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 __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 _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 _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_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()
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()
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()
