function'].__name__
name = cls._augment_name(options, name, kwargs)
log_dir = options['log_dir']
if log_dir is None:
log_dir = ('{data}/local/{prefix}/{name}'.format(
data=os.path.join(os.getcwd(), 'data'),
prefix=options['prefix'],
name=name))
if options['use_existing_dir']:
os.makedirs(log_dir, exist_ok=True)
else:
log_dir = _make_sequential_log_dir(log_dir)
tabular_log_file = os.path.join(log_dir, 'progress.csv')
text_log_file = os.path.join(log_dir, 'debug.log')
variant_log_file = os.path.join(log_dir, 'variant.json')
metadata_log_file = os.path.join(log_dir, 'metadata.json')
dump_json(variant_log_file, kwargs)
git_root_path, metadata = get_metadata()
dump_json(metadata_log_file, metadata)
if git_root_path and options['archive_launch_repo']:
make_launcher_archive(git_root_path=git_root_path, log_dir=log_dir)
logger.add_output(dowel.TextOutput(text_log_file))
logger.add_output(dowel.CsvOutput(tabular_log_file))
logger.add_output(
dowel.TensorBoardOutput(log_dir, x_axis='TotalEnvSteps'))
logger.add_output(dowel.StdOutput())
logger.push_prefix('[{}] '.format(name))
logger.log('Logging to {}'.format(log_dir))
return ExperimentContext(snapshot_dir=log_dir,
snapshot_mode=options['snapshot_mode'],
snapshot_gap=options['snapshot_gap'])
def restore_training(log_dir, exp_name, args, env_saved=True, env=None):
tabular_log_file = os.path.join(
log_dir, 'progress_restored.{}.{}.csv'.format(
str(time.time())[:10], socket.gethostname()))
text_log_file = os.path.join(
log_dir, 'debug_restored.{}.{}.log'.format(
str(time.time())[:10], socket.gethostname()))
logger.add_output(dowel.TextOutput(text_log_file))
logger.add_output(dowel.CsvOutput(tabular_log_file))
logger.add_output(dowel.TensorBoardOutput(log_dir))
logger.add_output(dowel.StdOutput())
logger.push_prefix('[%s] ' % exp_name)
ctxt = ExperimentContext(snapshot_dir=log_dir,
snapshot_mode='last',
snapshot_gap=1)
runner = LocalRunnerWrapper(ctxt,
eval=args.eval_during_training,
n_eval_episodes=args.n_eval_episodes,
eval_greedy=args.eval_greedy,
eval_epoch_freq=args.eval_epoch_freq,
save_env=env_saved)
saved = runner._snapshotter.load(log_dir, 'last')
runner._setup_args = saved['setup_args']
runner._train_args = saved['train_args']
runner._stats = saved['stats']
set_seed(runner._setup_args.seed)
algo = saved['algo']
# Compatibility patch
if not hasattr(algo, '_clip_grad_norm'):
setattr(algo, '_clip_grad_norm', args.clip_grad_norm)
if env_saved:
env = saved['env']
runner.setup(env=env,
algo=algo,
sampler_cls=runner._setup_args.sampler_cls,
sampler_args=runner._setup_args.sampler_args)
runner._train_args.start_epoch = runner._stats.total_epoch + 1
runner._train_args.n_epochs = runner._train_args.start_epoch + args.n_epochs
print('\nRestored checkpoint from epoch #{}...'.format(
runner._train_args.start_epoch))
print('To be trained for additional {} epochs...'.format(args.n_epochs))
print('Will be finished at epoch #{}...\n'.format(
runner._train_args.n_epochs))
return runner._algo.train(runner)
def run_garage_pytorch(env, seed, log_dir):
"""Create garage PyTorch PPO model and training.
Args:
env (dict): Environment of the task.
seed (int): Random positive integer for the trial.
log_dir (str): Log dir path.
Returns:
str: Path to output csv file
"""
env = TfEnv(normalize(env))
deterministic.set_seed(seed)
runner = LocalRunner(snapshot_config)
policy = PyTorch_GMP(env.spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
value_functions = LinearFeatureBaseline(env_spec=env.spec)
algo = PyTorch_PPO(env_spec=env.spec,
policy=policy,
value_function=value_functions,
optimizer=torch.optim.Adam,
policy_lr=3e-4,
max_path_length=hyper_parameters['max_path_length'],
discount=0.99,
gae_lambda=0.95,
center_adv=True,
lr_clip_range=0.2,
minibatch_size=128,
max_optimization_epochs=10)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(n_epochs=hyper_parameters['n_epochs'],
batch_size=hyper_parameters['batch_size'])
dowel_logger.remove_all()
return tabular_log_file
def _make_context(self, *args, **kwargs):
"""Make a context from the template information and variant args.
Currently, all arguments should be keyword arguments.
Args:
args (list): Should be empty.
kwargs (dict): Keyword arguments for the wrapped function. Will be
logged to `variant.json`
Returns:
ExperimentContext: The created experiment context.
Raises:
ValueError: If args is not empty.
"""
if args:
raise ValueError('metarl.experiment currently only supports '
'keyword arguments')
log_dir = self.log_dir
if log_dir is None:
name = self.name
if name is None:
name = self.function.__name__
self.name = self.function.__name__
log_dir = ('{data}/local/{prefix}/{name}/{time}'.format(
data=osp.join(os.getcwd(), 'data'),
prefix=self.prefix,
name=name,
time=timestamp))
log_dir = _make_sequential_log_dir(log_dir)
tabular_log_file = os.path.join(log_dir, 'progress.csv')
text_log_file = os.path.join(log_dir, 'debug.log')
variant_log_file = os.path.join(log_dir, 'variant.json')
dump_json(variant_log_file, kwargs)
logger.add_output(dowel.TextOutput(text_log_file))
logger.add_output(dowel.CsvOutput(tabular_log_file))
logger.add_output(
dowel.TensorBoardOutput(log_dir, x_axis='TotalEnvSteps'))
logger.add_output(dowel.StdOutput())
logger.push_prefix('[%s] ' % self.name)
return ExperimentContext(snapshot_dir=log_dir,
snapshot_mode=self.snapshot_mode,
snapshot_gap=self.snapshot_gap)
def run_garage_pytorch(env, seed, log_dir):
"""Create garage PyTorch PPO model and training.
Args:
env (dict): Environment of the task.
seed (int): Random positive integer for the trial.
log_dir (str): Log dir path.
Returns:
str: Path to output csv file
"""
env = TfEnv(normalize(env))
deterministic.set_seed(seed)
runner = LocalRunner(snapshot_config)
policy = PyTorch_GMP(env.spec,
hidden_sizes=hyper_parameters['hidden_sizes'],
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = PyTorch_TRPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=hyper_parameters['max_path_length'],
discount=hyper_parameters['discount'],
gae_lambda=hyper_parameters['gae_lambda'],
max_kl=hyper_parameters['max_kl'],
)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(n_epochs=hyper_parameters['n_epochs'],
batch_size=hyper_parameters['batch_size'])
dowel_logger.remove_all()
return tabular_log_file
def run_garage_tf(env, seed, log_dir):
"""Create garage TensorFlow VPG model and training."""
deterministic.set_seed(seed)
with LocalTFRunner(snapshot_config) as runner:
env = TfEnv(normalize(env))
policy = TF_GMP(
env_spec=env.spec,
hidden_sizes=hyper_parameters['hidden_sizes'],
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TF_VPG(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=hyper_parameters['max_path_length'],
discount=hyper_parameters['discount'],
center_adv=hyper_parameters['center_adv'],
optimizer_args=dict(
tf_optimizer_args=dict(
learning_rate=hyper_parameters['learning_rate']),
verbose=True)) # yapf: disable
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(n_epochs=hyper_parameters['n_epochs'], batch_size=2048)
dowel_logger.remove_all()
return tabular_log_file
def run_garage_pytorch(env, seed, log_dir):
"""Create garage PyTorch VPG model and training."""
env = TfEnv(normalize(env))
deterministic.set_seed(seed)
runner = LocalRunner(snapshot_config)
policy = PyTorch_GMP(
env.spec,
hidden_sizes=hyper_parameters['hidden_sizes'],
hidden_nonlinearity=torch.tanh,
# init_std=1e-6,
output_nonlinearity=None)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = PyTorch_VPG(env_spec=env.spec,
policy=policy,
optimizer=torch.optim.Adam,
baseline=baseline,
max_path_length=hyper_parameters['max_path_length'],
discount=hyper_parameters['discount'],
center_adv=hyper_parameters['center_adv'],
policy_lr=hyper_parameters['learning_rate'])
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(n_epochs=hyper_parameters['n_epochs'], batch_size=2048)
dowel_logger.remove_all()
return tabular_log_file
def run_experiment(argv):
"""Run experiment.
Args:
argv (list[str]): Command line arguments.
Raises:
BaseException: Propagate any exception in the experiment.
"""
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='last',
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(
'--resume_from_dir',
type=str,
default=None,
help='Directory of the pickle file to resume experiment from.')
parser.add_argument('--resume_from_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('--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('--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,
default=None,
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')
args = parser.parse_args(argv[1:])
if args.seed is not None:
deterministic.set_seed(args.seed)
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:
log_dir = os.path.join(os.path.join(os.getcwd(), 'data'),
args.exp_name)
else:
log_dir = args.log_dir
tabular_log_file = os.path.join(log_dir, args.tabular_log_file)
text_log_file = os.path.join(log_dir, args.text_log_file)
params_log_file = os.path.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 = os.path.join(log_dir, args.variant_log_file)
dump_variant(variant_log_file, variant_data)
else:
variant_data = None
log_parameters(params_log_file, args)
logger.add_output(dowel.TextOutput(text_log_file))
logger.add_output(dowel.CsvOutput(tabular_log_file))
logger.add_output(dowel.TensorBoardOutput(log_dir, x_axis='TotalEnvSteps'))
logger.add_output(dowel.StdOutput())
logger.push_prefix('[%s] ' % args.exp_name)
snapshot_config = SnapshotConfig(snapshot_dir=log_dir,
snapshot_mode=args.snapshot_mode,
snapshot_gap=args.snapshot_gap)
method_call = cloudpickle.loads(base64.b64decode(args.args_data))
try:
method_call(snapshot_config, variant_data, args.resume_from_dir,
args.resume_from_epoch)
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
logger.remove_all()
logger.pop_prefix()
def run_metarl(env, envs, tasks, seed, log_dir):
"""Create metarl Tensorflow PPO model and training.
Args:
env (dict): Environment of the task.
seed (int): Random positive integer for the trial.
log_dir (str): Log dir path.
Returns:
str: Path to output csv file
"""
deterministic.set_seed(seed)
snapshot_config = SnapshotConfig(snapshot_dir=log_dir,
snapshot_mode='gap',
snapshot_gap=10)
with LocalTFRunner(snapshot_config) as runner:
policy = GaussianGRUPolicy(
hidden_dims=hyper_parameters['hidden_sizes'],
env_spec=env.spec,
state_include_action=False)
baseline = MetaRLLinearFeatureBaseline(env_spec=env.spec)
inner_algo = RL2PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=hyper_parameters['max_path_length'] * hyper_parameters['rollout_per_task'],
discount=hyper_parameters['discount'],
gae_lambda=hyper_parameters['gae_lambda'],
lr_clip_range=hyper_parameters['lr_clip_range'],
optimizer_args=dict(
max_epochs=hyper_parameters['optimizer_max_epochs'],
tf_optimizer_args=dict(
learning_rate=hyper_parameters['optimizer_lr'],
),
)
)
# Need to pass this if meta_batch_size < num_of_tasks
task_names = list(ML45_ENVS['train'].keys())
algo = RL2(
policy=policy,
inner_algo=inner_algo,
max_path_length=hyper_parameters['max_path_length'],
meta_batch_size=hyper_parameters['meta_batch_size'],
task_sampler=tasks,
task_names=None if hyper_parameters['meta_batch_size'] >= len(task_names) else task_names)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
text_log_file = osp.join(log_dir, 'debug.log')
dowel_logger.add_output(dowel.TextOutput(text_log_file))
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(
algo,
envs,
sampler_cls=hyper_parameters['sampler_cls'],
n_workers=hyper_parameters['meta_batch_size'],
worker_class=RL2Worker,
sampler_args=dict(
use_all_workers=hyper_parameters['use_all_workers']),
worker_args=dict(
n_paths_per_trial=hyper_parameters['rollout_per_task']))
# meta evaluator
env_obs_dim = [env().observation_space.shape[0] for (_, env) in ML45_ENVS['test'].items()]
max_obs_dim = max(env_obs_dim)
ML_test_envs = [
TaskIdWrapper(NormalizedRewardEnv(RL2Env(env(*ML45_ARGS['test'][task]['args'],
**ML45_ARGS['test'][task]['kwargs']), max_obs_dim)), task_id=task_id, task_name=task)
for (task_id, (task, env)) in enumerate(ML45_ENVS['test'].items())
]
test_tasks = task_sampler.EnvPoolSampler(ML_test_envs)
test_tasks.grow_pool(hyper_parameters['n_test_tasks'])
test_task_names = list(ML45_ENVS['test'].keys())
runner.setup_meta_evaluator(test_task_sampler=test_tasks,
n_exploration_traj=hyper_parameters['rollout_per_task'],
n_test_rollouts=hyper_parameters['test_rollout_per_task'],
n_test_tasks=hyper_parameters['n_test_tasks'],
n_workers=hyper_parameters['n_test_tasks'],
test_task_names=None if hyper_parameters['n_test_tasks'] >= len(test_task_names) else test_task_names)
runner.train(n_epochs=hyper_parameters['n_itr'],
batch_size=hyper_parameters['meta_batch_size'] * hyper_parameters['rollout_per_task'] * hyper_parameters['max_path_length'])
dowel_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(normalize(env))
action_noise = OUStrategy(env.spec, sigma=params['sigma'])
policy = ContinuousMLPPolicyWithModel(
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'],
target_update_tau=params['tau'],
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(dowel.StdOutput())
logger.add_output(dowel.CsvOutput(tabular_log_file))
logger.add_output(dowel.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
"""Minimal example of dowel usage.
This example demonstrates how to log a simple progress metric using dowel.
The metric is simultaneously sent to the screen, a CSV files, a text log file
and TensorBoard.
"""
import time
import dowel
from dowel import logger, tabular
logger.add_output(dowel.StdOutput())
logger.add_output(dowel.CsvOutput('progress.csv'))
logger.add_output(dowel.TextOutput('progress.txt'))
logger.add_output(dowel.TensorBoardOutput('tensorboard_logdir'))
logger.log('Starting up...')
for i in range(1000):
logger.push_prefix('itr {}: '.format(i))
logger.log('Running training step')
time.sleep(0.01) # Tensorboard doesn't like output to be too fast.
tabular.record('itr', i)
tabular.record('loss', 100.0 / (2 + i))
logger.log(tabular)
logger.pop_prefix()
logger.dump_all()
def run_metarl(env, seed, log_dir):
"""Create metarl Tensorflow PPO model and training.
Args:
env (dict): Environment of the task.
seed (int): Random positive integer for the trial.
log_dir (str): Log dir path.
Returns:
str: Path to output csv file
"""
deterministic.set_seed(seed)
snapshot_config = SnapshotConfig(snapshot_dir=log_dir,
snapshot_mode='gap',
snapshot_gap=10)
with LocalTFRunner(snapshot_config) as runner:
env, task_samplers = _prepare_meta_env(env)
policy = GaussianGRUPolicy(
hidden_dims=hyper_parameters['hidden_sizes'],
env_spec=env.spec,
state_include_action=False)
baseline = MetaRLLinearFeatureBaseline(env_spec=env.spec)
inner_algo = RL2PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=hyper_parameters['max_path_length'] *
hyper_parameters['rollout_per_task'],
discount=hyper_parameters['discount'],
gae_lambda=hyper_parameters['gae_lambda'],
lr_clip_range=hyper_parameters['lr_clip_range'],
optimizer_args=dict(
max_epochs=hyper_parameters['optimizer_max_epochs'],
tf_optimizer_args=dict(
learning_rate=hyper_parameters['optimizer_lr'], ),
))
algo = RL2(policy=policy,
inner_algo=inner_algo,
max_path_length=hyper_parameters['max_path_length'],
meta_batch_size=hyper_parameters['meta_batch_size'],
task_sampler=task_samplers)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
text_log_file = osp.join(log_dir, 'debug.log')
dowel_logger.add_output(dowel.TextOutput(text_log_file))
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(
algo,
task_samplers.sample(hyper_parameters['meta_batch_size']),
sampler_cls=hyper_parameters['sampler_cls'],
n_workers=hyper_parameters['meta_batch_size'],
worker_class=RL2Worker,
sampler_args=dict(
use_all_workers=hyper_parameters['use_all_workers']),
worker_args=dict(
n_paths_per_trial=hyper_parameters['rollout_per_task']))
runner.setup_meta_evaluator(
test_task_sampler=task_samplers,
n_exploration_traj=hyper_parameters['rollout_per_task'],
n_test_rollouts=hyper_parameters['test_rollout_per_task'],
n_test_tasks=hyper_parameters['n_test_tasks'],
n_workers=hyper_parameters['n_test_tasks'])
runner.train(n_epochs=hyper_parameters['n_itr'],
batch_size=hyper_parameters['meta_batch_size'] *
hyper_parameters['rollout_per_task'] *
hyper_parameters['max_path_length'])
dowel_logger.remove_all()
return tabular_log_file
def main(args):
if args.output_folder is not None:
if not os.path.exists(args.output_folder):
raise ValueError(
"The folder with the training files does not exist")
policy_filename = os.path.join(args.output_folder, 'policy.th')
dynamics_filename = os.path.join(args.output_folder, 'dynamics.th')
config_filename = os.path.join(args.output_folder, 'config.json')
# eval_filename = os.path.join(args.output_folder, 'eval.npz')
text_log_file = os.path.join(args.output_folder, 'test_log.txt')
tabular_log_file = os.path.join(args.output_folder, 'test_result.csv')
output_test_folder = args.output_folder + "test" if args.output_folder[
-1] == '/' else args.output_folder + "/test"
if os.path.exists(output_test_folder):
shutil.rmtree(output_test_folder)
os.makedirs(output_test_folder)
# Set up logger
logger.add_output(dowel.StdOutput())
logger.add_output(dowel.TextOutput(text_log_file))
logger.add_output(dowel.CsvOutput(tabular_log_file))
logger.add_output(
dowel.TensorBoardOutput(output_test_folder, x_axis='Batch'))
logger.log('Logging to {}'.format(output_test_folder))
with open(config_filename, 'r') as f:
config = json.load(f)
seed = config["seed"] if "seed" in config else args.seed
if seed is not None:
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
random.seed(args.seed)
# Metaworld
if config['env-name'].startswith('Metaworld'):
env_name = config['env-name'].replace("Metaworld-", "")
metaworld = __import__('metaworld')
class_ = getattr(metaworld, env_name)
metaworld_benchmark = class_()
for name, env_cls in metaworld_benchmark.train_classes.items():
env = env_cls()
env.close()
benchmark = metaworld_benchmark
# Other gym envs
else:
env_name = config['env-name']
env = gym.make(config['env-name'], **config.get('env-kwargs', {}))
env.close()
benchmark = None
# Policy
policy = get_policy_for_env(env,
hidden_sizes=config['hidden-sizes'],
nonlinearity=config['nonlinearity'])
with open(policy_filename, 'rb') as f:
state_dict = torch.load(f, map_location=torch.device(args.device))
policy.load_state_dict(state_dict)
policy.share_memory()
# Dynamics
dynamics = get_dynamics_for_env(env,
config['use_vime'],
config['use_inv_vime'],
args.device,
config,
benchmark=benchmark)
inverse_dynamics = config['use_inv_vime']
use_dynamics = config["use_vime"] or config["use_inv_vime"]
if use_dynamics:
with open(dynamics_filename, 'rb') as f:
state_dict = torch.load(f, map_location=torch.device(args.device))
dynamics.load_state_dict(state_dict)
dynamics.share_memory()
# Eta
if config['adapt_eta']:
eta_value = torch.Tensor([config["adapted-eta"]])
else:
eta_value = torch.Tensor([config["eta"]])
eta_value = torch.log(eta_value / (1 - eta_value))
eta = EtaParameter(eta_value, adapt_eta=config['adapt_eta'])
eta.share_memory()
# Baseline
baseline = LinearFeatureBaseline(get_input_size(env))
# Sampler
normalize_spaces = config[
"normalize-spaces"] if "normalize-spaces" in config else True
act_prev_mean = mp.Manager().list()
obs_prev_mean = mp.Manager().list()
# Sampler
if normalize_spaces:
obs_prev_mean.append({
"mean": torch.Tensor(config["obs_mean"]),
"std": torch.Tensor(config["obs_std"])
})
act_prev_mean.append({
"mean": torch.Tensor(config["act_mean"]),
"std": torch.Tensor(config["act_std"])
})
epochs_counter = mp.Value('i', 100)
sampler = MultiTaskSampler(
config['env-name'],
env_kwargs=config.get('env-kwargs', {}),
batch_size=config['fast-batch-size'], # TODO
policy=policy,
baseline=baseline,
dynamics=dynamics,
inverse_dynamics=inverse_dynamics,
env=env,
seed=args.seed,
num_workers=args.num_workers,
epochs_counter=epochs_counter,
act_prev_mean=act_prev_mean,
obs_prev_mean=obs_prev_mean,
# rew_prev_mean=rew_prev_mean,
eta=eta,
benchmark=benchmark,
normalize_spaces=normalize_spaces)
logs = {'tasks': []}
train_returns, valid_returns = [], []
for batch in trange(args.num_batches):
tasks = sampler.sample_test_tasks(num_tasks=config['meta-batch-size'])
train_episodes, valid_episodes = sampler.sample(
tasks,
num_steps=args.num_steps,
fast_lr=config['fast-lr'],
gamma=config['gamma'],
gae_lambda=config['gae-lambda'],
device=args.device)
logs['tasks'].extend(tasks)
train_returns.append(get_returns(train_episodes[0]))
valid_returns.append(get_returns(valid_episodes))
logs['train_returns'] = np.concatenate(train_returns, axis=0)
logs['valid_returns'] = np.concatenate(valid_returns, axis=0)
tabular.record("Batch", batch)
log_returns(train_episodes,
valid_episodes,
batch,
log_dynamics=use_dynamics,
benchmark=benchmark,
env=env,
env_name=env_name,
is_testing=True)
log_trajectories(config['env-name'], output_test_folder,
train_episodes, valid_episodes, batch)
logger.log(tabular)
logger.dump_all()
# with open(eval_filename + "_" + str(batch), 'wb') as f:
# np.savez(f, **logs)
logger.remove_all()
def run_experiment(argv):
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='last',
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(
'--resume_from_dir',
type=str,
default=None,
help='Directory of the pickle file to resume experiment from.')
parser.add_argument('--resume_from_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('--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('--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,
default=None,
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:
log_dir = os.path.join(os.path.join(os.getcwd(), 'data'),
args.exp_name)
else:
log_dir = args.log_dir
tabular_log_file = os.path.join(log_dir, args.tabular_log_file)
text_log_file = os.path.join(log_dir, args.text_log_file)
params_log_file = os.path.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 = os.path.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(dowel.TextOutput(text_log_file))
logger.add_output(dowel.CsvOutput(tabular_log_file))
logger.add_output(dowel.TensorBoardOutput(log_dir))
logger.add_output(dowel.StdOutput())
logger.push_prefix('[%s] ' % args.exp_name)
snapshot_config = SnapshotConfig(snapshot_dir=log_dir,
snapshot_mode=args.snapshot_mode,
snapshot_gap=args.snapshot_gap)
# read from stdin
if args.use_cloudpickle:
import cloudpickle
method_call = cloudpickle.loads(base64.b64decode(args.args_data))
try:
method_call(snapshot_config, variant_data, args.resume_from_dir,
args.resume_from_epoch)
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
logger.remove_all()
logger.pop_prefix()
def run_metarl(env, seed, log_dir):
'''
Create metarl 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)
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=12,
inter_op_parallelism_threads=12)
sess = tf.Session(config=config)
with LocalTFRunner(snapshot_config, sess=sess, max_cpus=12) as runner:
env = TfEnv(normalize(env))
policy = CategoricalCNNPolicy(
env_spec=env.spec,
conv_filters=params['conv_filters'],
conv_filter_sizes=params['conv_filter_sizes'],
conv_strides=params['conv_strides'],
conv_pad=params['conv_pad'],
hidden_sizes=params['hidden_sizes'])
baseline = GaussianCNNBaseline(
env_spec=env.spec,
regressor_args=dict(num_filters=params['conv_filters'],
filter_dims=params['conv_filter_sizes'],
strides=params['conv_strides'],
padding=params['conv_pad'],
hidden_sizes=params['hidden_sizes'],
use_trust_region=params['use_trust_region']))
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),
),
flatten_input=False,
)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(n_epochs=params['n_epochs'],
batch_size=params['batch_size'])
dowel_logger.remove_all()
return tabular_log_file
def run_garage(env, seed, log_dir):
"""
Create garage model and training.
Replace the td3 with the algorithm you want to run.
:param env: Environment of the task.
:param seed: Random seed for the trail.
:param log_dir: Log dir path.
:return:
"""
deterministic.set_seed(seed)
with LocalTFRunner(snapshot_config) as runner:
env = TfEnv(normalize(env))
# Set up params for TD3
exploration_noise = GaussianStrategy(env.spec,
max_sigma=params['sigma'],
min_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(name='ContinuousMLPQFunction',
env_spec=env.spec,
hidden_sizes=params['qf_hidden_sizes'],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
qf2 = ContinuousMLPQFunction(name='ContinuousMLPQFunction2',
env_spec=env.spec,
hidden_sizes=params['qf_hidden_sizes'],
action_merge_layer=0,
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'])
td3 = TD3(env.spec,
policy=policy,
qf=qf,
qf2=qf2,
replay_buffer=replay_buffer,
steps_per_epoch=params['steps_per_epoch'],
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'],
smooth_return=params['smooth_return'],
min_buffer_size=params['min_buffer_size'],
buffer_batch_size=params['buffer_batch_size'],
exploration_strategy=exploration_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')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(td3, env)
runner.train(n_epochs=params['n_epochs'],
batch_size=params['n_rollout_steps'])
dowel_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.
Args:
env (gym.Env): Environment of the task.
seed (int): Random seed for the trial.
log_dir (str): Log dir path.
Returns:
str: Log file path.
"""
deterministic.set_seed(seed)
config = tf.compat.v1.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=12,
inter_op_parallelism_threads=12)
sess = tf.compat.v1.Session(config=config)
with LocalTFRunner(snapshot_config, sess=sess, max_cpus=12) as runner:
env = TfEnv(normalize(env))
# Set up params for ddpg
action_noise = OUStrategy(env.spec, sigma=params['sigma'])
policy = ContinuousMLPPolicy(
env_spec=env.spec,
name='ContinuousMLPPolicy',
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,
name='ContinuousMLPQFunction')
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,
steps_per_epoch=params['steps_per_epoch'],
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.compat.v1.train.AdamOptimizer,
qf_optimizer=tf.compat.v1.train.AdamOptimizer)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(ddpg, env, sampler_args=dict(n_envs=12))
runner.train(n_epochs=params['n_epochs'],
batch_size=params['n_rollout_steps'])
dowel_logger.remove_all()
return tabular_log_file
def run_metarl(env, test_env, seed, log_dir):
"""Create metarl model and training."""
deterministic.set_seed(seed)
snapshot_config = SnapshotConfig(snapshot_dir=log_dir,
snapshot_mode='gap',
snapshot_gap=10)
runner = LocalRunner(snapshot_config)
obs_dim = int(np.prod(env[0]().observation_space.shape))
action_dim = int(np.prod(env[0]().action_space.shape))
reward_dim = 1
# instantiate networks
encoder_in_dim = obs_dim + action_dim + reward_dim
encoder_out_dim = params['latent_size'] * 2
net_size = params['net_size']
context_encoder = MLPEncoder(input_dim=encoder_in_dim,
output_dim=encoder_out_dim,
hidden_sizes=[200, 200, 200])
space_a = akro.Box(low=-1,
high=1,
shape=(obs_dim + params['latent_size'], ),
dtype=np.float32)
space_b = akro.Box(low=-1, high=1, shape=(action_dim, ), dtype=np.float32)
augmented_env = EnvSpec(space_a, space_b)
qf1 = ContinuousMLPQFunction(env_spec=augmented_env,
hidden_sizes=[net_size, net_size, net_size])
qf2 = ContinuousMLPQFunction(env_spec=augmented_env,
hidden_sizes=[net_size, net_size, net_size])
obs_space = akro.Box(low=-1, high=1, shape=(obs_dim, ), dtype=np.float32)
action_space = akro.Box(low=-1,
high=1,
shape=(params['latent_size'], ),
dtype=np.float32)
vf_env = EnvSpec(obs_space, action_space)
vf = ContinuousMLPQFunction(env_spec=vf_env,
hidden_sizes=[net_size, net_size, net_size])
policy = TanhGaussianMLPPolicy2(
env_spec=augmented_env, hidden_sizes=[net_size, net_size, net_size])
context_conditioned_policy = ContextConditionedPolicy(
latent_dim=params['latent_size'],
context_encoder=context_encoder,
policy=policy,
use_ib=params['use_information_bottleneck'],
use_next_obs=params['use_next_obs_in_context'],
)
train_task_names = ML10.get_train_tasks()._task_names
test_task_names = ML10.get_test_tasks()._task_names
pearlsac = PEARLSAC(
env=env,
test_env=test_env,
policy=context_conditioned_policy,
qf1=qf1,
qf2=qf2,
vf=vf,
num_train_tasks=params['num_train_tasks'],
num_test_tasks=params['num_test_tasks'],
latent_dim=params['latent_size'],
meta_batch_size=params['meta_batch_size'],
num_steps_per_epoch=params['num_steps_per_epoch'],
num_initial_steps=params['num_initial_steps'],
num_tasks_sample=params['num_tasks_sample'],
num_steps_prior=params['num_steps_prior'],
num_extra_rl_steps_posterior=params['num_extra_rl_steps_posterior'],
num_evals=params['num_evals'],
num_steps_per_eval=params['num_steps_per_eval'],
batch_size=params['batch_size'],
embedding_batch_size=params['embedding_batch_size'],
embedding_mini_batch_size=params['embedding_mini_batch_size'],
max_path_length=params['max_path_length'],
reward_scale=params['reward_scale'],
train_task_names=train_task_names,
test_task_names=test_task_names,
)
tu.set_gpu_mode(params['use_gpu'], gpu_id=0)
if params['use_gpu']:
pearlsac.to()
tabular_log_file = osp.join(log_dir, 'progress.csv')
tensorboard_log_dir = osp.join(log_dir)
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(tensorboard_log_dir))
runner.setup(algo=pearlsac,
env=env,
sampler_cls=PEARLSampler,
sampler_args=dict(max_path_length=params['max_path_length']))
runner.train(n_epochs=params['num_epochs'],
batch_size=params['batch_size'])
dowel_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 LocalTFRunner(snapshot_config) 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),
),
)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(algo, env)
runner.train(n_epochs=488, batch_size=2048)
dowel_logger.remove_all()
return tabular_log_file