def test_time_limit(test_config):
config = DEFAULT_CONFIG.copy()
config.update(simple_addition.config)
config.update(reinforce_config)
config.update(max_time=120, max_steps=10000, seed=100)
config.update(test_config)
start = time.time()
with config:
training_loop()
elapsed = start - time.time()
assert elapsed < config.max_time + 1
def test_time_limit_between_stages(test_config):
config = DEFAULT_CONFIG.copy()
config.update(simple_addition.config)
config.update(reinforce_config)
config.update(max_time=120, max_steps=10, seed=100)
config.update(hooks=[AlarmHook(False, 0)])
config.update(test_config)
start = time.time()
with config:
result = training_loop()
print(result)
elapsed = start - time.time()
assert elapsed < 20
def _run(env_str, alg_str, _config=None, **kwargs):
env_config, alg_config = parse_env_alg(env_str, alg_str)
config = DEFAULT_CONFIG.copy()
config.update(alg_config)
config.update(env_config)
if _config is not None:
config.update(_config)
config.update(kwargs)
with config:
cfg.update_from_command_line()
return training_loop()
def __call__(self, new):
import os
stdout_path = f"./stdout_pid={os.getpid()}"
with redirect_stream('stdout', stdout_path, tee=True):
start_time = time.time()
print("Entered _RunTrainingLoop at: ")
print(datetime.datetime.now())
os.nice(10)
print("Sampled values: ")
print(new)
print("Base config: ")
print(self.base_config)
exp_name = '_'.join("{}={}".format(k, new[k])
for k in 'idx repeat'.split())
config = get_default_config()
config.update(self.base_config)
config.update(new)
config.update(
start_tensorboard=False,
show_plots=False,
update_latest=False,
git_record_mode='none',
in_parallel_session=True,
# Need these present so that they're are picked up when we get args from command line.
local_experiments_dir='',
backup_dir='',
env_name='',
max_time=0,
)
with config:
# This is used for passing args 'local_experiments_dir', 'backup_dir', 'env_name', and 'max_time'
cfg.update_from_command_line(strict=False)
from dps.train import training_loop
result = training_loop(exp_name=exp_name,
start_time=start_time)
print("Leaving _RunTrainingLoop at: ")
print(datetime.datetime.now())
return result
def test_stage_hook(test_config):
""" Test that we can safely use hooks to add new stages. """
config = DEFAULT_CONFIG.copy()
config.update(simple_addition.config)
config.update(reinforce_config)
config.update(
max_steps=11,
eval_step=10,
n_train=100,
seed=100,
hooks=[DummyHook(3, dict(max_steps=21))],
curriculum=[dict()],
width=1,
)
config.update(test_config)
with config:
data = training_loop()
assert data.n_stages == 3
assert not data.history[0]["stage_config"]
assert data.history[1]["stage_config"]["max_steps"] == 21
assert data.history[2]["stage_config"]["max_steps"] == 21
def __call__(self, new):
start_time = time.time()
print("Entered _RunTrainingLoop at: ")
print(datetime.datetime.now())
os.nice(10)
print("Sampled values: ")
print(new)
print("Base config: ")
print(self.base_config)
exp_name = '_'.join("{}={}".format(k, new[k]) for k in 'idx repeat'.split())
dps.reset_config()
config = DEFAULT_CONFIG.copy()
config.update(self.base_config)
config.update(new)
config.update(
start_tensorboard=False,
show_plots=False,
)
with config:
cfg.update_from_command_line()
from dps.train import training_loop
result = training_loop(exp_name=exp_name, start_time=start_time)
print("Leaving _RunTrainingLoop at: ")
print(datetime.datetime.now())
return result
def run_experiment(name,
config,
readme,
distributions=None,
durations=None,
alg=None,
task="grid",
name_variables=None,
env_kwargs=None):
name = sanitize(name)
durations = durations or {}
parser = argparse.ArgumentParser()
parser.add_argument("duration", choices=list(durations.keys()) + ["local"])
args, _ = parser.parse_known_args()
_config = DEFAULT_CONFIG.copy()
env_kwargs = env_kwargs or {}
env_kwargs['task'] = task
env_config = get_env_config(**env_kwargs)
_config.update(env_config)
if alg:
alg_config = getattr(alg_module, "{}_config".format(alg))
_config.update(alg_config)
alg_name = sanitize(alg_config.alg_name)
else:
alg_name = ""
_config.update(config)
_config.update_from_command_line()
_config.env_name = "{}_env={}".format(name, sanitize(env_config.env_name))
if args.duration == "local":
_config.exp_name = "alg={}".format(alg_name)
with _config:
return training_loop()
else:
run_kwargs = Config(
kind="slurm",
pmem=5000,
ignore_gpu=False,
)
duration_args = durations[args.duration]
if 'config' in duration_args:
_config.update(duration_args['config'])
del duration_args['config']
run_kwargs.update(durations[args.duration])
run_kwargs.update_from_command_line()
if name_variables is not None:
name_variables_str = "_".join("{}={}".format(
sanitize(str(k)), sanitize(str(getattr(_config, k))))
for k in name_variables.split(","))
_config.env_name = "{}_{}".format(_config.env_name, name_variables_str)
exp_name = "{}_alg={}_duration={}".format(_config.env_name, alg_name,
args.duration)
build_and_submit(name=exp_name,
config=_config,
distributions=distributions,
**run_kwargs)
def _raw_run(config):
with config:
return training_loop()
def run_experiment(name,
base_config,
readme,
distributions=None,
durations=None,
name_variables=None,
alg_configs=None,
env_configs=None,
late_config=None,
cl_mode='lax',
run_kwargs_base=None):
name = sanitize(name)
durations = durations or {}
parser = argparse.ArgumentParser()
if env_configs is not None:
parser.add_argument('env')
if alg_configs is not None:
parser.add_argument('alg')
parser.add_argument("duration",
choices=list(durations.keys()) + ["local"],
default="local",
nargs="?")
args, _ = parser.parse_known_args()
config = get_default_config()
config.update(base_config)
if env_configs is not None:
env_config = env_configs[args.env]
config.update(env_config)
if alg_configs is not None:
alg_config = alg_configs[args.alg]
config.update(alg_config)
if late_config is not None:
config.update(late_config)
env_name = sanitize(config.get('env_name', ''))
alg_name = sanitize(config.get("alg_name", ""))
run_kwargs = Config(
kind="slurm",
ignore_gpu=False,
)
if run_kwargs_base is not None:
run_kwargs.update(run_kwargs_base)
if args.duration == "local":
run_kwargs.update(durations.get('local', {}))
else:
run_kwargs.update(durations[args.duration])
if 'config' in run_kwargs:
config.update(run_kwargs.config)
del run_kwargs['config']
if cl_mode is not None:
if cl_mode == 'strict':
config.update_from_command_line(strict=True)
elif cl_mode == 'lax':
config.update_from_command_line(strict=False)
else:
raise Exception("Unknown value for cl_mode: {}".format(cl_mode))
if args.duration == "local":
config.exp_name = "alg={}".format(alg_name)
with config:
return training_loop()
else:
if 'distributions' in run_kwargs:
distributions = run_kwargs['distributions']
del run_kwargs['distributions']
if name_variables is not None:
name_variables_str = "_".join(
"{}={}".format(sanitize(k), sanitize(getattr(config, k)))
for k in name_variables.split(","))
env_name = "{}_{}".format(env_name, name_variables_str)
config.env_name = env_name
exp_name = "env={}_alg={}_duration={}".format(env_name, alg_name,
args.duration)
init()
build_and_submit(name,
exp_name,
config,
distributions=distributions,
**run_kwargs)
def build_and_submit(category,
exp_name,
config,
distributions,
n_param_settings=0,
n_repeats=1,
do_local_test=False,
kind="local",
readme="",
tasks_per_gpu=1,
**run_kwargs):
""" Build a job and submit it. Meant to be called from within a script.
Parameters
----------
category: str
High-level category of the experiment. Determines the ExperimentStore
where the experiment data will be stored.
exp_name: str
Low-level name of the experiment.
config: Config instance or dict
Configuration to use as the base config for all jobs.
distributions: dict
Object used to generate variations of the base config (so that different
jobs test different parameters).
n_param_settings: int
Number of different configurations to sample from `distributions`. If not
supplied, it is assumed that `distributions` actually specifies a grid
search, and an attempt is made to generate all possible configurations int
that grid search.
n_repeats: int
Number of experiments to run (with different random seeds) for each
generated configuration.
do_local_test: bool
If True, sample one of the generated configurations and use it to run a
short test locally, to ensure that the jobs will run properly.
kind: str
One of pbs, slurm, slurm-local, parallel, local. Specifies which method
should be used to run the jobs in parallel.
readme: str
A string outlining the purpose/context for the created experiment.
**run_kwargs:
Additional arguments that are ultimately passed to `ParallelSession` in
order to run the job.
"""
# Get run_kwargs from command line
sig = inspect.signature(ParallelSession.__init__)
default_run_kwargs = sig.bind_partial()
default_run_kwargs.apply_defaults()
cl_run_kwargs = clify.command_line(default_run_kwargs.arguments).parse()
run_kwargs.update(cl_run_kwargs)
if config.seed is None or config.seed < 0:
config.seed = gen_seed()
assert kind in "pbs slurm slurm-local parallel local".split()
assert 'build_command' not in config
config['build_command'] = ' '.join(sys.argv)
print(config['build_command'])
if kind == "local":
with config:
from dps.train import training_loop
return training_loop()
else:
config.name = category
config = config.copy()
if readme == "_vim_":
readme = edit_text(prefix="dps_readme_",
editor="vim",
initial_text="README.md: \n")
scratch = os.path.join(cfg.parallel_experiments_build_dir, category)
archive_path, n_tasks = build_search(scratch,
exp_name,
distributions,
config,
add_date=1,
_zip=True,
do_local_test=do_local_test,
n_param_settings=n_param_settings,
n_repeats=n_repeats,
readme=readme)
run_kwargs.update(archive_path=archive_path,
category=category,
exp_name=exp_name,
kind=kind)
gpu_kind = run_kwargs.get('gpu_kind', None)
resources = compute_required_resources(n_tasks, tasks_per_gpu,
gpu_kind)
run_kwargs.update(resources)
parallel_session = submit_job(**run_kwargs)
return parallel_session
x = int(sys.argv[1])
if x == 0:
print("TRPE")
config.update(name="TRPE",
delta_schedule='0.01',
max_cg_steps=10,
max_line_search_steps=10,
alg_class=TrustRegionPolicyEvaluation)
elif x == 1:
print("PPE")
config.update(name="PPE",
optimizer_spec="rmsprop",
lr_schedule="1e-2",
epsilon=0.2,
opt_steps_per_update=100,
S=1,
alg_class=ProximalPolicyEvaluation)
else:
print("PE")
config.update(name="PolicyEvaluation",
optimizer_spec='rmsprop',
lr_schedule='1e-5',
opt_steps_per_update=100,
alg_class=PolicyEvaluation)
with config:
cfg.update_from_command_line()
training_loop()
def run_experiment(
name, base_config, readme, distributions=None, durations=None,
name_variables=None, alg_configs=None, env_configs=None, late_config=None):
name = sanitize(name)
durations = durations or {}
parser = argparse.ArgumentParser()
if env_configs is not None:
parser.add_argument('env')
if alg_configs is not None:
parser.add_argument('alg')
parser.add_argument("duration", choices=list(durations.keys()) + ["local"], default="local", nargs="?")
args, _ = parser.parse_known_args()
config = DEFAULT_CONFIG.copy()
config.update(base_config)
if env_configs is not None:
env_config = env_configs[args.env]
config.update(env_config)
if alg_configs is not None:
alg_config = alg_configs[args.alg]
config.update(alg_config)
if late_config is not None:
config.update(late_config)
config.update_from_command_line()
env_name = sanitize(config.get('env_name', ''))
if name:
config.env_name = "{}_env={}".format(name, env_name)
else:
config.env_name = "env={}".format(env_name)
alg_name = sanitize(config.get("alg_name", ""))
if args.duration == "local":
config.exp_name = "alg={}".format(alg_name)
with config:
return training_loop()
run_kwargs = Config(
kind="slurm",
pmem=5000,
ignore_gpu=False,
)
duration_args = durations[args.duration]
if 'config' in duration_args:
config.update(duration_args['config'])
del duration_args['config']
run_kwargs.update(durations[args.duration])
run_kwargs.update_from_command_line()
if name_variables is not None:
name_variables_str = "_".join(
"{}={}".format(sanitize(str(k)), sanitize(str(getattr(config, k))))
for k in name_variables.split(","))
config.env_name = "{}_{}".format(config.env_name, name_variables_str)
exp_name = "{}_alg={}_duration={}".format(config.env_name, alg_name, args.duration)
build_and_submit(name=exp_name, config=config, distributions=distributions, **run_kwargs)
