def testMaxConcurrentSuggestions(self):
"""Checks that next_trials() supports throttling."""
experiment_spec = {
"run": "PPO",
"num_samples": 6,
}
experiments = [Experiment.from_json("test", experiment_spec)]
searcher = _MockSuggestionAlgorithm(max_concurrent=4)
searcher.add_configurations(experiments)
trials = searcher.next_trials()
self.assertEqual(len(trials), 4)
self.assertEqual(searcher.next_trials(), [])
finished_trial = trials.pop()
searcher.on_trial_complete(finished_trial.trial_id)
self.assertEqual(len(searcher.next_trials()), 1)
finished_trial = trials.pop()
searcher.on_trial_complete(finished_trial.trial_id)
finished_trial = trials.pop()
searcher.on_trial_complete(finished_trial.trial_id)
finished_trial = trials.pop()
searcher.on_trial_complete(finished_trial.trial_id)
self.assertEqual(len(searcher.next_trials()), 1)
self.assertEqual(len(searcher.next_trials()), 0)
def run_experiments(experiments,
scheduler=None,
with_server=False,
server_port=TuneServer.DEFAULT_PORT,
verbose=True):
"""Tunes experiments.
Args:
experiments (Experiment | list | dict): Experiments to run.
scheduler (TrialScheduler): Scheduler for executing
the experiment. Choose among FIFO (default), MedianStopping,
AsyncHyperBand, HyperBand, or HyperOpt.
with_server (bool): Starts a background Tune server. Needed for
using the Client API.
server_port (int): Port number for launching TuneServer.
verbose (bool): How much output should be printed for each trial.
"""
if scheduler is None:
scheduler = FIFOScheduler()
runner = TrialRunner(
scheduler,
launch_web_server=with_server,
server_port=server_port,
verbose=verbose)
exp_list = experiments
if isinstance(experiments, Experiment):
exp_list = [experiments]
elif type(experiments) is dict:
exp_list = [
Experiment.from_json(name, spec)
for name, spec in experiments.items()
]
if (type(exp_list) is list
and all(isinstance(exp, Experiment) for exp in exp_list)):
for experiment in exp_list:
scheduler.add_experiment(experiment, runner)
else:
raise TuneError("Invalid argument: {}".format(experiments))
print(runner.debug_string(max_debug=99999))
last_debug = 0
while not runner.is_finished():
runner.step()
if time.time() - last_debug > DEBUG_PRINT_INTERVAL:
print(runner.debug_string())
last_debug = time.time()
print(runner.debug_string(max_debug=99999))
for trial in runner.get_trials():
# TODO(rliaw): What about errored?
if trial.status != Trial.TERMINATED:
raise TuneError("Trial did not complete", trial)
wait_for_log_sync()
return runner.get_trials()
def create_searcher():
class TestSuggestion(Searcher):
def __init__(self, index):
self.index = index
self.returned_result = []
super().__init__(metric="episode_reward_mean", mode="max")
def suggest(self, trial_id):
self.index += 1
return {"test_variable": self.index}
def on_trial_complete(self, trial_id, result=None, **kwargs):
self.returned_result.append(result)
def save(self, checkpoint_path):
with open(checkpoint_path, "wb") as f:
pickle.dump(self.__dict__, f)
def restore(self, checkpoint_path):
with open(checkpoint_path, "rb") as f:
self.__dict__.update(pickle.load(f))
searcher = TestSuggestion(0)
searcher = ConcurrencyLimiter(searcher, max_concurrent=2)
searcher = Repeater(searcher, repeat=3, set_index=False)
search_alg = SearchGenerator(searcher)
experiment_spec = {
"run": "__fake",
"num_samples": 20,
"stop": {"training_iteration": 2},
}
experiments = [Experiment.from_json("test", experiment_spec)]
search_alg.add_configurations(experiments)
return search_alg
def testReporterNoUsage(self):
def run_task(config, reporter):
print("hello")
experiment = Experiment(run=run_task, name="ray_crash_repro")
[trial] = ray.tune.run(experiment).trials
print(trial.last_result)
self.assertEqual(trial.last_result[DONE], True)
def testExperimentList(self):
def train(config, reporter):
for i in range(100):
reporter(timesteps_total=i)
register_trainable("f1", train)
exp1 = Experiment(**{
"name": "foo",
"run": "f1",
})
exp2 = Experiment(**{
"name": "bar",
"run": "f1",
})
trials = run_experiments([exp1, exp2])
for trial in trials:
self.assertEqual(trial.status, Trial.TERMINATED)
self.assertEqual(trial.last_result[TIMESTEPS_TOTAL], 99)
def testRegisterTrainable(self):
def train(config, reporter):
pass
class A:
pass
class B(Trainable):
pass
register_trainable("foo", train)
Experiment("test", train)
register_trainable("foo", B)
Experiment("test", B)
self.assertRaises(TypeError, lambda: register_trainable("foo", B()))
self.assertRaises(TuneError, lambda: Experiment("foo", B()))
self.assertRaises(TypeError, lambda: register_trainable("foo", A))
self.assertRaises(TypeError, lambda: Experiment("foo", A))
def testConvertExperimentList(self):
exp1 = Experiment(**{
"name": "foo",
"run": "f1",
"config": {
"script_min_iter_time_s": 0
}
})
result = convert_to_experiment_list([exp1, exp1])
self.assertEqual(len(result), 2)
self.assertEqual(type(result), list)
def testRemoteRunExperiments(self):
def train(config, reporter):
for i in range(100):
reporter(timesteps_total=i)
register_trainable("f1", train)
exp1 = Experiment(**{
"name": "foo",
"run": "f1",
})
[trial] = run_experiments(exp1, _remote=True)
self.assertEqual(trial.status, Trial.TERMINATED)
self.assertEqual(trial.last_result[TIMESTEPS_TOTAL], 99)
def testCallbackSetupBackwardsCompatible(self, mocked_warning_method):
class NoExperimentInSetupCallback(Callback):
# Old method definition didn't take in **experiment.public_spec
def setup(self):
return
callback = NoExperimentInSetupCallback()
trial_runner = TrialRunner(callbacks=[callback])
trial_runner.setup_experiments(
experiments=[Experiment("", lambda x: x)], total_num_samples=1)
mocked_warning_method.assert_called_once()
self.assertIn("Please update",
mocked_warning_method.call_args_list[0][0][0])
def testSearchAlgFinishes(self):
"""Empty SearchAlg changing state in `next_trials` does not crash."""
os.environ["TUNE_MAX_PENDING_TRIALS_PG"] = "1"
class FinishFastAlg(_MockSuggestionAlgorithm):
_index = 0
def next_trial(self):
spec = self._experiment.spec
trial = None
if self._index < spec["num_samples"]:
trial = Trial(
spec.get("run"), stopping_criterion=spec.get("stop"))
self._index += 1
if self._index > 4:
self.set_finished()
return trial
def suggest(self, trial_id):
return {}
ray.init(num_cpus=2, local_mode=True, include_dashboard=False)
experiment_spec = {
"run": "__fake",
"num_samples": 2,
"stop": {
"training_iteration": 1
}
}
searcher = FinishFastAlg()
experiments = [Experiment.from_json("test", experiment_spec)]
searcher.add_configurations(experiments)
runner = TrialRunner(search_alg=searcher)
self.assertFalse(runner.is_finished())
runner.step() # This launches a new run
runner.step() # This launches a 2nd run
self.assertFalse(searcher.is_finished())
self.assertFalse(runner.is_finished())
runner.step() # This kills the first run
self.assertFalse(searcher.is_finished())
self.assertFalse(runner.is_finished())
runner.step() # This kills the 2nd run
self.assertFalse(searcher.is_finished())
self.assertFalse(runner.is_finished())
runner.step() # this converts self._finished to True
self.assertTrue(searcher.is_finished())
self.assertRaises(TuneError, runner.step)
def testSearchAlgStalled(self):
"""Checks that runner and searcher state is maintained when stalled."""
ray.init(num_cpus=4, num_gpus=2)
experiment_spec = {
"run": "__fake",
"num_samples": 3,
"stop": {
"training_iteration": 1
}
}
experiments = [Experiment.from_json("test", experiment_spec)]
search_alg = _MockSuggestionAlgorithm(max_concurrent=1)
search_alg.add_configurations(experiments)
searcher = search_alg.searcher
runner = TrialRunner(search_alg=search_alg)
runner.step()
trials = runner.get_trials()
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
trials = runner.get_trials()
runner.step()
self.assertEqual(trials[1].status, Trial.RUNNING)
self.assertEqual(len(searcher.live_trials), 1)
searcher.stall = True
runner.step()
self.assertEqual(trials[1].status, Trial.TERMINATED)
self.assertEqual(len(searcher.live_trials), 0)
self.assertTrue(all(trial.is_finished() for trial in trials))
self.assertFalse(search_alg.is_finished())
self.assertFalse(runner.is_finished())
searcher.stall = False
runner.step()
trials = runner.get_trials()
self.assertEqual(trials[2].status, Trial.RUNNING)
self.assertEqual(len(searcher.live_trials), 1)
runner.step()
self.assertEqual(trials[2].status, Trial.TERMINATED)
self.assertEqual(len(searcher.live_trials), 0)
self.assertTrue(search_alg.is_finished())
self.assertTrue(runner.is_finished())
def testExperiment(self):
def train(config, reporter):
for i in range(100):
reporter(timesteps_total=i)
register_trainable("f1", train)
exp1 = Experiment(**{
"name": "foo",
"run": "f1",
"config": {
"script_min_iter_time_s": 0
}
})
[trial] = run_experiments(exp1)
self.assertEqual(trial.status, Trial.TERMINATED)
self.assertEqual(trial.last_result.timesteps_total, 99)
def testSearchAlgFinishes(self):
"""Empty SearchAlg changing state in `next_trials` does not crash."""
class FinishFastAlg(SuggestionAlgorithm):
_index = 0
def next_trials(self):
trials = []
self._index += 1
for trial in self._trial_generator:
trials += [trial]
break
if self._index > 4:
self._finished = True
return trials
def _suggest(self, trial_id):
return {}
ray.init(num_cpus=2)
experiment_spec = {
"run": "__fake",
"num_samples": 2,
"stop": {
"training_iteration": 1
}
}
searcher = FinishFastAlg()
experiments = [Experiment.from_json("test", experiment_spec)]
searcher.add_configurations(experiments)
runner = TrialRunner(search_alg=searcher)
self.assertFalse(runner.is_finished())
runner.step() # This launches a new run
runner.step() # This launches a 2nd run
self.assertFalse(searcher.is_finished())
self.assertFalse(runner.is_finished())
runner.step() # This kills the first run
self.assertFalse(searcher.is_finished())
self.assertFalse(runner.is_finished())
runner.step() # This kills the 2nd run
self.assertFalse(searcher.is_finished())
self.assertFalse(runner.is_finished())
runner.step() # this converts self._finished to True
self.assertTrue(searcher.is_finished())
self.assertRaises(TuneError, runner.step)
def testSearchAlgNotification(self):
"""Checks notification of trial to the Search Algorithm."""
os.environ["TUNE_RESULT_BUFFER_LENGTH"] = "1" # Don't finish early
os.environ["TUNE_MAX_PENDING_TRIALS_PG"] = "1"
ray.init(num_cpus=4, num_gpus=2)
experiment_spec = {"run": "__fake", "stop": {"training_iteration": 2}}
experiments = [Experiment.from_json("test", experiment_spec)]
search_alg = _MockSuggestionAlgorithm()
searcher = search_alg.searcher
search_alg.add_configurations(experiments)
runner = TrialRunner(search_alg=search_alg)
while not runner.is_finished():
runner.step()
self.assertEqual(searcher.counter["result"], 1)
self.assertEqual(searcher.counter["complete"], 1)
def testSearchAlgFinished(self):
"""Checks that SearchAlg is Finished before all trials are done."""
ray.init(num_cpus=4, num_gpus=2)
experiment_spec = {"run": "__fake", "stop": {"training_iteration": 1}}
experiments = [Experiment.from_json("test", experiment_spec)]
searcher = _MockSuggestionAlgorithm(experiments, max_concurrent=10)
runner = TrialRunner(search_alg=searcher)
runner.step()
trials = runner.get_trials()
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertTrue(searcher.is_finished())
self.assertFalse(runner.is_finished())
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(len(searcher.live_trials), 0)
self.assertTrue(searcher.is_finished())
self.assertTrue(runner.is_finished())
def testRemoteRunExperimentsInClient(self):
ray.init()
assert not ray.util.client.ray.is_connected()
with ray_start_client_server():
assert ray.util.client.ray.is_connected()
def train(config, reporter):
for i in range(100):
reporter(timesteps_total=i)
register_trainable("f1", train)
exp1 = Experiment(**{
"name": "foo",
"run": "f1",
})
[trial] = run_experiments(exp1)
self.assertEqual(trial.status, Trial.TERMINATED)
self.assertEqual(trial.last_result[TIMESTEPS_TOTAL], 99)
def testSearchAlgNotification(self):
"""Checks notification of trial to the Search Algorithm."""
ray.init(num_cpus=4, num_gpus=2)
experiment_spec = {"run": "__fake", "stop": {"training_iteration": 2}}
experiments = [Experiment.from_json("test", experiment_spec)]
searcher = _MockSuggestionAlgorithm(experiments, max_concurrent=10)
runner = TrialRunner(search_alg=searcher)
runner.step()
trials = runner.get_trials()
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(searcher.counter["result"], 1)
self.assertEqual(searcher.counter["complete"], 1)
def get_tune_experiment(config, agent, episodes, root_dir, is_schedule):
scheduler = None
agent_class = get_agent(agent)
ex_conf = {}
ex_conf["name"] = agent
ex_conf["run"] = agent_class
ex_conf["local_dir"] = config["env_config"]["output_dir"]
ex_conf["stop"] = {"episodes_total": episodes}
if is_schedule:
ex_conf["stop"] = {"time_total_s": 300}
ex_conf["num_samples"] = 2
config["env_config"]["parallel_envs"] = True
# custom changes to experiment
log.info("Performing tune experiment")
config, scheduler = set_tuning_parameters(agent, config)
ex_conf["config"] = config
experiment = Experiment(**ex_conf)
return experiment, scheduler
def testSearchAlgSchedulerInteraction(self):
"""Checks that TrialScheduler killing trial will notify SearchAlg."""
class _MockScheduler(FIFOScheduler):
def on_trial_result(self, *args, **kwargs):
return TrialScheduler.STOP
ray.init(num_cpus=4, num_gpus=2)
experiment_spec = {"run": "__fake", "stop": {"training_iteration": 2}}
experiments = [Experiment.from_json("test", experiment_spec)]
searcher = _MockSuggestionAlgorithm(experiments, max_concurrent=10)
runner = TrialRunner(search_alg=searcher, scheduler=_MockScheduler())
runner.step()
trials = runner.get_trials()
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertTrue(searcher.is_finished())
self.assertFalse(runner.is_finished())
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(len(searcher.live_trials), 0)
self.assertTrue(searcher.is_finished())
self.assertTrue(runner.is_finished())
def get_tune_experiment(config, agent):
SCHEDULE = False
scheduler = None
agent_class = get_agent(agent)
ex_conf = {}
ex_conf["name"] = agent
ex_conf["run"] = agent_class
ex_conf["local_dir"] = ARGS.output_dir
ex_conf["stop"] = {"timesteps_total": ARGS.timesteps}
ex_conf["checkpoint_freq"] = ARGS.checkpoint_freq
ex_conf["checkpoint_at_end"] = True
ex_conf["restore"] = ARGS.restore
if SCHEDULE:
ex_conf["stop"] = {"time_total_s": ARGS.timesteps / 2}
ex_conf["num_samples"] = 2
config["env_config"]["parallel_envs"] = True
# custom changes to experiment
log.info("Performing tune experiment")
config, scheduler = set_tuning_parameters(agent, config)
ex_conf["config"] = config
experiment = Experiment(**ex_conf)
return experiment, scheduler
def testSearchAlgSchedulerEarlyStop(self):
"""Early termination notif to Searcher can be turned off."""
class _MockScheduler(FIFOScheduler):
def on_trial_result(self, *args, **kwargs):
return TrialScheduler.STOP
ray.init(num_cpus=4, num_gpus=2)
experiment_spec = {"run": "__fake", "stop": {"training_iteration": 2}}
experiments = [Experiment.from_json("test", experiment_spec)]
searcher = _MockSuggestionAlgorithm(use_early_stopped_trials=True)
searcher.add_configurations(experiments)
runner = TrialRunner(search_alg=searcher, scheduler=_MockScheduler())
runner.step()
runner.step()
self.assertEqual(len(searcher.final_results), 1)
searcher = _MockSuggestionAlgorithm(use_early_stopped_trials=False)
searcher.add_configurations(experiments)
runner = TrialRunner(search_alg=searcher, scheduler=_MockScheduler())
runner.step()
runner.step()
self.assertEqual(len(searcher.final_results), 0)
def testSearchAlgNotification(self):
"""Checks notification of trial to the Search Algorithm."""
os.environ["TUNE_RESULT_BUFFER_LENGTH"] = "1" # Don't finish early
ray.init(num_cpus=4, num_gpus=2)
experiment_spec = {"run": "__fake", "stop": {"training_iteration": 2}}
experiments = [Experiment.from_json("test", experiment_spec)]
search_alg = _MockSuggestionAlgorithm()
searcher = search_alg.searcher
search_alg.add_configurations(experiments)
runner = TrialRunner(search_alg=search_alg)
runner.step()
trials = runner.get_trials()
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(searcher.counter["result"], 1)
self.assertEqual(searcher.counter["complete"], 1)
def testSearchAlgFinishes(self):
"""SearchAlg changing state in `next_trials` does not crash."""
class FinishFastAlg(SuggestionAlgorithm):
def next_trials(self):
self._finished = True
return []
ray.init(num_cpus=4, num_gpus=2)
experiment_spec = {
"run": "__fake",
"num_samples": 3,
"stop": {
"training_iteration": 1
}
}
searcher = FinishFastAlg()
experiments = [Experiment.from_json("test", experiment_spec)]
searcher.add_configurations(experiments)
runner = TrialRunner(search_alg=searcher)
runner.step() # This should not fail
self.assertTrue(searcher.is_finished())
self.assertTrue(runner.is_finished())
def run(
run_or_experiment,
name=None,
metric=None,
mode=None,
stop=None,
time_budget_s=None,
config=None,
resources_per_trial=None,
num_samples=1,
local_dir=None,
search_alg=None,
scheduler=None,
keep_checkpoints_num=None,
checkpoint_score_attr=None,
checkpoint_freq=0,
checkpoint_at_end=False,
verbose=2,
progress_reporter=None,
loggers=None,
log_to_file=False,
trial_name_creator=None,
trial_dirname_creator=None,
sync_config=None,
export_formats=None,
max_failures=0,
fail_fast=False,
restore=None,
server_port=None,
resume=False,
reuse_actors=False,
trial_executor=None,
raise_on_failed_trial=True,
# Deprecated args
ray_auto_init=None,
run_errored_only=None,
queue_trials=None,
global_checkpoint_period=None,
with_server=None,
upload_dir=None,
sync_to_cloud=None,
sync_to_driver=None,
sync_on_checkpoint=None,
):
"""Executes training.
Examples:
.. code-block:: python
# Run 10 trials (each trial is one instance of a Trainable). Tune runs
# in parallel and automatically determines concurrency.
tune.run(trainable, num_samples=10)
# Run 1 trial, stop when trial has reached 10 iterations
tune.run(my_trainable, stop={"training_iteration": 10})
# automatically retry failed trials up to 3 times
tune.run(my_trainable, stop={"training_iteration": 10}, max_failures=3)
# Run 1 trial, search over hyperparameters, stop after 10 iterations.
space = {"lr": tune.uniform(0, 1), "momentum": tune.uniform(0, 1)}
tune.run(my_trainable, config=space, stop={"training_iteration": 10})
# Resumes training if a previous machine crashed
tune.run(my_trainable, config=space,
local_dir=<path/to/dir>, resume=True)
# Rerun ONLY failed trials after an experiment is finished.
tune.run(my_trainable, config=space,
local_dir=<path/to/dir>, resume="ERRORED_ONLY")
Args:
run_or_experiment (function | class | str | :class:`Experiment`): If
function|class|str, this is the algorithm or model to train.
This may refer to the name of a built-on algorithm
(e.g. RLLib's DQN or PPO), a user-defined trainable
function or class, or the string identifier of a
trainable function or class registered in the tune registry.
If Experiment, then Tune will execute training based on
Experiment.spec. If you want to pass in a Python lambda, you
will need to first register the function:
``tune.register_trainable("lambda_id", lambda x: ...)``. You can
then use ``tune.run("lambda_id")``.
metric (str): Metric to optimize. This metric should be reported
with `tune.report()`. If set, will be passed to the search
algorithm and scheduler.
mode (str): Must be one of [min, max]. Determines whether objective is
minimizing or maximizing the metric attribute. If set, will be
passed to the search algorithm and scheduler.
name (str): Name of experiment.
stop (dict | callable | :class:`Stopper`): Stopping criteria. If dict,
the keys may be any field in the return result of 'train()',
whichever is reached first. If function, it must take (trial_id,
result) as arguments and return a boolean (True if trial should be
stopped, False otherwise). This can also be a subclass of
``ray.tune.Stopper``, which allows users to implement
custom experiment-wide stopping (i.e., stopping an entire Tune
run based on some time constraint).
time_budget_s (int|float|datetime.timedelta): Global time budget in
seconds after which all trials are stopped. Can also be a
``datetime.timedelta`` object.
config (dict): Algorithm-specific configuration for Tune variant
generation (e.g. env, hyperparams). Defaults to empty dict.
Custom search algorithms may ignore this.
resources_per_trial (dict): Machine resources to allocate per trial,
e.g. ``{"cpu": 64, "gpu": 8}``. Note that GPUs will not be
assigned unless you specify them here. Defaults to 1 CPU and 0
GPUs in ``Trainable.default_resource_request()``.
num_samples (int): Number of times to sample from the
hyperparameter space. Defaults to 1. If `grid_search` is
provided as an argument, the grid will be repeated
`num_samples` of times.
local_dir (str): Local dir to save training results to.
Defaults to ``~/ray_results``.
search_alg (Searcher): Search algorithm for optimization.
scheduler (TrialScheduler): Scheduler for executing
the experiment. Choose among FIFO (default), MedianStopping,
AsyncHyperBand, HyperBand and PopulationBasedTraining. Refer to
ray.tune.schedulers for more options.
keep_checkpoints_num (int): Number of checkpoints to keep. A value of
`None` keeps all checkpoints. Defaults to `None`. If set, need
to provide `checkpoint_score_attr`.
checkpoint_score_attr (str): Specifies by which attribute to rank the
best checkpoint. Default is increasing order. If attribute starts
with `min-` it will rank attribute in decreasing order, i.e.
`min-validation_loss`.
checkpoint_freq (int): How many training iterations between
checkpoints. A value of 0 (default) disables checkpointing.
This has no effect when using the Functional Training API.
checkpoint_at_end (bool): Whether to checkpoint at the end of the
experiment regardless of the checkpoint_freq. Default is False.
This has no effect when using the Functional Training API.
verbose (int): 0, 1, or 2. Verbosity mode. 0 = silent,
1 = only status updates, 2 = status and trial results.
progress_reporter (ProgressReporter): Progress reporter for reporting
intermediate experiment progress. Defaults to CLIReporter if
running in command-line, or JupyterNotebookReporter if running in
a Jupyter notebook.
loggers (list): List of logger creators to be used with
each Trial. If None, defaults to ray.tune.logger.DEFAULT_LOGGERS.
See `ray/tune/logger.py`.
log_to_file (bool|str|Sequence): Log stdout and stderr to files in
Tune's trial directories. If this is `False` (default), no files
are written. If `true`, outputs are written to `trialdir/stdout`
and `trialdir/stderr`, respectively. If this is a single string,
this is interpreted as a file relative to the trialdir, to which
both streams are written. If this is a Sequence (e.g. a Tuple),
it has to have length 2 and the elements indicate the files to
which stdout and stderr are written, respectively.
trial_name_creator (Callable[[Trial], str]): Optional function
for generating the trial string representation.
trial_dirname_creator (Callable[[Trial], str]): Function
for generating the trial dirname. This function should take
in a Trial object and return a string representing the
name of the directory. The return value cannot be a path.
sync_config (SyncConfig): Configuration object for syncing. See
tune.SyncConfig.
export_formats (list): List of formats that exported at the end of
the experiment. Default is None.
max_failures (int): Try to recover a trial at least this many times.
Ray will recover from the latest checkpoint if present.
Setting to -1 will lead to infinite recovery retries.
Setting to 0 will disable retries. Defaults to 3.
fail_fast (bool | str): Whether to fail upon the first error.
If fail_fast='raise' provided, Tune will automatically
raise the exception received by the Trainable. fail_fast='raise'
can easily leak resources and should be used with caution (it
is best used with `ray.init(local_mode=True)`).
restore (str): Path to checkpoint. Only makes sense to set if
running 1 trial. Defaults to None.
server_port (int): Port number for launching TuneServer.
resume (str|bool): One of "LOCAL", "REMOTE", "PROMPT", "ERRORED_ONLY",
or bool. LOCAL/True restores the checkpoint from the
local_checkpoint_dir, determined
by `name` and `local_dir`. REMOTE restores the checkpoint
from remote_checkpoint_dir. PROMPT provides CLI feedback.
False forces a new experiment. ERRORED_ONLY resets and reruns
ERRORED trials upon resume - previous trial artifacts will
be left untouched. If resume is set but checkpoint does not exist,
ValueError will be thrown.
reuse_actors (bool): Whether to reuse actors between different trials
when possible. This can drastically speed up experiments that start
and stop actors often (e.g., PBT in time-multiplexing mode). This
requires trials to have the same resource requirements.
trial_executor (TrialExecutor): Manage the execution of trials.
raise_on_failed_trial (bool): Raise TuneError if there exists failed
trial (of ERROR state) when the experiments complete.
Returns:
ExperimentAnalysis: Object for experiment analysis.
Raises:
TuneError: Any trials failed and `raise_on_failed_trial` is True.
"""
if global_checkpoint_period:
raise ValueError("global_checkpoint_period is deprecated. Set env var "
"'TUNE_GLOBAL_CHECKPOINT_S' instead.")
if queue_trials:
raise ValueError(
"queue_trials is deprecated. "
"Set env var 'TUNE_DISABLE_QUEUE_TRIALS=1' instead to "
"disable queuing behavior.")
if ray_auto_init:
raise ValueError("ray_auto_init is deprecated. "
"Set env var 'TUNE_DISABLE_AUTO_INIT=1' instead or "
"call 'ray.init' before calling 'tune.run'.")
if with_server:
raise ValueError(
"with_server is deprecated. It is now enabled by default "
"if 'server_port' is not None.")
if sync_on_checkpoint or sync_to_cloud or sync_to_driver or upload_dir:
raise ValueError(
"sync_on_checkpoint / sync_to_cloud / sync_to_driver / "
"upload_dir must now be set via `tune.run("
"sync_config=SyncConfig(...)`. See `ray.tune.SyncConfig` for "
"more details.")
if mode and mode not in ["min", "max"]:
raise ValueError(
"The `mode` parameter passed to `tune.run()` has to be one of "
"['min', 'max']")
config = config or {}
sync_config = sync_config or SyncConfig()
set_sync_periods(sync_config)
trial_executor = trial_executor or RayTrialExecutor(
reuse_actors=reuse_actors)
if isinstance(run_or_experiment, list):
experiments = run_or_experiment
else:
experiments = [run_or_experiment]
for i, exp in enumerate(experiments):
if not isinstance(exp, Experiment):
experiments[i] = Experiment(
name=name,
run=exp,
stop=stop,
time_budget_s=time_budget_s,
config=config,
resources_per_trial=resources_per_trial,
num_samples=num_samples,
local_dir=local_dir,
upload_dir=sync_config.upload_dir,
sync_to_driver=sync_config.sync_to_driver,
trial_name_creator=trial_name_creator,
trial_dirname_creator=trial_dirname_creator,
loggers=loggers,
log_to_file=log_to_file,
checkpoint_freq=checkpoint_freq,
checkpoint_at_end=checkpoint_at_end,
sync_on_checkpoint=sync_config.sync_on_checkpoint,
keep_checkpoints_num=keep_checkpoints_num,
checkpoint_score_attr=checkpoint_score_attr,
export_formats=export_formats,
max_failures=max_failures,
restore=restore)
else:
logger.debug("Ignoring some parameters passed into tune.run.")
if sync_config.sync_to_cloud:
for exp in experiments:
assert exp.remote_checkpoint_dir, (
"Need `upload_dir` if `sync_to_cloud` given.")
if fail_fast and max_failures != 0:
raise ValueError("max_failures must be 0 if fail_fast=True.")
if issubclass(type(search_alg), Searcher):
search_alg = SearchGenerator(search_alg)
if not search_alg:
search_alg = BasicVariantGenerator()
if config and not search_alg.set_search_properties(metric, mode, config):
if has_unresolved_values(config):
raise ValueError(
"You passed a `config` parameter to `tune.run()` with "
"unresolved parameters, but the search algorithm was already "
"instantiated with a search space. Make sure that `config` "
"does not contain any more parameter definitions - include "
"them in the search algorithm's search space if necessary.")
scheduler = scheduler or FIFOScheduler()
if not scheduler.set_search_properties(metric, mode):
raise ValueError(
"You passed a `metric` or `mode` argument to `tune.run()`, but "
"the scheduler you are using was already instantiated with their "
"own `metric` and `mode` parameters. Either remove the arguments "
"from your scheduler or from your call to `tune.run()`")
runner = TrialRunner(
search_alg=search_alg,
scheduler=scheduler,
local_checkpoint_dir=experiments[0].checkpoint_dir,
remote_checkpoint_dir=experiments[0].remote_checkpoint_dir,
sync_to_cloud=sync_config.sync_to_cloud,
stopper=experiments[0].stopper,
resume=resume,
server_port=server_port,
verbose=bool(verbose > 1),
fail_fast=fail_fast,
trial_executor=trial_executor)
if not runner.resumed:
for exp in experiments:
search_alg.add_configurations([exp])
else:
logger.info("TrialRunner resumed, ignoring new add_experiment.")
if progress_reporter is None:
if IS_NOTEBOOK:
progress_reporter = JupyterNotebookReporter(overwrite=verbose < 2)
else:
progress_reporter = CLIReporter()
# User Warning for GPUs
if trial_executor.has_gpus():
if isinstance(resources_per_trial,
dict) and "gpu" in resources_per_trial:
# "gpu" is manually set.
pass
elif _check_default_resources_override(experiments[0].run_identifier):
# "default_resources" is manually overriden.
pass
else:
logger.warning("Tune detects GPUs, but no trials are using GPUs. "
"To enable trials to use GPUs, set "
"tune.run(resources_per_trial={'gpu': 1}...) "
"which allows Tune to expose 1 GPU to each trial. "
"You can also override "
"`Trainable.default_resource_request` if using the "
"Trainable API.")
while not runner.is_finished():
runner.step()
if verbose:
_report_progress(runner, progress_reporter)
try:
runner.checkpoint(force=True)
except Exception as e:
logger.warning(f"Trial Runner checkpointing failed: {str(e)}")
if verbose:
_report_progress(runner, progress_reporter, done=True)
wait_for_sync()
runner.cleanup_trials()
incomplete_trials = []
for trial in runner.get_trials():
if trial.status != Trial.TERMINATED:
incomplete_trials += [trial]
if incomplete_trials:
if raise_on_failed_trial:
raise TuneError("Trials did not complete", incomplete_trials)
else:
logger.error("Trials did not complete: %s", incomplete_trials)
trials = runner.get_trials()
return ExperimentAnalysis(runner.checkpoint_file,
trials=trials,
default_metric=metric,
default_mode=mode)
def test_cluster_interrupt(start_connected_cluster, tmpdir):
"""Tests run_experiment on cluster shutdown with actual interrupt.
This is an end-to-end test.
"""
cluster = start_connected_cluster
dirpath = str(tmpdir)
# Needs to be in scope for pytest
class _Mock(tune.Trainable):
"""Finishes on the 4th iteration."""
def _setup(self, config):
self.state = {"hi": 0}
def _train(self):
self.state["hi"] += 1
time.sleep(0.5)
return {"done": self.state["hi"] >= 4}
def _save(self, path):
return self.state
def _restore(self, state):
self.state = state
# Removes indent from class.
reformatted = "\n".join(line[4:] if len(line) else line
for line in inspect.getsource(_Mock).split("\n"))
script = """
import time
import ray
from ray import tune
ray.init(redis_address="{redis_address}")
{fail_class_code}
kwargs = dict(
run={fail_class},
stop=dict(training_iteration=5),
local_dir="{checkpoint_dir}",
checkpoint_freq=1,
max_failures=1)
tune.run_experiments(
dict(experiment=kwargs),
raise_on_failed_trial=False)
""".format(
redis_address=cluster.redis_address,
checkpoint_dir=dirpath,
fail_class_code=reformatted,
fail_class=_Mock.__name__)
run_string_as_driver_nonblocking(script)
# Wait until the right checkpoint is saved.
# The trainable returns every 0.5 seconds, so this should not miss
# the checkpoint.
metadata_checkpoint_dir = os.path.join(dirpath, "experiment")
for i in range(50):
if TrialRunner.checkpoint_exists(metadata_checkpoint_dir):
# Inspect the internal trialrunner
runner = TrialRunner.restore(metadata_checkpoint_dir)
trials = runner.get_trials()
last_res = trials[0].last_result
if last_res and last_res.get("training_iteration") == 3:
break
time.sleep(0.2)
if not TrialRunner.checkpoint_exists(metadata_checkpoint_dir):
raise RuntimeError("Checkpoint file didn't appear.")
ray.shutdown()
cluster.shutdown()
cluster = _start_new_cluster()
Experiment._register_if_needed(_Mock)
# Inspect the internal trialrunner
runner = TrialRunner.restore(metadata_checkpoint_dir)
trials = runner.get_trials()
assert trials[0].last_result["training_iteration"] == 3
assert trials[0].status == Trial.PENDING
# Restore properly from checkpoint
trials2 = tune.run_experiments(
{
"experiment": {
"run": _Mock,
"local_dir": dirpath,
"checkpoint_freq": 1
}
},
resume=True,
raise_on_failed_trial=False)
assert all(t.status == Trial.TERMINATED for t in trials2)
assert {t.trial_id for t in trials2} == {t.trial_id for t in trials}
cluster.shutdown()
def run(
run_or_experiment: Union[str, Callable, Type],
name: Optional[str] = None,
metric: Optional[str] = None,
mode: Optional[str] = None,
stop: Union[None, Mapping, Stopper, Callable[[str, Mapping],
bool]] = None,
time_budget_s: Union[None, int, float, datetime.timedelta] = None,
config: Optional[Dict[str, Any]] = None,
resources_per_trial: Union[None, Mapping[str, Union[
float, int, Mapping]], PlacementGroupFactory] = None,
num_samples: int = 1,
local_dir: Optional[str] = None,
search_alg: Optional[Union[Searcher, SearchAlgorithm]] = None,
scheduler: Optional[TrialScheduler] = None,
keep_checkpoints_num: Optional[int] = None,
checkpoint_score_attr: Optional[str] = None,
checkpoint_freq: int = 0,
checkpoint_at_end: bool = False,
verbose: Union[int, Verbosity] = Verbosity.V3_TRIAL_DETAILS,
progress_reporter: Optional[ProgressReporter] = None,
log_to_file: bool = False,
trial_name_creator: Optional[Callable[[Trial], str]] = None,
trial_dirname_creator: Optional[Callable[[Trial], str]] = None,
sync_config: Optional[SyncConfig] = None,
export_formats: Optional[Sequence] = None,
max_failures: int = 0,
fail_fast: bool = False,
restore: Optional[str] = None,
server_port: Optional[int] = None,
resume: bool = False,
queue_trials: bool = False,
reuse_actors: bool = False,
trial_executor: Optional[RayTrialExecutor] = None,
raise_on_failed_trial: bool = True,
callbacks: Optional[Sequence[Callback]] = None,
# Deprecated args
loggers: Optional[Sequence[Type[Logger]]] = None,
ray_auto_init: Optional = None,
run_errored_only: Optional = None,
global_checkpoint_period: Optional = None,
with_server: Optional = None,
upload_dir: Optional = None,
sync_to_cloud: Optional = None,
sync_to_driver: Optional = None,
sync_on_checkpoint: Optional = None,
_remote: bool = None,
) -> ExperimentAnalysis:
"""Executes training.
When a SIGINT signal is received (e.g. through Ctrl+C), the tuning run
will gracefully shut down and checkpoint the latest experiment state.
Sending SIGINT again (or SIGKILL/SIGTERM instead) will skip this step.
Examples:
.. code-block:: python
# Run 10 trials (each trial is one instance of a Trainable). Tune runs
# in parallel and automatically determines concurrency.
tune.run(trainable, num_samples=10)
# Run 1 trial, stop when trial has reached 10 iterations
tune.run(my_trainable, stop={"training_iteration": 10})
# automatically retry failed trials up to 3 times
tune.run(my_trainable, stop={"training_iteration": 10}, max_failures=3)
# Run 1 trial, search over hyperparameters, stop after 10 iterations.
space = {"lr": tune.uniform(0, 1), "momentum": tune.uniform(0, 1)}
tune.run(my_trainable, config=space, stop={"training_iteration": 10})
# Resumes training if a previous machine crashed
tune.run(my_trainable, config=space,
local_dir=<path/to/dir>, resume=True)
# Rerun ONLY failed trials after an experiment is finished.
tune.run(my_trainable, config=space,
local_dir=<path/to/dir>, resume="ERRORED_ONLY")
Args:
run_or_experiment (function | class | str | :class:`Experiment`): If
function|class|str, this is the algorithm or model to train.
This may refer to the name of a built-on algorithm
(e.g. RLLib's DQN or PPO), a user-defined trainable
function or class, or the string identifier of a
trainable function or class registered in the tune registry.
If Experiment, then Tune will execute training based on
Experiment.spec. If you want to pass in a Python lambda, you
will need to first register the function:
``tune.register_trainable("lambda_id", lambda x: ...)``. You can
then use ``tune.run("lambda_id")``.
metric (str): Metric to optimize. This metric should be reported
with `tune.report()`. If set, will be passed to the search
algorithm and scheduler.
mode (str): Must be one of [min, max]. Determines whether objective is
minimizing or maximizing the metric attribute. If set, will be
passed to the search algorithm and scheduler.
name (str): Name of experiment.
stop (dict | callable | :class:`Stopper`): Stopping criteria. If dict,
the keys may be any field in the return result of 'train()',
whichever is reached first. If function, it must take (trial_id,
result) as arguments and return a boolean (True if trial should be
stopped, False otherwise). This can also be a subclass of
``ray.tune.Stopper``, which allows users to implement
custom experiment-wide stopping (i.e., stopping an entire Tune
run based on some time constraint).
time_budget_s (int|float|datetime.timedelta): Global time budget in
seconds after which all trials are stopped. Can also be a
``datetime.timedelta`` object.
config (dict): Algorithm-specific configuration for Tune variant
generation (e.g. env, hyperparams). Defaults to empty dict.
Custom search algorithms may ignore this.
resources_per_trial (dict|PlacementGroupFactory): Machine resources
to allocate per trial, e.g. ``{"cpu": 64, "gpu": 8}``.
Note that GPUs will not be assigned unless you specify them here.
Defaults to 1 CPU and 0 GPUs in
``Trainable.default_resource_request()``. This can also
be a PlacementGroupFactory object wrapping arguments to create a
per-trial placement group.
num_samples (int): Number of times to sample from the
hyperparameter space. Defaults to 1. If `grid_search` is
provided as an argument, the grid will be repeated
`num_samples` of times. If this is -1, (virtually) infinite
samples are generated until a stopping condition is met.
local_dir (str): Local dir to save training results to.
Defaults to ``~/ray_results``.
search_alg (Searcher|SearchAlgorithm): Search algorithm for
optimization.
scheduler (TrialScheduler): Scheduler for executing
the experiment. Choose among FIFO (default), MedianStopping,
AsyncHyperBand, HyperBand and PopulationBasedTraining. Refer to
ray.tune.schedulers for more options.
keep_checkpoints_num (int): Number of checkpoints to keep. A value of
`None` keeps all checkpoints. Defaults to `None`. If set, need
to provide `checkpoint_score_attr`.
checkpoint_score_attr (str): Specifies by which attribute to rank the
best checkpoint. Default is increasing order. If attribute starts
with `min-` it will rank attribute in decreasing order, i.e.
`min-validation_loss`.
checkpoint_freq (int): How many training iterations between
checkpoints. A value of 0 (default) disables checkpointing.
This has no effect when using the Functional Training API.
checkpoint_at_end (bool): Whether to checkpoint at the end of the
experiment regardless of the checkpoint_freq. Default is False.
This has no effect when using the Functional Training API.
verbose (Union[int, Verbosity]): 0, 1, 2, or 3. Verbosity mode.
0 = silent, 1 = only status updates, 2 = status and brief trial
results, 3 = status and detailed trial results. Defaults to 3.
progress_reporter (ProgressReporter): Progress reporter for reporting
intermediate experiment progress. Defaults to CLIReporter if
running in command-line, or JupyterNotebookReporter if running in
a Jupyter notebook.
log_to_file (bool|str|Sequence): Log stdout and stderr to files in
Tune's trial directories. If this is `False` (default), no files
are written. If `true`, outputs are written to `trialdir/stdout`
and `trialdir/stderr`, respectively. If this is a single string,
this is interpreted as a file relative to the trialdir, to which
both streams are written. If this is a Sequence (e.g. a Tuple),
it has to have length 2 and the elements indicate the files to
which stdout and stderr are written, respectively.
trial_name_creator (Callable[[Trial], str]): Optional function
for generating the trial string representation.
trial_dirname_creator (Callable[[Trial], str]): Function
for generating the trial dirname. This function should take
in a Trial object and return a string representing the
name of the directory. The return value cannot be a path.
sync_config (SyncConfig): Configuration object for syncing. See
tune.SyncConfig.
export_formats (list): List of formats that exported at the end of
the experiment. Default is None.
max_failures (int): Try to recover a trial at least this many times.
Ray will recover from the latest checkpoint if present.
Setting to -1 will lead to infinite recovery retries.
Setting to 0 will disable retries. Defaults to 0.
fail_fast (bool | str): Whether to fail upon the first error.
If fail_fast='raise' provided, Tune will automatically
raise the exception received by the Trainable. fail_fast='raise'
can easily leak resources and should be used with caution (it
is best used with `ray.init(local_mode=True)`).
restore (str): Path to checkpoint. Only makes sense to set if
running 1 trial. Defaults to None.
server_port (int): Port number for launching TuneServer.
resume (str|bool): One of "LOCAL", "REMOTE", "PROMPT", "ERRORED_ONLY",
or bool. LOCAL/True restores the checkpoint from the
local_checkpoint_dir, determined
by `name` and `local_dir`. REMOTE restores the checkpoint
from remote_checkpoint_dir. PROMPT provides CLI feedback.
False forces a new experiment. ERRORED_ONLY resets and reruns
ERRORED trials upon resume - previous trial artifacts will
be left untouched. If resume is set but checkpoint does not exist,
ValueError will be thrown.
queue_trials (bool): Whether to queue trials when the cluster does
not currently have enough resources to launch one. This should
be set to True when running on an autoscaling cluster to enable
automatic scale-up.
reuse_actors (bool): Whether to reuse actors between different trials
when possible. This can drastically speed up experiments that start
and stop actors often (e.g., PBT in time-multiplexing mode). This
requires trials to have the same resource requirements.
trial_executor (TrialExecutor): Manage the execution of trials.
raise_on_failed_trial (bool): Raise TuneError if there exists failed
trial (of ERROR state) when the experiments complete.
callbacks (list): List of callbacks that will be called at different
times in the training loop. Must be instances of the
``ray.tune.callback.Callback`` class. If not passed,
`LoggerCallback` and `SyncerCallback` callbacks are automatically
added.
_remote (bool): Whether to run the Tune driver in a remote function.
This is disabled automatically if a custom trial executor is
passed in. This is enabled by default in Ray client mode.
Returns:
ExperimentAnalysis: Object for experiment analysis.
Raises:
TuneError: Any trials failed and `raise_on_failed_trial` is True.
"""
if _remote is None:
_remote = ray.util.client.ray.is_connected()
if _remote is True and trial_executor:
raise ValueError("cannot use custom trial executor")
if not trial_executor or isinstance(trial_executor, RayTrialExecutor):
_ray_auto_init()
if _remote:
return ray.get(
ray.remote(num_cpus=0)(run).remote(
run_or_experiment,
name,
metric,
mode,
stop,
time_budget_s,
config,
resources_per_trial,
num_samples,
local_dir,
search_alg,
scheduler,
keep_checkpoints_num,
checkpoint_score_attr,
checkpoint_freq,
checkpoint_at_end,
verbose,
progress_reporter,
log_to_file,
trial_name_creator,
trial_dirname_creator,
sync_config,
export_formats,
max_failures,
fail_fast,
restore,
server_port,
resume,
queue_trials,
reuse_actors,
trial_executor,
raise_on_failed_trial,
callbacks,
# Deprecated args
loggers,
ray_auto_init,
run_errored_only,
global_checkpoint_period,
with_server,
upload_dir,
sync_to_cloud,
sync_to_driver,
sync_on_checkpoint,
_remote=False))
all_start = time.time()
if global_checkpoint_period:
raise ValueError("global_checkpoint_period is deprecated. Set env var "
"'TUNE_GLOBAL_CHECKPOINT_S' instead.")
if ray_auto_init:
raise ValueError("ray_auto_init is deprecated. "
"Set env var 'TUNE_DISABLE_AUTO_INIT=1' instead or "
"call 'ray.init' before calling 'tune.run'.")
if with_server:
raise ValueError(
"with_server is deprecated. It is now enabled by default "
"if 'server_port' is not None.")
if sync_on_checkpoint or sync_to_cloud or sync_to_driver or upload_dir:
raise ValueError(
"sync_on_checkpoint / sync_to_cloud / sync_to_driver / "
"upload_dir must now be set via `tune.run("
"sync_config=SyncConfig(...)`. See `ray.tune.SyncConfig` for "
"more details.")
if mode and mode not in ["min", "max"]:
raise ValueError(
"The `mode` parameter passed to `tune.run()` has to be one of "
"['min', 'max']")
set_verbosity(verbose)
config = config or {}
sync_config = sync_config or SyncConfig()
set_sync_periods(sync_config)
if num_samples == -1:
num_samples = sys.maxsize
trial_executor = trial_executor or RayTrialExecutor(
reuse_actors=reuse_actors, queue_trials=queue_trials)
if isinstance(run_or_experiment, list):
experiments = run_or_experiment
else:
experiments = [run_or_experiment]
for i, exp in enumerate(experiments):
if not isinstance(exp, Experiment):
experiments[i] = Experiment(
name=name,
run=exp,
stop=stop,
time_budget_s=time_budget_s,
config=config,
resources_per_trial=resources_per_trial,
num_samples=num_samples,
local_dir=local_dir,
upload_dir=sync_config.upload_dir,
sync_to_driver=sync_config.sync_to_driver,
trial_name_creator=trial_name_creator,
trial_dirname_creator=trial_dirname_creator,
log_to_file=log_to_file,
checkpoint_freq=checkpoint_freq,
checkpoint_at_end=checkpoint_at_end,
sync_on_checkpoint=sync_config.sync_on_checkpoint,
keep_checkpoints_num=keep_checkpoints_num,
checkpoint_score_attr=checkpoint_score_attr,
export_formats=export_formats,
max_failures=max_failures,
restore=restore)
else:
logger.debug("Ignoring some parameters passed into tune.run.")
if sync_config.sync_to_cloud:
for exp in experiments:
assert exp.remote_checkpoint_dir, (
"Need `upload_dir` if `sync_to_cloud` given.")
if fail_fast and max_failures != 0:
raise ValueError("max_failures must be 0 if fail_fast=True.")
if issubclass(type(search_alg), Searcher):
search_alg = SearchGenerator(search_alg)
if not search_alg:
search_alg = BasicVariantGenerator()
if config and not search_alg.set_search_properties(metric, mode, config):
if has_unresolved_values(config):
raise ValueError(
"You passed a `config` parameter to `tune.run()` with "
"unresolved parameters, but the search algorithm was already "
"instantiated with a search space. Make sure that `config` "
"does not contain any more parameter definitions - include "
"them in the search algorithm's search space if necessary.")
scheduler = scheduler or FIFOScheduler()
if not scheduler.set_search_properties(metric, mode):
raise ValueError(
"You passed a `metric` or `mode` argument to `tune.run()`, but "
"the scheduler you are using was already instantiated with their "
"own `metric` and `mode` parameters. Either remove the arguments "
"from your scheduler or from your call to `tune.run()`")
# Create syncer callbacks
callbacks = create_default_callbacks(
callbacks, sync_config, metric=metric, loggers=loggers)
runner = TrialRunner(
search_alg=search_alg,
scheduler=scheduler,
local_checkpoint_dir=experiments[0].checkpoint_dir,
remote_checkpoint_dir=experiments[0].remote_checkpoint_dir,
sync_to_cloud=sync_config.sync_to_cloud,
stopper=experiments[0].stopper,
resume=resume,
server_port=server_port,
fail_fast=fail_fast,
trial_executor=trial_executor,
callbacks=callbacks,
metric=metric)
if not runner.resumed:
for exp in experiments:
search_alg.add_configurations([exp])
else:
logger.info("TrialRunner resumed, ignoring new add_experiment.")
if progress_reporter is None:
if IS_NOTEBOOK:
progress_reporter = JupyterNotebookReporter(
overwrite=not has_verbosity(Verbosity.V2_TRIAL_NORM))
else:
progress_reporter = CLIReporter()
if not progress_reporter.set_search_properties(metric, mode):
raise ValueError(
"You passed a `metric` or `mode` argument to `tune.run()`, but "
"the reporter you are using was already instantiated with their "
"own `metric` and `mode` parameters. Either remove the arguments "
"from your reporter or from your call to `tune.run()`")
progress_reporter.set_total_samples(search_alg.total_samples)
# User Warning for GPUs
if trial_executor.has_gpus():
if isinstance(resources_per_trial,
dict) and "gpu" in resources_per_trial:
# "gpu" is manually set.
pass
elif _check_default_resources_override(experiments[0].run_identifier):
# "default_resources" is manually overridden.
pass
else:
logger.warning("Tune detects GPUs, but no trials are using GPUs. "
"To enable trials to use GPUs, set "
"tune.run(resources_per_trial={'gpu': 1}...) "
"which allows Tune to expose 1 GPU to each trial. "
"You can also override "
"`Trainable.default_resource_request` if using the "
"Trainable API.")
original_handler = signal.getsignal(signal.SIGINT)
state = {signal.SIGINT: False}
def sigint_handler(sig, frame):
logger.warning(
"SIGINT received (e.g. via Ctrl+C), ending Ray Tune run. "
"This will try to checkpoint the experiment state one last time. "
"Press CTRL+C one more time (or send SIGINT/SIGKILL/SIGTERM) "
"to skip. ")
state[signal.SIGINT] = True
# Restore original signal handler to react to future SIGINT signals
signal.signal(signal.SIGINT, original_handler)
if not int(os.getenv("TUNE_DISABLE_SIGINT_HANDLER", "0")):
signal.signal(signal.SIGINT, sigint_handler)
tune_start = time.time()
while not runner.is_finished() and not state[signal.SIGINT]:
runner.step()
if has_verbosity(Verbosity.V1_EXPERIMENT):
_report_progress(runner, progress_reporter)
tune_taken = time.time() - tune_start
try:
runner.checkpoint(force=True)
except Exception as e:
logger.warning(f"Trial Runner checkpointing failed: {str(e)}")
if has_verbosity(Verbosity.V1_EXPERIMENT):
_report_progress(runner, progress_reporter, done=True)
wait_for_sync()
runner.cleanup_trials()
incomplete_trials = []
for trial in runner.get_trials():
if trial.status != Trial.TERMINATED:
incomplete_trials += [trial]
if incomplete_trials:
if raise_on_failed_trial and not state[signal.SIGINT]:
raise TuneError("Trials did not complete", incomplete_trials)
else:
logger.error("Trials did not complete: %s", incomplete_trials)
all_taken = time.time() - all_start
if has_verbosity(Verbosity.V1_EXPERIMENT):
logger.info(f"Total run time: {all_taken:.2f} seconds "
f"({tune_taken:.2f} seconds for the tuning loop).")
if state[signal.SIGINT]:
logger.warning(
"Experiment has been interrupted, but the most recent state was "
"saved. You can continue running this experiment by passing "
"`resume=True` to `tune.run()`")
trials = runner.get_trials()
return ExperimentAnalysis(
runner.checkpoint_file,
trials=trials,
default_metric=metric,
default_mode=mode)
def test_cluster_interrupt(start_connected_cluster, tmpdir):
"""Tests run_experiment on cluster shutdown with actual interrupt.
This is an end-to-end test.
"""
cluster = start_connected_cluster
dirpath = str(tmpdir)
# Needs to be in scope for pytest
class _Mock(tune.Trainable):
"""Finishes on the 4th iteration."""
def _setup(self, config):
self.state = {"hi": 0}
def _train(self):
self.state["hi"] += 1
time.sleep(0.5)
return {"done": self.state["hi"] >= 4}
def _save(self, path):
return self.state
def _restore(self, state):
self.state = state
# Removes indent from class.
reformatted = "\n".join(line[4:] if len(line) else line
for line in inspect.getsource(_Mock).split("\n"))
script = """
import time
import ray
from ray import tune
ray.init(address="{address}")
{fail_class_code}
tune.run(
{fail_class},
name="experiment",
stop=dict(training_iteration=5),
local_dir="{checkpoint_dir}",
checkpoint_freq=1,
global_checkpoint_period=0,
max_failures=1,
raise_on_failed_trial=False)
""".format(address=cluster.address,
checkpoint_dir=dirpath,
fail_class_code=reformatted,
fail_class=_Mock.__name__)
run_string_as_driver_nonblocking(script)
# Wait until the right checkpoint is saved.
# The trainable returns every 0.5 seconds, so this should not miss
# the checkpoint.
local_checkpoint_dir = os.path.join(dirpath, "experiment")
for i in range(50):
if TrialRunner.checkpoint_exists(local_checkpoint_dir):
# Inspect the internal trialrunner
runner = TrialRunner(resume="LOCAL",
local_checkpoint_dir=local_checkpoint_dir)
trials = runner.get_trials()
last_res = trials[0].last_result
if last_res and last_res.get("training_iteration") == 3:
break
time.sleep(0.2)
if not TrialRunner.checkpoint_exists(local_checkpoint_dir):
raise RuntimeError("Checkpoint file didn't appear.")
ray.shutdown()
cluster.shutdown()
cluster = _start_new_cluster()
Experiment.register_if_needed(_Mock)
# Inspect the internal trialrunner
runner = TrialRunner(resume="LOCAL",
local_checkpoint_dir=local_checkpoint_dir)
trials = runner.get_trials()
assert trials[0].last_result["training_iteration"] == 3
assert trials[0].status == Trial.PENDING
# Restore properly from checkpoint
trials2 = tune.run_experiments(
{
"experiment": {
"run": _Mock,
"local_dir": dirpath,
"checkpoint_freq": 1
}
},
resume=True,
raise_on_failed_trial=False)
assert all(t.status == Trial.TERMINATED for t in trials2)
assert {t.trial_id for t in trials2} == {t.trial_id for t in trials}
ray.shutdown()
cluster.shutdown()
def run(run_or_experiment,
name=None,
stop=None,
config=None,
resources_per_trial=None,
num_samples=1,
local_dir=None,
upload_dir=None,
trial_name_creator=None,
loggers=None,
sync_to_cloud=None,
sync_to_driver=None,
checkpoint_freq=0,
checkpoint_at_end=False,
export_formats=None,
max_failures=3,
restore=None,
search_alg=None,
scheduler=None,
with_server=False,
server_port=TuneServer.DEFAULT_PORT,
verbose=2,
resume=False,
queue_trials=False,
reuse_actors=True,
trial_executor=None,
raise_on_failed_trial=True,
return_trials=True,
ray_auto_init=True,
sync_function=None):
"""Executes training.
Args:
run_or_experiment (function|class|str|Experiment): If
function|class|str, this is the algorithm or model to train.
This may refer to the name of a built-on algorithm
(e.g. RLLib's DQN or PPO), a user-defined trainable
function or class, or the string identifier of a
trainable function or class registered in the tune registry.
If Experiment, then Tune will execute training based on
Experiment.spec.
name (str): Name of experiment.
stop (dict): The stopping criteria. The keys may be any field in
the return result of 'train()', whichever is reached first.
Defaults to empty dict.
config (dict): Algorithm-specific configuration for Tune variant
generation (e.g. env, hyperparams). Defaults to empty dict.
Custom search algorithms may ignore this.
resources_per_trial (dict): Machine resources to allocate per trial,
e.g. ``{"cpu": 64, "gpu": 8}``. Note that GPUs will not be
assigned unless you specify them here. Defaults to 1 CPU and 0
GPUs in ``Trainable.default_resource_request()``.
num_samples (int): Number of times to sample from the
hyperparameter space. Defaults to 1. If `grid_search` is
provided as an argument, the grid will be repeated
`num_samples` of times.
local_dir (str): Local dir to save training results to.
Defaults to ``~/ray_results``.
upload_dir (str): Optional URI to sync training results
to (e.g. ``s3://bucket``).
trial_name_creator (func): Optional function for generating
the trial string representation.
loggers (list): List of logger creators to be used with
each Trial. If None, defaults to ray.tune.logger.DEFAULT_LOGGERS.
See `ray/tune/logger.py`.
sync_to_cloud (func|str): Function for syncing the local_dir to and
from upload_dir. If string, then it must be a string template
that includes `{source}` and `{target}` for the syncer to run.
If not provided, the sync command defaults to standard
S3 or gsutil sync comamnds.
sync_to_driver (func|str): Function for syncing trial logdir from
remote node to local. If string, then it must be a string template
that includes `{source}` and `{target}` for the syncer to run.
If not provided, defaults to using rsync.
checkpoint_freq (int): How many training iterations between
checkpoints. A value of 0 (default) disables checkpointing.
checkpoint_at_end (bool): Whether to checkpoint at the end of the
experiment regardless of the checkpoint_freq. Default is False.
export_formats (list): List of formats that exported at the end of
the experiment. Default is None.
max_failures (int): Try to recover a trial from its last
checkpoint at least this many times. Only applies if
checkpointing is enabled. Setting to -1 will lead to infinite
recovery retries. Defaults to 3.
restore (str): Path to checkpoint. Only makes sense to set if
running 1 trial. Defaults to None.
search_alg (SearchAlgorithm): Search Algorithm. Defaults to
BasicVariantGenerator.
scheduler (TrialScheduler): Scheduler for executing
the experiment. Choose among FIFO (default), MedianStopping,
AsyncHyperBand, and HyperBand.
with_server (bool): Starts a background Tune server. Needed for
using the Client API.
server_port (int): Port number for launching TuneServer.
verbose (int): 0, 1, or 2. Verbosity mode. 0 = silent,
1 = only status updates, 2 = status and trial results.
resume (str|bool): One of "LOCAL", "REMOTE", "PROMPT", or bool.
LOCAL/True restores the checkpoint from the local_checkpoint_dir.
REMOTE restores the checkpoint from remote_checkpoint_dir.
PROMPT provides CLI feedback. False forces a new
experiment. If resume is set but checkpoint does not exist,
ValueError will be thrown.
queue_trials (bool): Whether to queue trials when the cluster does
not currently have enough resources to launch one. This should
be set to True when running on an autoscaling cluster to enable
automatic scale-up.
reuse_actors (bool): Whether to reuse actors between different trials
when possible. This can drastically speed up experiments that start
and stop actors often (e.g., PBT in time-multiplexing mode). This
requires trials to have the same resource requirements.
trial_executor (TrialExecutor): Manage the execution of trials.
raise_on_failed_trial (bool): Raise TuneError if there exists failed
trial (of ERROR state) when the experiments complete.
ray_auto_init (bool): Automatically starts a local Ray cluster
if using a RayTrialExecutor (which is the default) and
if Ray is not initialized. Defaults to True.
sync_function: Deprecated. See `sync_to_cloud` and
`sync_to_driver`.
Returns:
List of Trial objects.
Raises:
TuneError if any trials failed and `raise_on_failed_trial` is True.
Examples:
>>> tune.run(mytrainable, scheduler=PopulationBasedTraining())
>>> tune.run(mytrainable, num_samples=5, reuse_actors=True)
>>> tune.run(
"PG",
num_samples=5,
config={
"env": "CartPole-v0",
"lr": tune.sample_from(lambda _: np.random.rand())
}
)
"""
trial_executor = trial_executor or RayTrialExecutor(
queue_trials=queue_trials,
reuse_actors=reuse_actors,
ray_auto_init=ray_auto_init)
experiment = run_or_experiment
if not isinstance(run_or_experiment, Experiment):
run_identifier = Experiment._register_if_needed(run_or_experiment)
experiment = Experiment(
name=name,
run=run_identifier,
stop=stop,
config=config,
resources_per_trial=resources_per_trial,
num_samples=num_samples,
local_dir=local_dir,
upload_dir=upload_dir,
sync_to_driver=sync_to_driver,
trial_name_creator=trial_name_creator,
loggers=loggers,
checkpoint_freq=checkpoint_freq,
checkpoint_at_end=checkpoint_at_end,
export_formats=export_formats,
max_failures=max_failures,
restore=restore,
sync_function=sync_function)
else:
logger.debug("Ignoring some parameters passed into tune.run.")
if sync_to_cloud:
assert experiment.remote_checkpoint_dir, (
"Need `upload_dir` if `sync_to_cloud` given.")
runner = TrialRunner(
search_alg=search_alg or BasicVariantGenerator(),
scheduler=scheduler or FIFOScheduler(),
local_checkpoint_dir=experiment.checkpoint_dir,
remote_checkpoint_dir=experiment.remote_checkpoint_dir,
sync_to_cloud=sync_to_cloud,
resume=resume,
launch_web_server=with_server,
server_port=server_port,
verbose=bool(verbose > 1),
trial_executor=trial_executor)
runner.add_experiment(experiment)
if verbose:
print(runner.debug_string(max_debug=99999))
last_debug = 0
while not runner.is_finished():
runner.step()
if time.time() - last_debug > DEBUG_PRINT_INTERVAL:
if verbose:
print(runner.debug_string())
last_debug = time.time()
if verbose:
print(runner.debug_string(max_debug=99999))
wait_for_sync()
errored_trials = []
for trial in runner.get_trials():
if trial.status != Trial.TERMINATED:
errored_trials += [trial]
if errored_trials:
if raise_on_failed_trial:
raise TuneError("Trials did not complete", errored_trials)
else:
logger.error("Trials did not complete: %s", errored_trials)
if return_trials:
return runner.get_trials()
return ExperimentAnalysis(experiment.checkpoint_dir)
def run(
run_or_experiment: Union[str, Callable, Type],
name: Optional[str] = None,
metric: Optional[str] = None,
mode: Optional[str] = None,
stop: Union[None, Mapping, Stopper, Callable[[str, Mapping],
bool]] = None,
time_budget_s: Union[None, int, float, datetime.timedelta] = None,
config: Optional[Dict[str, Any]] = None,
resources_per_trial: Union[None, Mapping[str, Union[
float, int, Mapping]], PlacementGroupFactory] = None,
num_samples: int = 1,
local_dir: Optional[str] = None,
search_alg: Optional[Union[Searcher, SearchAlgorithm, str]] = None,
scheduler: Optional[Union[TrialScheduler, str]] = None,
keep_checkpoints_num: Optional[int] = None,
checkpoint_score_attr: Optional[str] = None,
checkpoint_freq: int = 0,
checkpoint_at_end: bool = False,
verbose: Union[int, Verbosity] = Verbosity.V3_TRIAL_DETAILS,
progress_reporter: Optional[ProgressReporter] = None,
log_to_file: bool = False,
trial_name_creator: Optional[Callable[[Trial], str]] = None,
trial_dirname_creator: Optional[Callable[[Trial], str]] = None,
sync_config: Optional[SyncConfig] = None,
export_formats: Optional[Sequence] = None,
max_failures: int = 0,
fail_fast: bool = False,
restore: Optional[str] = None,
server_port: Optional[int] = None,
resume: bool = False,
reuse_actors: bool = False,
trial_executor: Optional[RayTrialExecutor] = None,
raise_on_failed_trial: bool = True,
callbacks: Optional[Sequence[Callback]] = None,
max_concurrent_trials: Optional[int] = None,
# Deprecated args
queue_trials: Optional[bool] = None,
loggers: Optional[Sequence[Type[Logger]]] = None,
_remote: Optional[bool] = None,
) -> ExperimentAnalysis:
"""Executes training.
When a SIGINT signal is received (e.g. through Ctrl+C), the tuning run
will gracefully shut down and checkpoint the latest experiment state.
Sending SIGINT again (or SIGKILL/SIGTERM instead) will skip this step.
Many aspects of Tune, such as the frequency of global checkpointing,
maximum pending placement group trials and the path of the result
directory be configured through environment variables. Refer to
:ref:`tune-env-vars` for a list of environment variables available.
Examples:
.. code-block:: python
# Run 10 trials (each trial is one instance of a Trainable). Tune runs
# in parallel and automatically determines concurrency.
tune.run(trainable, num_samples=10)
# Run 1 trial, stop when trial has reached 10 iterations
tune.run(my_trainable, stop={"training_iteration": 10})
# automatically retry failed trials up to 3 times
tune.run(my_trainable, stop={"training_iteration": 10}, max_failures=3)
# Run 1 trial, search over hyperparameters, stop after 10 iterations.
space = {"lr": tune.uniform(0, 1), "momentum": tune.uniform(0, 1)}
tune.run(my_trainable, config=space, stop={"training_iteration": 10})
# Resumes training if a previous machine crashed
tune.run(my_trainable, config=space,
local_dir=<path/to/dir>, resume=True)
# Rerun ONLY failed trials after an experiment is finished.
tune.run(my_trainable, config=space,
local_dir=<path/to/dir>, resume="ERRORED_ONLY")
Args:
run_or_experiment (function | class | str | :class:`Experiment`): If
function|class|str, this is the algorithm or model to train.
This may refer to the name of a built-on algorithm
(e.g. RLLib's DQN or PPO), a user-defined trainable
function or class, or the string identifier of a
trainable function or class registered in the tune registry.
If Experiment, then Tune will execute training based on
Experiment.spec. If you want to pass in a Python lambda, you
will need to first register the function:
``tune.register_trainable("lambda_id", lambda x: ...)``. You can
then use ``tune.run("lambda_id")``.
metric (str): Metric to optimize. This metric should be reported
with `tune.report()`. If set, will be passed to the search
algorithm and scheduler.
mode (str): Must be one of [min, max]. Determines whether objective is
minimizing or maximizing the metric attribute. If set, will be
passed to the search algorithm and scheduler.
name (str): Name of experiment.
stop (dict | callable | :class:`Stopper`): Stopping criteria. If dict,
the keys may be any field in the return result of 'train()',
whichever is reached first. If function, it must take (trial_id,
result) as arguments and return a boolean (True if trial should be
stopped, False otherwise). This can also be a subclass of
``ray.tune.Stopper``, which allows users to implement
custom experiment-wide stopping (i.e., stopping an entire Tune
run based on some time constraint).
time_budget_s (int|float|datetime.timedelta): Global time budget in
seconds after which all trials are stopped. Can also be a
``datetime.timedelta`` object.
config (dict): Algorithm-specific configuration for Tune variant
generation (e.g. env, hyperparams). Defaults to empty dict.
Custom search algorithms may ignore this.
resources_per_trial (dict|PlacementGroupFactory): Machine resources
to allocate per trial, e.g. ``{"cpu": 64, "gpu": 8}``.
Note that GPUs will not be assigned unless you specify them here.
Defaults to 1 CPU and 0 GPUs in
``Trainable.default_resource_request()``. This can also
be a PlacementGroupFactory object wrapping arguments to create a
per-trial placement group.
num_samples (int): Number of times to sample from the
hyperparameter space. Defaults to 1. If `grid_search` is
provided as an argument, the grid will be repeated
`num_samples` of times. If this is -1, (virtually) infinite
samples are generated until a stopping condition is met.
local_dir (str): Local dir to save training results to.
Defaults to ``~/ray_results``.
search_alg (Searcher|SearchAlgorithm|str): Search algorithm for
optimization. You can also use the name of the algorithm.
scheduler (TrialScheduler|str): Scheduler for executing
the experiment. Choose among FIFO (default), MedianStopping,
AsyncHyperBand, HyperBand and PopulationBasedTraining. Refer to
ray.tune.schedulers for more options. You can also use the
name of the scheduler.
keep_checkpoints_num (int): Number of checkpoints to keep. A value of
`None` keeps all checkpoints. Defaults to `None`. If set, need
to provide `checkpoint_score_attr`.
checkpoint_score_attr (str): Specifies by which attribute to rank the
best checkpoint. Default is increasing order. If attribute starts
with `min-` it will rank attribute in decreasing order, i.e.
`min-validation_loss`.
checkpoint_freq (int): How many training iterations between
checkpoints. A value of 0 (default) disables checkpointing.
This has no effect when using the Functional Training API.
checkpoint_at_end (bool): Whether to checkpoint at the end of the
experiment regardless of the checkpoint_freq. Default is False.
This has no effect when using the Functional Training API.
verbose (Union[int, Verbosity]): 0, 1, 2, or 3. Verbosity mode.
0 = silent, 1 = only status updates, 2 = status and brief trial
results, 3 = status and detailed trial results. Defaults to 3.
progress_reporter (ProgressReporter): Progress reporter for reporting
intermediate experiment progress. Defaults to CLIReporter if
running in command-line, or JupyterNotebookReporter if running in
a Jupyter notebook.
log_to_file (bool|str|Sequence): Log stdout and stderr to files in
Tune's trial directories. If this is `False` (default), no files
are written. If `true`, outputs are written to `trialdir/stdout`
and `trialdir/stderr`, respectively. If this is a single string,
this is interpreted as a file relative to the trialdir, to which
both streams are written. If this is a Sequence (e.g. a Tuple),
it has to have length 2 and the elements indicate the files to
which stdout and stderr are written, respectively.
trial_name_creator (Callable[[Trial], str]): Optional function
for generating the trial string representation.
trial_dirname_creator (Callable[[Trial], str]): Function
for generating the trial dirname. This function should take
in a Trial object and return a string representing the
name of the directory. The return value cannot be a path.
sync_config (SyncConfig): Configuration object for syncing. See
tune.SyncConfig.
export_formats (list): List of formats that exported at the end of
the experiment. Default is None.
max_failures (int): Try to recover a trial at least this many times.
Ray will recover from the latest checkpoint if present.
Setting to -1 will lead to infinite recovery retries.
Setting to 0 will disable retries. Defaults to 0.
fail_fast (bool | str): Whether to fail upon the first error.
If fail_fast='raise' provided, Tune will automatically
raise the exception received by the Trainable. fail_fast='raise'
can easily leak resources and should be used with caution (it
is best used with `ray.init(local_mode=True)`).
restore (str): Path to checkpoint. Only makes sense to set if
running 1 trial. Defaults to None.
server_port (int): Port number for launching TuneServer.
resume (str|bool): One of "LOCAL", "REMOTE", "PROMPT", "ERRORED_ONLY",
or bool. LOCAL/True restores the checkpoint from the
local experiment directory, determined
by ``name`` and ``local_dir``. REMOTE restores the checkpoint
from ``upload_dir`` (as passed to ``sync_config``).
PROMPT provides CLI feedback.
False forces a new experiment. ERRORED_ONLY resets and reruns
ERRORED trials upon resume - previous trial artifacts will
be left untouched. If resume is set but checkpoint does not exist,
ValueError will be thrown.
reuse_actors (bool): Whether to reuse actors between different trials
when possible. This can drastically speed up experiments that start
and stop actors often (e.g., PBT in time-multiplexing mode). This
requires trials to have the same resource requirements.
trial_executor (TrialExecutor): Manage the execution of trials.
raise_on_failed_trial (bool): Raise TuneError if there exists failed
trial (of ERROR state) when the experiments complete.
callbacks (list): List of callbacks that will be called at different
times in the training loop. Must be instances of the
``ray.tune.callback.Callback`` class. If not passed,
`LoggerCallback` and `SyncerCallback` callbacks are automatically
added.
max_concurrent_trials (int): Maximum number of trials to run
concurrently. Must be non-negative. If None or 0, no limit will
be applied. This is achieved by wrapping the ``search_alg`` in
a :class:`ConcurrencyLimiter`, and thus setting this argument
will raise an exception if the ``search_alg`` is already a
:class:`ConcurrencyLimiter`. Defaults to None.
_remote (bool): Whether to run the Tune driver in a remote function.
This is disabled automatically if a custom trial executor is
passed in. This is enabled by default in Ray client mode.
Returns:
ExperimentAnalysis: Object for experiment analysis.
Raises:
TuneError: Any trials failed and `raise_on_failed_trial` is True.
"""
# To be removed in 1.9.
if queue_trials is not None:
raise DeprecationWarning(
"`queue_trials` has been deprecated and is replaced by "
"the `TUNE_MAX_PENDING_TRIALS_PG` environment variable. "
"Per default at least one Trial is queued at all times, "
"so you likely don't need to change anything other than "
"removing this argument from your call to `tune.run()`")
# NO CODE IS TO BE ADDED ABOVE THIS COMMENT
# remote_run_kwargs must be defined before any other
# code is ran to ensure that at this point,
# `locals()` is equal to args and kwargs
remote_run_kwargs = locals().copy()
remote_run_kwargs.pop("_remote")
if _remote is None:
_remote = ray.util.client.ray.is_connected()
if _remote is True and trial_executor:
raise ValueError("cannot use custom trial executor")
if not trial_executor or isinstance(trial_executor, RayTrialExecutor):
_ray_auto_init()
if _remote:
remote_run = ray.remote(num_cpus=0)(run)
# Make sure tune.run is called on the sever node.
remote_run = force_on_current_node(remote_run)
# JupyterNotebooks don't work with remote tune runs out of the box
# (e.g. via Ray client) as they don't have access to the main
# process stdout. So we introduce a queue here that accepts
# callables, which will then be executed on the driver side.
if isinstance(progress_reporter, JupyterNotebookReporter):
execute_queue = Queue(actor_options={
"num_cpus": 0,
**force_on_current_node(None)
})
progress_reporter.set_output_queue(execute_queue)
def get_next_queue_item():
try:
return execute_queue.get(block=False)
except Empty:
return None
else:
# If we don't need a queue, use this dummy get fn instead of
# scheduling an unneeded actor
def get_next_queue_item():
return None
def _handle_execute_queue():
execute_item = get_next_queue_item()
while execute_item:
if isinstance(execute_item, Callable):
execute_item()
execute_item = get_next_queue_item()
remote_future = remote_run.remote(_remote=False, **remote_run_kwargs)
# ray.wait(...)[1] returns futures that are not ready, yet
while ray.wait([remote_future], timeout=0.2)[1]:
# Check if we have items to execute
_handle_execute_queue()
# Handle queue one last time
_handle_execute_queue()
return ray.get(remote_future)
del remote_run_kwargs
all_start = time.time()
if loggers:
# Raise DeprecationWarning in 1.9, remove in 1.10/1.11
warnings.warn(
"The `loggers` argument is deprecated. Please pass the respective "
"`LoggerCallback` classes to the `callbacks` argument instead. "
"See https://docs.ray.io/en/latest/tune/api_docs/logging.html")
if mode and mode not in ["min", "max"]:
raise ValueError(
"The `mode` parameter passed to `tune.run()` has to be one of "
"['min', 'max']")
set_verbosity(verbose)
config = config or {}
sync_config = sync_config or SyncConfig()
set_sync_periods(sync_config)
if num_samples == -1:
num_samples = sys.maxsize
result_buffer_length = None
# Create scheduler here as we need access to some of its properties
if isinstance(scheduler, str):
# importing at top level causes a recursive dependency
from ray.tune.schedulers import create_scheduler
scheduler = create_scheduler(scheduler)
scheduler = scheduler or FIFOScheduler()
if not scheduler.supports_buffered_results:
# Result buffering with e.g. a Hyperband scheduler is a bad idea, as
# hyperband tries to stop trials when processing brackets. With result
# buffering, we might trigger this multiple times when evaluating
# a single trial, which leads to unexpected behavior.
env_result_buffer_length = os.getenv("TUNE_RESULT_BUFFER_LENGTH", "")
if env_result_buffer_length:
warnings.warn(
f"You are using a {type(scheduler)} scheduler, but "
f"TUNE_RESULT_BUFFER_LENGTH is set "
f"({env_result_buffer_length}). This can lead to undesired "
f"and faulty behavior, so the buffer length was forcibly set "
f"to 1 instead.")
result_buffer_length = 1
if isinstance(scheduler,
(PopulationBasedTraining,
PopulationBasedTrainingReplay)) and not reuse_actors:
warnings.warn(
"Consider boosting PBT performance by enabling `reuse_actors` as "
"well as implementing `reset_config` for Trainable.")
trial_executor = trial_executor or RayTrialExecutor(
reuse_actors=reuse_actors, result_buffer_length=result_buffer_length)
if isinstance(run_or_experiment, list):
experiments = run_or_experiment
else:
experiments = [run_or_experiment]
for i, exp in enumerate(experiments):
if not isinstance(exp, Experiment):
experiments[i] = Experiment(
name=name,
run=exp,
stop=stop,
time_budget_s=time_budget_s,
config=config,
resources_per_trial=resources_per_trial,
num_samples=num_samples,
local_dir=local_dir,
sync_config=sync_config,
trial_name_creator=trial_name_creator,
trial_dirname_creator=trial_dirname_creator,
log_to_file=log_to_file,
checkpoint_freq=checkpoint_freq,
checkpoint_at_end=checkpoint_at_end,
keep_checkpoints_num=keep_checkpoints_num,
checkpoint_score_attr=checkpoint_score_attr,
export_formats=export_formats,
max_failures=max_failures,
restore=restore)
else:
logger.debug("Ignoring some parameters passed into tune.run.")
if fail_fast and max_failures != 0:
raise ValueError("max_failures must be 0 if fail_fast=True.")
if isinstance(search_alg, str):
# importing at top level causes a recursive dependency
from ray.tune.suggest import create_searcher
search_alg = create_searcher(search_alg)
# if local_mode=True is set during ray.init().
is_local_mode = ray.worker._mode() == ray.worker.LOCAL_MODE
if is_local_mode:
max_concurrent_trials = 1
if not search_alg:
search_alg = BasicVariantGenerator(
max_concurrent=max_concurrent_trials or 0)
elif max_concurrent_trials:
if isinstance(search_alg, ConcurrencyLimiter):
if search_alg.max_concurrent != max_concurrent_trials:
raise ValueError(
"You have specified `max_concurrent_trials="
f"{max_concurrent_trials}`, but the `search_alg` is "
"already a `ConcurrencyLimiter` with `max_concurrent="
f"{search_alg.max_concurrent}. FIX THIS by setting "
"`max_concurrent_trials=None`.")
else:
logger.warning(
"You have specified `max_concurrent_trials="
f"{max_concurrent_trials}`, but the `search_alg` is "
"already a `ConcurrencyLimiter`. `max_concurrent_trials` "
"will be ignored.")
else:
if max_concurrent_trials < 1:
raise ValueError(
"`max_concurrent_trials` must be greater or equal than 1, "
f"got {max_concurrent_trials}.")
if isinstance(search_alg, Searcher):
search_alg = ConcurrencyLimiter(
search_alg, max_concurrent=max_concurrent_trials)
elif not is_local_mode:
logger.warning(
"You have passed a `SearchGenerator` instance as the "
"`search_alg`, but `max_concurrent_trials` requires a "
"`Searcher` instance`. `max_concurrent_trials` "
"will be ignored.")
if isinstance(search_alg, Searcher):
search_alg = SearchGenerator(search_alg)
if config and not set_search_properties_backwards_compatible(
search_alg.set_search_properties, metric, mode, config, **
experiments[0].public_spec):
if has_unresolved_values(config):
raise ValueError(
"You passed a `config` parameter to `tune.run()` with "
"unresolved parameters, but the search algorithm was already "
"instantiated with a search space. Make sure that `config` "
"does not contain any more parameter definitions - include "
"them in the search algorithm's search space if necessary.")
if not scheduler.set_search_properties(metric, mode):
raise ValueError(
"You passed a `metric` or `mode` argument to `tune.run()`, but "
"the scheduler you are using was already instantiated with their "
"own `metric` and `mode` parameters. Either remove the arguments "
"from your scheduler or from your call to `tune.run()`")
# Create syncer callbacks
callbacks = create_default_callbacks(
callbacks, sync_config, metric=metric, loggers=loggers)
runner = TrialRunner(
search_alg=search_alg,
scheduler=scheduler,
local_checkpoint_dir=experiments[0].checkpoint_dir,
remote_checkpoint_dir=experiments[0].remote_checkpoint_dir,
sync_config=sync_config,
stopper=experiments[0].stopper,
resume=resume,
server_port=server_port,
fail_fast=fail_fast,
trial_executor=trial_executor,
callbacks=callbacks,
metric=metric,
# Driver should only sync trial checkpoints if
# checkpoints are not synced to cloud
driver_sync_trial_checkpoints=not bool(sync_config.upload_dir))
if not runner.resumed:
for exp in experiments:
search_alg.add_configurations([exp])
else:
logger.info("TrialRunner resumed, ignoring new add_experiment but "
"updating trial resources.")
if resources_per_trial:
runner.update_pending_trial_resources(resources_per_trial)
progress_reporter = progress_reporter or detect_reporter()
if not progress_reporter.set_search_properties(metric, mode):
raise ValueError(
"You passed a `metric` or `mode` argument to `tune.run()`, but "
"the reporter you are using was already instantiated with their "
"own `metric` and `mode` parameters. Either remove the arguments "
"from your reporter or from your call to `tune.run()`")
progress_reporter.set_total_samples(search_alg.total_samples)
# Calls setup on callbacks
runner.setup_experiments(
experiments=experiments, total_num_samples=search_alg.total_samples)
# User Warning for GPUs
if trial_executor.has_gpus():
if isinstance(resources_per_trial,
dict) and "gpu" in resources_per_trial:
# "gpu" is manually set.
pass
elif _check_default_resources_override(experiments[0].run_identifier):
# "default_resources" is manually overridden.
pass
else:
logger.warning("Tune detects GPUs, but no trials are using GPUs. "
"To enable trials to use GPUs, set "
"tune.run(resources_per_trial={'gpu': 1}...) "
"which allows Tune to expose 1 GPU to each trial. "
"You can also override "
"`Trainable.default_resource_request` if using the "
"Trainable API.")
original_handler = signal.getsignal(signal.SIGINT)
state = {signal.SIGINT: False}
def sigint_handler(sig, frame):
logger.warning(
"SIGINT received (e.g. via Ctrl+C), ending Ray Tune run. "
"This will try to checkpoint the experiment state one last time. "
"Press CTRL+C one more time (or send SIGINT/SIGKILL/SIGTERM) "
"to skip. ")
state[signal.SIGINT] = True
# Restore original signal handler to react to future SIGINT signals
signal.signal(signal.SIGINT, original_handler)
if not int(os.getenv("TUNE_DISABLE_SIGINT_HANDLER", "0")):
signal.signal(signal.SIGINT, sigint_handler)
tune_start = time.time()
progress_reporter.set_start_time(tune_start)
while not runner.is_finished() and not state[signal.SIGINT]:
runner.step()
if has_verbosity(Verbosity.V1_EXPERIMENT):
_report_progress(runner, progress_reporter)
tune_taken = time.time() - tune_start
try:
runner.checkpoint(force=True)
except Exception as e:
logger.warning(f"Trial Runner checkpointing failed: {str(e)}")
if has_verbosity(Verbosity.V1_EXPERIMENT):
_report_progress(runner, progress_reporter, done=True)
wait_for_sync()
runner.cleanup()
incomplete_trials = []
for trial in runner.get_trials():
if trial.status != Trial.TERMINATED:
incomplete_trials += [trial]
if incomplete_trials:
if raise_on_failed_trial and not state[signal.SIGINT]:
raise TuneError("Trials did not complete", incomplete_trials)
else:
logger.error("Trials did not complete: %s", incomplete_trials)
all_taken = time.time() - all_start
if has_verbosity(Verbosity.V1_EXPERIMENT):
logger.info(f"Total run time: {all_taken:.2f} seconds "
f"({tune_taken:.2f} seconds for the tuning loop).")
if state[signal.SIGINT]:
logger.warning(
"Experiment has been interrupted, but the most recent state was "
"saved. You can continue running this experiment by passing "
"`resume=True` to `tune.run()`")
trials = runner.get_trials()
return ExperimentAnalysis(
runner.checkpoint_file,
trials=trials,
default_metric=metric,
default_mode=mode,
sync_config=sync_config)
def run(run_or_experiment,
name=None,
stop=None,
config=None,
resources_per_trial=None,
num_samples=1,
local_dir=None,
upload_dir=None,
trial_name_creator=None,
loggers=None,
log_to_file=False,
sync_to_cloud=None,
sync_to_driver=None,
checkpoint_freq=0,
checkpoint_at_end=False,
sync_on_checkpoint=True,
keep_checkpoints_num=None,
checkpoint_score_attr=None,
global_checkpoint_period=10,
export_formats=None,
max_failures=0,
fail_fast=False,
restore=None,
search_alg=None,
scheduler=None,
with_server=False,
server_port=TuneServer.DEFAULT_PORT,
verbose=2,
progress_reporter=None,
resume=False,
queue_trials=False,
reuse_actors=False,
trial_executor=None,
raise_on_failed_trial=True,
return_trials=False,
ray_auto_init=True):
"""Executes training.
Args:
run_or_experiment (function | class | str | :class:`Experiment`): If
function|class|str, this is the algorithm or model to train.
This may refer to the name of a built-on algorithm
(e.g. RLLib's DQN or PPO), a user-defined trainable
function or class, or the string identifier of a
trainable function or class registered in the tune registry.
If Experiment, then Tune will execute training based on
Experiment.spec. If you want to pass in a Python lambda, you
will need to first register the function:
``tune.register_trainable("lambda_id", lambda x: ...)``. You can
then use ``tune.run("lambda_id")``.
name (str): Name of experiment.
stop (dict | callable | :class:`Stopper`): Stopping criteria. If dict,
the keys may be any field in the return result of 'train()',
whichever is reached first. If function, it must take (trial_id,
result) as arguments and return a boolean (True if trial should be
stopped, False otherwise). This can also be a subclass of
``ray.tune.Stopper``, which allows users to implement
custom experiment-wide stopping (i.e., stopping an entire Tune
run based on some time constraint).
config (dict): Algorithm-specific configuration for Tune variant
generation (e.g. env, hyperparams). Defaults to empty dict.
Custom search algorithms may ignore this.
resources_per_trial (dict): Machine resources to allocate per trial,
e.g. ``{"cpu": 64, "gpu": 8}``. Note that GPUs will not be
assigned unless you specify them here. Defaults to 1 CPU and 0
GPUs in ``Trainable.default_resource_request()``.
num_samples (int): Number of times to sample from the
hyperparameter space. Defaults to 1. If `grid_search` is
provided as an argument, the grid will be repeated
`num_samples` of times.
local_dir (str): Local dir to save training results to.
Defaults to ``~/ray_results``.
upload_dir (str): Optional URI to sync training results and checkpoints
to (e.g. ``s3://bucket`` or ``gs://bucket``).
trial_name_creator (func): Optional function for generating
the trial string representation.
loggers (list): List of logger creators to be used with
each Trial. If None, defaults to ray.tune.logger.DEFAULT_LOGGERS.
See `ray/tune/logger.py`.
log_to_file (bool|str|Sequence): Log stdout and stderr to files in
Tune's trial directories. If this is `False` (default), no files
are written. If `true`, outputs are written to `trialdir/stdout`
and `trialdir/stderr`, respectively. If this is a single string,
this is interpreted as a file relative to the trialdir, to which
both streams are written. If this is a Sequence (e.g. a Tuple),
it has to have length 2 and the elements indicate the files to
which stdout and stderr are written, respectively.
sync_to_cloud (func|str): Function for syncing the local_dir to and
from upload_dir. If string, then it must be a string template that
includes `{source}` and `{target}` for the syncer to run. If not
provided, the sync command defaults to standard S3 or gsutil sync
commands. By default local_dir is synced to remote_dir every 300
seconds. To change this, set the TUNE_CLOUD_SYNC_S
environment variable in the driver machine.
sync_to_driver (func|str|bool): Function for syncing trial logdir from
remote node to local. If string, then it must be a string template
that includes `{source}` and `{target}` for the syncer to run.
If True or not provided, it defaults to using rsync. If False,
syncing to driver is disabled.
checkpoint_freq (int): How many training iterations between
checkpoints. A value of 0 (default) disables checkpointing.
This has no effect when using the Functional Training API.
checkpoint_at_end (bool): Whether to checkpoint at the end of the
experiment regardless of the checkpoint_freq. Default is False.
This has no effect when using the Functional Training API.
sync_on_checkpoint (bool): Force sync-down of trial checkpoint to
driver. If set to False, checkpoint syncing from worker to driver
is asynchronous and best-effort. This does not affect persistent
storage syncing. Defaults to True.
keep_checkpoints_num (int): Number of checkpoints to keep. A value of
`None` keeps all checkpoints. Defaults to `None`. If set, need
to provide `checkpoint_score_attr`.
checkpoint_score_attr (str): Specifies by which attribute to rank the
best checkpoint. Default is increasing order. If attribute starts
with `min-` it will rank attribute in decreasing order, i.e.
`min-validation_loss`.
global_checkpoint_period (int): Seconds between global checkpointing.
This does not affect `checkpoint_freq`, which specifies frequency
for individual trials.
export_formats (list): List of formats that exported at the end of
the experiment. Default is None.
max_failures (int): Try to recover a trial at least this many times.
Ray will recover from the latest checkpoint if present.
Setting to -1 will lead to infinite recovery retries.
Setting to 0 will disable retries. Defaults to 3.
fail_fast (bool): Whether to fail upon the first error.
restore (str): Path to checkpoint. Only makes sense to set if
running 1 trial. Defaults to None.
search_alg (Searcher): Search algorithm for optimization.
scheduler (TrialScheduler): Scheduler for executing
the experiment. Choose among FIFO (default), MedianStopping,
AsyncHyperBand, HyperBand and PopulationBasedTraining. Refer to
ray.tune.schedulers for more options.
with_server (bool): Starts a background Tune server. Needed for
using the Client API.
server_port (int): Port number for launching TuneServer.
verbose (int): 0, 1, or 2. Verbosity mode. 0 = silent,
1 = only status updates, 2 = status and trial results.
progress_reporter (ProgressReporter): Progress reporter for reporting
intermediate experiment progress. Defaults to CLIReporter if
running in command-line, or JupyterNotebookReporter if running in
a Jupyter notebook.
resume (str|bool): One of "LOCAL", "REMOTE", "PROMPT", or bool.
LOCAL/True restores the checkpoint from the local_checkpoint_dir.
REMOTE restores the checkpoint from remote_checkpoint_dir.
PROMPT provides CLI feedback. False forces a new
experiment. If resume is set but checkpoint does not exist,
ValueError will be thrown.
queue_trials (bool): Whether to queue trials when the cluster does
not currently have enough resources to launch one. This should
be set to True when running on an autoscaling cluster to enable
automatic scale-up.
reuse_actors (bool): Whether to reuse actors between different trials
when possible. This can drastically speed up experiments that start
and stop actors often (e.g., PBT in time-multiplexing mode). This
requires trials to have the same resource requirements.
trial_executor (TrialExecutor): Manage the execution of trials.
raise_on_failed_trial (bool): Raise TuneError if there exists failed
trial (of ERROR state) when the experiments complete.
ray_auto_init (bool): Automatically starts a local Ray cluster
if using a RayTrialExecutor (which is the default) and
if Ray is not initialized. Defaults to True.
Returns:
ExperimentAnalysis: Object for experiment analysis.
Raises:
TuneError: Any trials failed and `raise_on_failed_trial` is True.
Examples:
.. code-block:: python
# Run 10 trials (each trial is one instance of a Trainable). Tune runs
# in parallel and automatically determines concurrency.
tune.run(trainable, num_samples=10)
# Run 1 trial, stop when trial has reached 10 iterations
tune.run(my_trainable, stop={"training_iteration": 10})
# Run 1 trial, search over hyperparameters, stop after 10 iterations.
space = {"lr": tune.uniform(0, 1), "momentum": tune.uniform(0, 1)}
tune.run(my_trainable, config=space, stop={"training_iteration": 10})
"""
config = config or {}
trial_executor = trial_executor or RayTrialExecutor(
queue_trials=queue_trials,
reuse_actors=reuse_actors,
ray_auto_init=ray_auto_init)
if isinstance(run_or_experiment, list):
experiments = run_or_experiment
else:
experiments = [run_or_experiment]
for i, exp in enumerate(experiments):
if not isinstance(exp, Experiment):
experiments[i] = Experiment(
name=name,
run=exp,
stop=stop,
config=config,
resources_per_trial=resources_per_trial,
num_samples=num_samples,
local_dir=local_dir,
upload_dir=upload_dir,
sync_to_driver=sync_to_driver,
trial_name_creator=trial_name_creator,
loggers=loggers,
log_to_file=log_to_file,
checkpoint_freq=checkpoint_freq,
checkpoint_at_end=checkpoint_at_end,
sync_on_checkpoint=sync_on_checkpoint,
keep_checkpoints_num=keep_checkpoints_num,
checkpoint_score_attr=checkpoint_score_attr,
export_formats=export_formats,
max_failures=max_failures,
restore=restore)
else:
logger.debug("Ignoring some parameters passed into tune.run.")
if sync_to_cloud:
for exp in experiments:
assert exp.remote_checkpoint_dir, (
"Need `upload_dir` if `sync_to_cloud` given.")
if fail_fast and max_failures != 0:
raise ValueError("max_failures must be 0 if fail_fast=True.")
if issubclass(type(search_alg), Searcher):
search_alg = SearchGenerator(search_alg)
if not search_alg:
search_alg = BasicVariantGenerator()
runner = TrialRunner(
search_alg=search_alg,
scheduler=scheduler or FIFOScheduler(),
local_checkpoint_dir=experiments[0].checkpoint_dir,
remote_checkpoint_dir=experiments[0].remote_checkpoint_dir,
sync_to_cloud=sync_to_cloud,
stopper=experiments[0].stopper,
checkpoint_period=global_checkpoint_period,
resume=resume,
launch_web_server=with_server,
server_port=server_port,
verbose=bool(verbose > 1),
fail_fast=fail_fast,
trial_executor=trial_executor)
if not runner.resumed:
for exp in experiments:
search_alg.add_configurations([exp])
else:
logger.info("TrialRunner resumed, ignoring new add_experiment.")
if progress_reporter is None:
if IS_NOTEBOOK:
progress_reporter = JupyterNotebookReporter(overwrite=verbose < 2)
else:
progress_reporter = CLIReporter()
# User Warning for GPUs
if trial_executor.has_gpus():
if isinstance(resources_per_trial,
dict) and "gpu" in resources_per_trial:
# "gpu" is manually set.
pass
elif _check_default_resources_override(experiments[0].run_identifier):
# "default_resources" is manually overriden.
pass
else:
logger.warning("Tune detects GPUs, but no trials are using GPUs. "
"To enable trials to use GPUs, set "
"tune.run(resources_per_trial={'gpu': 1}...) "
"which allows Tune to expose 1 GPU to each trial. "
"You can also override "
"`Trainable.default_resource_request` if using the "
"Trainable API.")
while not runner.is_finished():
runner.step()
if verbose:
_report_progress(runner, progress_reporter)
try:
runner.checkpoint(force=True)
except Exception:
logger.exception("Trial Runner checkpointing failed.")
if verbose:
_report_progress(runner, progress_reporter, done=True)
wait_for_sync()
runner.cleanup_trials()
incomplete_trials = []
for trial in runner.get_trials():
if trial.status != Trial.TERMINATED:
incomplete_trials += [trial]
if incomplete_trials:
if raise_on_failed_trial:
raise TuneError("Trials did not complete", incomplete_trials)
else:
logger.error("Trials did not complete: %s", incomplete_trials)
trials = runner.get_trials()
if return_trials:
return trials
return ExperimentAnalysis(runner.checkpoint_file, trials=trials)
num_cpus=Params.n_cpus,
num_gpus=Params.n_gpus,
local_mode=Params.
debug, # use local mode when debugging, remove it for performance increase
)
loggers = list(DEFAULT_LOGGERS)
loggers[0] = CustomJsonLogger
# initialize experiment
exp = Experiment(
name=f"Sim-{Params.training_alg}",
run=Params.training_alg, # must be the same as the default config
config=config,
stop=Params.stop_conditions,
local_dir=Params.ray_results_dir,
max_failures=9999,
checkpoint_freq=Params.checkpoint_freq,
checkpoint_at_end=True,
loggers=loggers,
)
# defining population scheduler
pbt_scheduler = PopulationBasedTraining(
time_attr='training_iteration',
metric='episode_reward_mean',
mode='max',
perturbation_interval=Params.training_iteration //
100, # perturbate a total of N times during the training
hyperparam_mutations={ # fixme: get correct params
"lr": [1e-4],
