def run_experiments(experiments, scheduler=None, with_server=False,
server_port=TuneServer.DEFAULT_PORT, verbose=True):
# Make sure rllib agents are registered
from ray import rllib # noqa # pylint: disable=unused-import
if scheduler is None:
scheduler = FIFOScheduler()
runner = TrialRunner(
scheduler, launch_web_server=with_server, server_port=server_port)
for name, spec in experiments.items():
for trial in generate_trials(spec, name):
trial.set_verbose(verbose)
runner.add_trial(trial)
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 test_migration_checkpoint_removal(start_connected_emptyhead_cluster):
"""Test checks that trial restarts if checkpoint is lost w/ node fail."""
cluster = start_connected_emptyhead_cluster
node = cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
runner = TrialRunner(BasicVariantGenerator())
kwargs = {
"stopping_criterion": {
"training_iteration": 3
},
"checkpoint_freq": 2,
"max_failures": 2
}
# Test recovery of trial that has been checkpointed
t1 = Trial("__fake", **kwargs)
runner.add_trial(t1)
runner.step() # start
runner.step() # 1 result
runner.step() # 2 result and checkpoint
assert t1.has_checkpoint()
cluster.add_node(num_cpus=1)
cluster.remove_node(node)
cluster.wait_for_nodes()
shutil.rmtree(os.path.dirname(t1._checkpoint.value))
runner.step() # Recovery step
for i in range(3):
runner.step()
assert t1.status == Trial.TERMINATED
def test_cluster_down_simple(start_connected_cluster, tmpdir):
"""Tests that TrialRunner save/restore works on cluster shutdown."""
cluster = start_connected_cluster
cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
dirpath = str(tmpdir)
runner = TrialRunner(
BasicVariantGenerator(), metadata_checkpoint_dir=dirpath)
kwargs = {
"stopping_criterion": {
"training_iteration": 2
},
"checkpoint_freq": 1,
"max_failures": 1
}
trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step() # start
runner.step() # start2
runner.step() # step
assert all(t.status == Trial.RUNNING for t in runner.get_trials())
runner.checkpoint()
cluster.shutdown()
ray.shutdown()
cluster = _start_new_cluster()
runner = TrialRunner.restore(dirpath)
runner.step() # start
runner.step() # start2
for i in range(3):
runner.step()
with pytest.raises(TuneError):
runner.step()
assert all(t.status == Trial.TERMINATED for t in runner.get_trials())
cluster.shutdown()
def test_counting_resources(start_connected_cluster):
"""Tests that Tune accounting is consistent with actual cluster."""
cluster = start_connected_cluster
nodes = []
assert ray.global_state.cluster_resources()["CPU"] == 1
runner = TrialRunner(BasicVariantGenerator())
kwargs = {"stopping_criterion": {"training_iteration": 10}}
trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step() # run 1
nodes += [cluster.add_node(num_cpus=1)]
cluster.wait_for_nodes()
assert ray.global_state.cluster_resources()["CPU"] == 2
cluster.remove_node(nodes.pop())
cluster.wait_for_nodes()
assert ray.global_state.cluster_resources()["CPU"] == 1
runner.step() # run 2
assert sum(t.status == Trial.RUNNING for t in runner.get_trials()) == 1
for i in range(5):
nodes += [cluster.add_node(num_cpus=1)]
cluster.wait_for_nodes()
assert ray.global_state.cluster_resources()["CPU"] == 6
runner.step() # 1 result
assert sum(t.status == Trial.RUNNING for t in runner.get_trials()) == 2
def test_trial_requeue(start_connected_emptyhead_cluster):
"""Removing a node in full cluster causes Trial to be requeued."""
cluster = start_connected_emptyhead_cluster
node = cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
runner = TrialRunner(BasicVariantGenerator())
kwargs = {
"stopping_criterion": {
"training_iteration": 5
},
"checkpoint_freq": 1,
"max_failures": 1
}
trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step() # start
runner.step() # 1 result
cluster.remove_node(node)
cluster.wait_for_nodes()
runner.step()
assert all(t.status == Trial.PENDING for t in trials)
with pytest.raises(TuneError):
runner.step()
def test_remove_node_before_result(start_connected_emptyhead_cluster):
"""Tune continues when node is removed before trial returns."""
cluster = start_connected_emptyhead_cluster
node = cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
runner = TrialRunner(BasicVariantGenerator())
kwargs = {
"stopping_criterion": {
"training_iteration": 3
},
"checkpoint_freq": 2,
"max_failures": 2
}
trial = Trial("__fake", **kwargs)
runner.add_trial(trial)
runner.step() # run 1
assert trial.status == Trial.RUNNING
cluster.remove_node(node)
cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
assert ray.global_state.cluster_resources()["CPU"] == 1
for i in range(3):
runner.step()
assert trial.status == Trial.TERMINATED
with pytest.raises(TuneError):
runner.step()
def testTrialNoSave(self):
"""Check that non-checkpointing trials are not saved."""
ray.init(num_cpus=3)
tmpdir = tempfile.mkdtemp()
runner = TrialRunner(local_checkpoint_dir=tmpdir, checkpoint_period=0)
runner.add_trial(
Trial("__fake",
trial_id="non_checkpoint",
stopping_criterion={"training_iteration": 2}))
while not all(t.status == Trial.TERMINATED
for t in runner.get_trials()):
runner.step()
runner.add_trial(
Trial("__fake",
trial_id="checkpoint",
checkpoint_at_end=True,
stopping_criterion={"training_iteration": 2}))
while not all(t.status == Trial.TERMINATED
for t in runner.get_trials()):
runner.step()
runner.add_trial(
Trial("__fake",
trial_id="pending",
stopping_criterion={"training_iteration": 2}))
runner.step()
runner.step()
runner2 = TrialRunner(resume="LOCAL", local_checkpoint_dir=tmpdir)
new_trials = runner2.get_trials()
self.assertEquals(len(new_trials), 3)
self.assertTrue(
runner2.get_trial("non_checkpoint").status == Trial.TERMINATED)
self.assertTrue(
runner2.get_trial("checkpoint").status == Trial.TERMINATED)
self.assertTrue(runner2.get_trial("pending").status == Trial.PENDING)
self.assertTrue(not runner2.get_trial("pending").last_result)
runner2.step()
shutil.rmtree(tmpdir)
def test_remove_node_before_result(start_connected_emptyhead_cluster):
"""Tune continues when node is removed before trial returns."""
cluster = start_connected_emptyhead_cluster
node = cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
runner = TrialRunner(BasicVariantGenerator())
kwargs = {
"stopping_criterion": {
"training_iteration": 3
},
"checkpoint_freq": 2,
"max_failures": 2,
}
trial = Trial("__fake", **kwargs)
runner.add_trial(trial)
runner.step() # Start trial, call _train once
running_trials = _get_running_trials(runner)
assert len(running_trials) == 1
assert _check_trial_running(running_trials[0])
assert not trial.has_reported_at_least_once
assert trial.status == Trial.RUNNING
cluster.remove_node(node)
cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
assert ray.cluster_resources()["CPU"] == 1
# Process result: fetch data, invoke _train again
runner.step()
assert trial.last_result.get("training_iteration") == 1
# Process result: discover failure, recover, _train (from scratch)
while trial.status != Trial.TERMINATED:
runner.step()
assert trial.last_result.get("training_iteration") > 1
with pytest.raises(TuneError):
runner.step()
def test_trial_requeue(start_connected_emptyhead_cluster, trainable_id):
"""Removing a node in full cluster causes Trial to be requeued."""
cluster = start_connected_emptyhead_cluster
node = cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
syncer_callback = _PerTrialSyncerCallback(
lambda trial: trial.trainable_name == "__fake")
runner = TrialRunner(BasicVariantGenerator(), callbacks=[syncer_callback])
kwargs = {
"stopping_criterion": {
"training_iteration": 5
},
"checkpoint_freq": 1,
"max_failures": 1,
"remote_checkpoint_dir": MOCK_REMOTE_DIR,
}
trials = [Trial(trainable_id, **kwargs), Trial(trainable_id, **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step() # Start trial
runner.step() # Process result, dispatch save
runner.step() # Process save
running_trials = _get_running_trials(runner)
assert len(running_trials) == 1
assert _check_trial_running(running_trials[0])
cluster.remove_node(node)
cluster.wait_for_nodes()
runner.step() # Process result, dispatch save
runner.step() # Process save (detect error), requeue trial
assert all(t.status == Trial.PENDING
for t in trials), runner.debug_string()
with pytest.raises(TuneError):
runner.step()
def testFailureRecoveryEnabled(self):
ray.init(num_cpus=1, num_gpus=1)
runner = TrialRunner(BasicVariantGenerator())
kwargs = {
"resources": Resources(cpu=1, gpu=1),
"checkpoint_freq": 1,
"max_failures": 1,
"config": {
"mock_error": True,
},
}
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(trials[0].num_failures, 1)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
def testFailFastRaise(self):
ray.init(num_cpus=1, num_gpus=1)
runner = TrialRunner(fail_fast=TrialRunner.RAISE)
kwargs = {
"resources": Resources(cpu=1, gpu=1),
"checkpoint_freq": 1,
"max_failures": 0,
"config": {
"mock_error": True,
"persistent_error": True,
},
}
runner.add_trial(Trial("__fake", **kwargs))
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
runner.step() # Start trial
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step() # Process result, dispatch save
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step() # Process save
with self.assertRaises(Exception):
runner.step() # Error
def testChangeResources(self):
"""Checks that resource requirements can be changed on fly."""
ray.init(num_cpus=2)
class ChangingScheduler(FIFOScheduler):
def __init__(self):
self._has_received_one_trial_result = False
# For figuring out how many runner.step there are.
def has_received_one_trial_result(self):
return self._has_received_one_trial_result
def on_trial_result(self, trial_runner, trial, result):
if result["training_iteration"] == 1:
self._has_received_one_trial_result = True
executor = trial_runner.trial_executor
executor.pause_trial(trial)
trial.update_resources(dict(cpu=2, gpu=0))
return TrialScheduler.NOOP
scheduler = ChangingScheduler()
runner = TrialRunner(scheduler=scheduler)
kwargs = {
"stopping_criterion": {"training_iteration": 2},
"resources": Resources(cpu=1, gpu=0),
}
trials = [Trial("__fake", **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(
runner.trial_executor._pg_manager.occupied_resources().get("CPU"), 1
)
self.assertRaises(
ValueError, lambda: trials[0].update_resources(dict(cpu=2, gpu=0))
)
while not scheduler.has_received_one_trial_result():
runner.step()
self.assertEqual(trials[0].status, Trial.PAUSED)
# extra step for tune loop to stage the resource requests.
runner.step()
self.assertEqual(
runner.trial_executor._pg_manager.occupied_resources().get("CPU"), 2
)
def testResultDone(self):
"""Tests that last_result is marked `done` after trial is complete."""
ray.init(num_cpus=1, num_gpus=1)
runner = TrialRunner()
kwargs = {
"stopping_criterion": {"training_iteration": 2},
"resources": Resources(cpu=1, gpu=1),
}
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step()
self.assertNotEqual(trials[0].last_result.done, True)
runner.step()
self.assertEqual(trials[0].last_result.done, True)
def testStopTrial(self):
ray.init(num_cpus=4, num_gpus=2)
runner = TrialRunner()
kwargs = {
"stopping_criterion": {
"training_iteration": 5
},
"resources": Resources(cpu=1, gpu=1),
}
trials = [
Trial("__fake", **kwargs),
Trial("__fake", **kwargs),
Trial("__fake", **kwargs),
Trial("__fake", **kwargs)
]
for t in trials:
runner.add_trial(t)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(trials[1].status, Trial.PENDING)
# Stop trial while running
runner.stop_trial(trials[0])
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.PENDING)
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.RUNNING)
self.assertEqual(trials[-1].status, Trial.PENDING)
# Stop trial while pending
runner.stop_trial(trials[-1])
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.RUNNING)
self.assertEqual(trials[-1].status, Trial.TERMINATED)
time.sleep(2) # Wait for stopped placement group to free resources
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.RUNNING)
self.assertEqual(trials[2].status, Trial.RUNNING)
self.assertEqual(trials[-1].status, Trial.TERMINATED)
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 testStopTrial(self):
ray.init(num_cpus=4, num_gpus=2)
runner = TrialRunner(BasicVariantGenerator())
kwargs = {
"stopping_criterion": {
"training_iteration": 5
},
"resources": Resources(cpu=1, gpu=1),
}
trials = [
Trial("__fake", **kwargs),
Trial("__fake", **kwargs),
Trial("__fake", **kwargs),
Trial("__fake", **kwargs)
]
for t in trials:
runner.add_trial(t)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(trials[1].status, Trial.PENDING)
# Stop trial while running
runner.stop_trial(trials[0])
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.PENDING)
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.RUNNING)
self.assertEqual(trials[-1].status, Trial.PENDING)
# Stop trial while pending
runner.stop_trial(trials[-1])
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.RUNNING)
self.assertEqual(trials[-1].status, Trial.TERMINATED)
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.RUNNING)
self.assertEqual(trials[2].status, Trial.RUNNING)
self.assertEqual(trials[-1].status, Trial.TERMINATED)
def test_trial_processed_after_node_failure(start_connected_emptyhead_cluster):
"""Tests that Tune processes a trial as failed if its node died."""
cluster = start_connected_emptyhead_cluster
node = cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
runner = TrialRunner(BasicVariantGenerator())
mock_process_failure = MagicMock(side_effect=runner._process_trial_failure)
runner._process_trial_failure = mock_process_failure
runner.add_trial(Trial("__fake"))
runner.step()
runner.step()
assert not mock_process_failure.called
cluster.remove_node(node)
runner.step()
if not mock_process_failure.called:
runner.step()
assert mock_process_failure.called
def testCheckpointingAtEnd(self):
ray.init(num_cpus=1, num_gpus=1)
runner = TrialRunner()
kwargs = {
"stopping_criterion": {
"training_iteration": 2
},
"checkpoint_at_end": True,
"resources": Resources(cpu=1, gpu=1),
}
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
runner.step() # Start trial
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step() # Process result
runner.step() # Process result, dispatch save
self.assertEqual(trials[0].last_result[DONE], True)
runner.step() # Process save
self.assertEqual(trials[0].has_checkpoint(), True)
def test_trial_requeue(start_connected_emptyhead_cluster, trainable_id):
"""Removing a node in full cluster causes Trial to be requeued."""
os.environ["TUNE_MAX_PENDING_TRIALS_PG"] = "1"
cluster = start_connected_emptyhead_cluster
node = cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
syncer_callback = _PerTrialSyncerCallback(
lambda trial: trial.trainable_name == "__fake")
runner = TrialRunner(BasicVariantGenerator(),
callbacks=[syncer_callback]) # noqa
kwargs = {
"stopping_criterion": {
"training_iteration": 5
},
"checkpoint_freq": 1,
"max_failures": 1,
}
if trainable_id == "__fake_durable":
kwargs["remote_checkpoint_dir"] = MOCK_REMOTE_DIR
trials = [Trial(trainable_id, **kwargs), Trial(trainable_id, **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step() # Start trial
runner.step() # Process result, dispatch save
runner.step() # Process save
running_trials = _get_running_trials(runner)
assert len(running_trials) == 1
assert _check_trial_running(running_trials[0])
cluster.remove_node(node)
cluster.wait_for_nodes()
time.sleep(0.1) # Sleep so that next step() refreshes cluster resources
runner.step() # Process result, dispatch save
runner.step() # Process save (detect error), requeue trial
assert all(t.status == Trial.PENDING
for t in trials), runner.debug_string()
def testRestoreMetricsAfterCheckpointing(self):
ray.init(num_cpus=1, num_gpus=1)
observer = TrialResultObserver()
runner = TrialRunner(callbacks=[observer])
kwargs = {
"stopping_criterion": {
"training_iteration": 2
},
"resources": Resources(cpu=1, gpu=1),
"checkpoint_freq": 1,
}
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
while not runner.is_finished():
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
kwargs["restore_path"] = trials[0].checkpoint.value
kwargs.pop("stopping_criterion")
kwargs.pop("checkpoint_freq") # No checkpointing for next trial
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
observer.reset()
while not observer.just_received_a_result():
runner.step()
self.assertEqual(trials[1].last_result["timesteps_since_restore"], 10)
self.assertEqual(trials[1].last_result["iterations_since_restore"], 1)
self.assertGreater(trials[1].last_result["time_since_restore"], 0)
while not observer.just_received_a_result():
runner.step()
self.assertEqual(trials[1].last_result["timesteps_since_restore"], 20)
self.assertEqual(trials[1].last_result["iterations_since_restore"], 2)
self.assertGreater(trials[1].last_result["time_since_restore"], 0)
self.addCleanup(os.remove, trials[0].checkpoint.value)
def testQueueFilling(self):
os.environ["TUNE_MAX_PENDING_TRIALS_PG"] = "1"
ray.init(num_cpus=4)
def f1(config):
for i in range(10):
yield i
tune.register_trainable("f1", f1)
search_alg = BasicVariantGenerator()
search_alg.add_configurations(
{
"foo": {
"run": "f1",
"num_samples": 100,
"config": {
"a": tune.sample_from(lambda spec: 5.0 / 7),
"b": tune.sample_from(lambda spec: "long" * 40),
},
"resources_per_trial": {"cpu": 2},
}
}
)
runner = TrialRunner(search_alg=search_alg)
runner.step()
runner.step()
runner.step()
self.assertEqual(len(runner._trials), 3)
runner.step()
self.assertEqual(len(runner._trials), 3)
self.assertEqual(runner._trials[0].status, Trial.RUNNING)
self.assertEqual(runner._trials[1].status, Trial.RUNNING)
self.assertEqual(runner._trials[2].status, Trial.PENDING)
def test_counting_resources(start_connected_cluster):
"""Tests that Tune accounting is consistent with actual cluster."""
cluster = start_connected_cluster
nodes = []
assert ray.cluster_resources()["CPU"] == 1
runner = TrialRunner(BasicVariantGenerator())
kwargs = {"stopping_criterion": {"training_iteration": 10}}
trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step()
running_trials = _get_running_trials(runner)
assert len(running_trials) == 1
assert _check_trial_running(running_trials[0])
assert ray.available_resources().get("CPU", 0) == 0
nodes += [cluster.add_node(num_cpus=1)]
cluster.wait_for_nodes()
assert ray.cluster_resources()["CPU"] == 2
cluster.remove_node(nodes.pop())
cluster.wait_for_nodes()
assert ray.cluster_resources()["CPU"] == 1
runner.step()
# Only 1 trial can be running due to resource limitation.
assert sum(t.status == Trial.RUNNING for t in runner.get_trials()) == 1
for i in range(5):
nodes += [cluster.add_node(num_cpus=1)]
cluster.wait_for_nodes()
assert ray.cluster_resources()["CPU"] == 6
# This is to make sure that pg is ready for the previous pending trial,
# so that when runner.step() is called next, the trial can be started in
# the same event loop.
time.sleep(5)
runner.step()
assert sum(t.status == Trial.RUNNING for t in runner.get_trials()) == 2
def testMultiStepRun2(self):
"""Checks that runner.step throws when overstepping."""
ray.init(num_cpus=1)
runner = TrialRunner(BasicVariantGenerator())
kwargs = {
"stopping_criterion": {
"training_iteration": 2
},
"resources": Resources(cpu=1, gpu=0),
}
trials = [Trial("__fake", **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step()
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.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
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 testFailureRecoveryNodeRemoval(self):
# Node removal simulation only works with resource requests
os.environ["TUNE_PLACEMENT_GROUP_AUTO_DISABLED"] = "1"
ray.init(num_cpus=1, num_gpus=1)
searchalg, scheduler = create_mock_components()
runner = TrialRunner(searchalg, scheduler=scheduler)
kwargs = {
"resources": Resources(cpu=1, gpu=1),
"checkpoint_freq": 1,
"max_failures": 1,
"config": {
"mock_error": True,
},
}
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
with patch("ray.cluster_resources") as resource_mock:
resource_mock.return_value = {"CPU": 1, "GPU": 1}
runner.step() # Start trial
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step() # Process result, dispatch save
runner.step() # Process save
self.assertEqual(trials[0].status, Trial.RUNNING)
# Mimic a node failure
resource_mock.return_value = {"CPU": 0, "GPU": 0}
runner.step() # Detect node failure
self.assertEqual(trials[0].status, Trial.PENDING)
self.assertEqual(trials[0].num_failures, 1)
self.assertEqual(len(searchalg.errored_trials), 0)
self.assertEqual(len(scheduler.errored_trials), 1)
class TrialRunnerCallbacks(unittest.TestCase):
def setUp(self):
self.tmpdir = tempfile.mkdtemp()
self.callback = TestCallback()
self.executor = _MockTrialExecutor()
self.trial_runner = TrialRunner(trial_executor=self.executor,
callbacks=[self.callback])
def tearDown(self):
ray.shutdown()
_register_all() # re-register the evicted objects
if "CUDA_VISIBLE_DEVICES" in os.environ:
del os.environ["CUDA_VISIBLE_DEVICES"]
shutil.rmtree(self.tmpdir)
def testCallbackSteps(self):
trials = [
Trial("__fake", trial_id="one"),
Trial("__fake", trial_id="two")
]
for t in trials:
self.trial_runner.add_trial(t)
self.executor.next_trial = trials[0]
self.trial_runner.step()
# Trial 1 has been started
self.assertEqual(self.callback.state["trial_start"]["iteration"], 0)
self.assertEqual(self.callback.state["trial_start"]["trial"].trial_id,
"one")
# All these events haven't happened, yet
self.assertTrue(
all(k not in self.callback.state for k in [
"trial_restore", "trial_save", "trial_result",
"trial_complete", "trial_fail"
]))
self.executor.next_trial = trials[1]
self.trial_runner.step()
# Iteration not increased yet
self.assertEqual(self.callback.state["step_begin"]["iteration"], 1)
# Iteration increased
self.assertEqual(self.callback.state["step_end"]["iteration"], 2)
# Second trial has been just started
self.assertEqual(self.callback.state["trial_start"]["iteration"], 1)
self.assertEqual(self.callback.state["trial_start"]["trial"].trial_id,
"two")
cp = Checkpoint(Checkpoint.PERSISTENT, "__checkpoint",
{TRAINING_ITERATION: 0})
# Let the first trial save a checkpoint
self.executor.next_trial = trials[0]
trials[0].saving_to = cp
self.trial_runner.step()
self.assertEqual(self.callback.state["trial_save"]["iteration"], 2)
self.assertEqual(self.callback.state["trial_save"]["trial"].trial_id,
"one")
# Let the second trial send a result
result = {TRAINING_ITERATION: 1, "metric": 800, "done": False}
self.executor.results[trials[1]] = result
self.executor.next_trial = trials[1]
self.assertEqual(trials[1].last_result, {})
self.trial_runner.step()
self.assertEqual(self.callback.state["trial_result"]["iteration"], 3)
self.assertEqual(self.callback.state["trial_result"]["trial"].trial_id,
"two")
self.assertEqual(
self.callback.state["trial_result"]["result"]["metric"], 800)
self.assertEqual(trials[1].last_result["metric"], 800)
# Let the second trial restore from a checkpoint
trials[1].restoring_from = cp
self.executor.results[trials[1]] = trials[1].last_result
self.trial_runner.step()
self.assertEqual(self.callback.state["trial_restore"]["iteration"], 4)
self.assertEqual(
self.callback.state["trial_restore"]["trial"].trial_id, "two")
# Let the second trial finish
trials[1].restoring_from = None
self.executor.results[trials[1]] = {
TRAINING_ITERATION: 2,
"metric": 900,
"done": True
}
self.trial_runner.step()
self.assertEqual(self.callback.state["trial_complete"]["iteration"], 5)
self.assertEqual(
self.callback.state["trial_complete"]["trial"].trial_id, "two")
# Let the first trial error
self.executor.failed_trial = trials[0]
self.trial_runner.step()
self.assertEqual(self.callback.state["trial_fail"]["iteration"], 6)
self.assertEqual(self.callback.state["trial_fail"]["trial"].trial_id,
"one")
def testCallbacksEndToEnd(self):
def train(config):
if config["do"] == "save":
with tune.checkpoint_dir(0):
pass
tune.report(metric=1)
elif config["do"] == "fail":
raise RuntimeError("I am failing on purpose.")
elif config["do"] == "delay":
time.sleep(2)
tune.report(metric=20)
config = {"do": tune.grid_search(["save", "fail", "delay"])}
tune.run(train,
config=config,
raise_on_failed_trial=False,
callbacks=[self.callback])
self.assertEqual(
self.callback.state["trial_fail"]["trial"].config["do"], "fail")
self.assertEqual(
self.callback.state["trial_save"]["trial"].config["do"], "save")
self.assertEqual(
self.callback.state["trial_result"]["trial"].config["do"], "delay")
self.assertEqual(
self.callback.state["trial_complete"]["trial"].config["do"],
"delay")
def test_trial_migration(start_connected_emptyhead_cluster):
"""Removing a node while cluster has space should migrate trial.
The trial state should also be consistent with the checkpoint.
"""
cluster = start_connected_emptyhead_cluster
node = cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
runner = TrialRunner(BasicVariantGenerator())
kwargs = {
"stopping_criterion": {
"training_iteration": 3
},
"checkpoint_freq": 2,
"max_failures": 2
}
# Test recovery of trial that hasn't been checkpointed
t = Trial("__fake", **kwargs)
runner.add_trial(t)
runner.step() # start
runner.step() # 1 result
assert t.last_result
node2 = cluster.add_node(num_cpus=1)
cluster.remove_node(node)
cluster.wait_for_nodes()
runner.step() # Recovery step
# TODO(rliaw): This assertion is not critical but will not pass
# because checkpoint handling is messy and should be refactored
# rather than hotfixed.
# assert t.last_result is None, "Trial result not restored correctly."
for i in range(3):
runner.step()
assert t.status == Trial.TERMINATED
# Test recovery of trial that has been checkpointed
t2 = Trial("__fake", **kwargs)
runner.add_trial(t2)
runner.step() # start
runner.step() # 1 result
runner.step() # 2 result and checkpoint
assert t2.has_checkpoint()
node3 = cluster.add_node(num_cpus=1)
cluster.remove_node(node2)
cluster.wait_for_nodes()
runner.step() # Recovery step
assert t2.last_result["training_iteration"] == 2
for i in range(1):
runner.step()
assert t2.status == Trial.TERMINATED
# Test recovery of trial that won't be checkpointed
t3 = Trial("__fake", **{"stopping_criterion": {"training_iteration": 3}})
runner.add_trial(t3)
runner.step() # start
runner.step() # 1 result
cluster.add_node(num_cpus=1)
cluster.remove_node(node3)
cluster.wait_for_nodes()
runner.step() # Error handling step
assert t3.status == Trial.ERROR
with pytest.raises(TuneError):
runner.step()
def testUserCheckpointBuffered(self):
os.environ["TUNE_RESULT_BUFFER_LENGTH"] = "8"
os.environ["TUNE_RESULT_BUFFER_MIN_TIME_S"] = "1"
def num_checkpoints(trial):
return sum(
item.startswith("checkpoint_")
for item in os.listdir(trial.logdir))
ray.init(num_cpus=3)
runner = TrialRunner(
local_checkpoint_dir=self.tmpdir, checkpoint_period=0)
runner.add_trial(Trial("__fake", config={"user_checkpoint_freq": 10}))
trials = runner.get_trials()
runner.step() # Start trial, schedule 1-8
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(ray.get(trials[0].runner.set_info.remote(1)), 1)
self.assertEqual(num_checkpoints(trials[0]), 0)
runner.step() # Process results 0-8, schedule 9-11 (CP)
self.assertEqual(trials[0].last_result.get(TRAINING_ITERATION), 8)
self.assertFalse(trials[0].has_checkpoint())
self.assertEqual(num_checkpoints(trials[0]), 0)
runner.step() # Process results 9-11
runner.step() # handle CP, schedule 12-19
self.assertEqual(trials[0].last_result.get(TRAINING_ITERATION), 11)
self.assertTrue(trials[0].has_checkpoint())
self.assertEqual(num_checkpoints(trials[0]), 1)
runner.step() # Process results 12-19, schedule 20-21
self.assertEqual(trials[0].last_result.get(TRAINING_ITERATION), 19)
self.assertTrue(trials[0].has_checkpoint())
self.assertEqual(num_checkpoints(trials[0]), 1)
runner.step() # Process results 20-21
runner.step() # handle CP, schedule 21-29
self.assertEqual(trials[0].last_result.get(TRAINING_ITERATION), 21)
self.assertTrue(trials[0].has_checkpoint())
self.assertEqual(num_checkpoints(trials[0]), 2)
runner.step() # Process results 21-29, schedule 30-31
self.assertEqual(trials[0].last_result.get(TRAINING_ITERATION), 29)
self.assertTrue(trials[0].has_checkpoint())
self.assertTrue(trials[0].has_checkpoint())
self.assertEqual(num_checkpoints(trials[0]), 2)
def testTrialSaveRestore(self):
"""Creates different trials to test runner.checkpoint/restore."""
ray.init(num_cpus=3)
runner = TrialRunner(
local_checkpoint_dir=self.tmpdir, checkpoint_period=0)
trials = [
Trial(
"__fake",
trial_id="trial_terminate",
stopping_criterion={"training_iteration": 1},
checkpoint_freq=1)
]
runner.add_trial(trials[0])
runner.step() # Start trial
runner.step() # Process result, dispatch save
runner.step() # Process save
self.assertEqual(trials[0].status, Trial.TERMINATED)
trials += [
Trial(
"__fake",
trial_id="trial_fail",
stopping_criterion={"training_iteration": 3},
checkpoint_freq=1,
config={"mock_error": True})
]
runner.add_trial(trials[1])
runner.step() # Start trial
runner.step() # Process result, dispatch save
runner.step() # Process save
runner.step() # Error
self.assertEqual(trials[1].status, Trial.ERROR)
trials += [
Trial(
"__fake",
trial_id="trial_succ",
stopping_criterion={"training_iteration": 2},
checkpoint_freq=1)
]
runner.add_trial(trials[2])
runner.step() # Start trial
self.assertEqual(len(runner.trial_executor.get_checkpoints()), 3)
self.assertEqual(trials[2].status, Trial.RUNNING)
runner2 = TrialRunner(resume="LOCAL", local_checkpoint_dir=self.tmpdir)
for tid in ["trial_terminate", "trial_fail"]:
original_trial = runner.get_trial(tid)
restored_trial = runner2.get_trial(tid)
self.assertEqual(original_trial.status, restored_trial.status)
restored_trial = runner2.get_trial("trial_succ")
self.assertEqual(Trial.PENDING, restored_trial.status)
runner2.step() # Start trial
runner2.step() # Process result, dispatch save
runner2.step() # Process save
runner2.step() # Process result, dispatch save
runner2.step() # Process save
self.assertRaises(TuneError, runner2.step)
def testSearcherSaveRestore(self):
ray.init(num_cpus=8, local_mode=True)
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
searcher = create_searcher()
runner = TrialRunner(
search_alg=searcher,
local_checkpoint_dir=self.tmpdir,
checkpoint_period=-1)
for i in range(6):
runner.step()
assert len(
runner.get_trials()) == 6, [t.config for t in runner.get_trials()]
runner.checkpoint()
trials = runner.get_trials()
[
runner.trial_executor.stop_trial(t) for t in trials
if t.status is not Trial.ERROR
]
del runner
# stop_all(runner.get_trials())
searcher = create_searcher()
runner2 = TrialRunner(
search_alg=searcher,
local_checkpoint_dir=self.tmpdir,
resume="LOCAL")
assert len(runner2.get_trials()) == 6, [
t.config for t in runner2.get_trials()
]
def trial_statuses():
return [t.status for t in runner2.get_trials()]
def num_running_trials():
return sum(t.status == Trial.RUNNING for t in runner2.get_trials())
for i in range(6):
runner2.step()
assert len(set(trial_statuses())) == 1
assert Trial.RUNNING in trial_statuses()
for i in range(20):
runner2.step()
assert 1 <= num_running_trials() <= 6
evaluated = [
t.evaluated_params["test_variable"] for t in runner2.get_trials()
]
count = Counter(evaluated)
assert all(v <= 3 for v in count.values())
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 test_cluster_down_simple(start_connected_cluster, tmpdir, trainable_id):
"""Tests that TrialRunner save/restore works on cluster shutdown."""
cluster = start_connected_cluster
cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
dirpath = str(tmpdir)
runner = TrialRunner(local_checkpoint_dir=dirpath, checkpoint_period=0)
kwargs = {
"stopping_criterion": {
"training_iteration": 2
},
"checkpoint_freq": 1,
"max_failures": 1,
"remote_checkpoint_dir": MOCK_REMOTE_DIR,
"sync_to_driver_fn": trainable_id == "__fake",
}
trials = [Trial(trainable_id, **kwargs), Trial(trainable_id, **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step() # start
runner.step() # start2
runner.step() # step
assert all(t.status == Trial.RUNNING for t in runner.get_trials())
runner.checkpoint()
ray.shutdown()
cluster.shutdown()
cluster = _start_new_cluster()
runner = TrialRunner(resume="LOCAL", local_checkpoint_dir=dirpath)
runner.step() # start
runner.step() # process restore
runner.step() # start2
for i in range(3):
runner.step()
with pytest.raises(TuneError):
runner.step()
assert all(t.status == Trial.TERMINATED for t in runner.get_trials())
ray.shutdown()
cluster.shutdown()
def test_trial_migration(start_connected_emptyhead_cluster, trainable_id):
"""Removing a node while cluster has space should migrate trial.
The trial state should also be consistent with the checkpoint.
"""
cluster = start_connected_emptyhead_cluster
node = cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
runner = TrialRunner(BasicVariantGenerator())
kwargs = {
"stopping_criterion": {
"training_iteration": 4
},
"checkpoint_freq": 2,
"max_failures": 2,
"remote_checkpoint_dir": MOCK_REMOTE_DIR,
"sync_to_driver_fn": trainable_id == "__fake",
}
# Test recovery of trial that hasn't been checkpointed
t = Trial(trainable_id, **kwargs)
runner.add_trial(t)
runner.step() # start
runner.step() # 1 result
assert t.last_result
node2 = cluster.add_node(num_cpus=1)
cluster.remove_node(node)
cluster.wait_for_nodes()
runner.step() # Recovery step
# TODO(rliaw): This assertion is not critical but will not pass
# because checkpoint handling is messy and should be refactored
# rather than hotfixed.
# assert t.last_result is None, "Trial result not restored correctly."
for i in range(4):
runner.step()
assert t.status == Trial.TERMINATED
# Test recovery of trial that has been checkpointed
t2 = Trial(trainable_id, **kwargs)
runner.add_trial(t2)
runner.step() # start
runner.step() # 1 result
runner.step() # 2 result and checkpoint
assert t2.has_checkpoint()
node3 = cluster.add_node(num_cpus=1)
cluster.remove_node(node2)
cluster.wait_for_nodes()
runner.step() # 3 result + start and fail 4 result
runner.step() # Recovery step
runner.step() # Process recovery
runner.step() # result
if t2.status != Trial.TERMINATED:
runner.step()
assert t2.status == Trial.TERMINATED, runner.debug_string()
# Test recovery of trial that won't be checkpointed
kwargs = {
"stopping_criterion": {
"training_iteration": 3
},
"remote_checkpoint_dir": MOCK_REMOTE_DIR,
"sync_to_driver_fn": trainable_id == "__fake",
}
t3 = Trial(trainable_id, **kwargs)
runner.add_trial(t3)
runner.step() # start
runner.step() # 1 result
cluster.add_node(num_cpus=1)
cluster.remove_node(node3)
cluster.wait_for_nodes()
runner.step() # Error handling step
if t3.status != Trial.ERROR:
runner.step()
assert t3.status == Trial.ERROR, runner.debug_string()
with pytest.raises(TuneError):
runner.step()
def run_experiments(experiments,
search_alg=None,
scheduler=None,
with_server=False,
server_port=TuneServer.DEFAULT_PORT,
verbose=True,
resume=False,
queue_trials=False,
trial_executor=None,
raise_on_failed_trial=True):
"""Runs and blocks until all trials finish.
Args:
experiments (Experiment | list | dict): Experiments to run. Will be
passed to `search_alg` via `add_configurations`.
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 (bool): How much output should be printed for each trial.
resume (bool|"prompt"): If checkpoint exists, the experiment will
resume from there. If resume is "prompt", Tune will prompt if
checkpoint detected.
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.
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.
Examples:
>>> experiment_spec = Experiment("experiment", my_func)
>>> run_experiments(experiments=experiment_spec)
>>> experiment_spec = {"experiment": {"run": my_func}}
>>> run_experiments(experiments=experiment_spec)
>>> run_experiments(
>>> experiments=experiment_spec,
>>> scheduler=MedianStoppingRule(...))
>>> run_experiments(
>>> experiments=experiment_spec,
>>> search_alg=SearchAlgorithm(),
>>> scheduler=MedianStoppingRule(...))
Returns:
List of Trial objects, holding data for each executed trial.
"""
# This is important to do this here
# because it schematize the experiments
# and it conducts the implicit registration.
experiments = convert_to_experiment_list(experiments)
checkpoint_dir = _find_checkpoint_dir(experiments)
runner = None
restore = False
if os.path.exists(
os.path.join(checkpoint_dir, TrialRunner.CKPT_FILE_NAME)):
if resume == "prompt":
msg = ("Found incomplete experiment at {}. "
"Would you like to resume it?".format(checkpoint_dir))
restore = click.confirm(msg, default=False)
if restore:
logger.info("Tip: to always resume, "
"pass resume=True to run_experiments()")
else:
logger.info("Tip: to always start a new experiment, "
"pass resume=False to run_experiments()")
elif resume:
restore = True
else:
logger.info(
"Tip: to resume incomplete experiments, "
"pass resume='prompt' or resume=True to run_experiments()")
else:
logger.info(
"Did not find checkpoint file in {}.".format(checkpoint_dir))
if restore:
runner = try_restore_runner(checkpoint_dir, search_alg, scheduler,
trial_executor)
else:
logger.info("Starting a new experiment.")
if not runner:
if scheduler is None:
scheduler = FIFOScheduler()
if search_alg is None:
search_alg = BasicVariantGenerator()
search_alg.add_configurations(experiments)
runner = TrialRunner(
search_alg,
scheduler=scheduler,
metadata_checkpoint_dir=checkpoint_dir,
launch_web_server=with_server,
server_port=server_port,
verbose=verbose,
queue_trials=queue_trials,
trial_executor=trial_executor)
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))
wait_for_log_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)
return runner.get_trials()
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_experiments(experiments,
search_alg=None,
scheduler=None,
with_server=False,
server_port=TuneServer.DEFAULT_PORT,
verbose=2,
resume=False,
queue_trials=False,
trial_executor=None,
raise_on_failed_trial=True):
"""Runs and blocks until all trials finish.
Args:
experiments (Experiment | list | dict): Experiments to run. Will be
passed to `search_alg` via `add_configurations`.
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 (bool|"prompt"): If checkpoint exists, the experiment will
resume from there. If resume is "prompt", Tune will prompt if
checkpoint detected.
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.
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.
Examples:
>>> experiment_spec = Experiment("experiment", my_func)
>>> run_experiments(experiments=experiment_spec)
>>> experiment_spec = {"experiment": {"run": my_func}}
>>> run_experiments(experiments=experiment_spec)
>>> run_experiments(
>>> experiments=experiment_spec,
>>> scheduler=MedianStoppingRule(...))
>>> run_experiments(
>>> experiments=experiment_spec,
>>> search_alg=SearchAlgorithm(),
>>> scheduler=MedianStoppingRule(...))
Returns:
List of Trial objects, holding data for each executed trial.
"""
# This is important to do this here
# because it schematize the experiments
# and it conducts the implicit registration.
experiments = convert_to_experiment_list(experiments)
checkpoint_dir = _find_checkpoint_dir(experiments)
runner = None
restore = False
if TrialRunner.checkpoint_exists(checkpoint_dir):
if resume == "prompt":
msg = ("Found incomplete experiment at {}. "
"Would you like to resume it?".format(checkpoint_dir))
restore = click.confirm(msg, default=False)
if restore:
logger.info("Tip: to always resume, "
"pass resume=True to run_experiments()")
else:
logger.info("Tip: to always start a new experiment, "
"pass resume=False to run_experiments()")
elif resume:
restore = True
else:
logger.info(
"Tip: to resume incomplete experiments, "
"pass resume='prompt' or resume=True to run_experiments()")
else:
logger.info(
"Did not find checkpoint file in {}.".format(checkpoint_dir))
if restore:
runner = try_restore_runner(checkpoint_dir, search_alg, scheduler,
trial_executor)
else:
logger.info("Starting a new experiment.")
if not runner:
if scheduler is None:
scheduler = FIFOScheduler()
if search_alg is None:
search_alg = BasicVariantGenerator()
search_alg.add_configurations(experiments)
runner = TrialRunner(search_alg,
scheduler=scheduler,
metadata_checkpoint_dir=checkpoint_dir,
launch_web_server=with_server,
server_port=server_port,
verbose=bool(verbose > 1),
queue_trials=queue_trials,
trial_executor=trial_executor)
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_log_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)
return runner.get_trials()
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 test_migration_checkpoint_removal(start_connected_emptyhead_cluster,
trainable_id):
"""Test checks that trial restarts if checkpoint is lost w/ node fail."""
cluster = start_connected_emptyhead_cluster
node = cluster.add_node(num_cpus=1)
cluster.wait_for_nodes()
runner = TrialRunner(BasicVariantGenerator())
kwargs = {
"stopping_criterion": {
"training_iteration": 4
},
"checkpoint_freq": 2,
"max_failures": 2,
"remote_checkpoint_dir": MOCK_REMOTE_DIR,
"sync_to_driver_fn": trainable_id == "__fake_remote",
}
# The following patches only affect __fake_remote.
find_checkpoint_dir = TrainableUtil.find_checkpoint_dir
with patch("ray.tune.logger.get_node_syncer") as mock_get_node_syncer:
trainable_util = "ray.tune.ray_trial_executor.TrainableUtil"
with patch(trainable_util + ".find_checkpoint_dir") as mock_find_dir:
def mock_get_syncer_fn(local_dir, remote_dir, sync_function):
client = mock_storage_client()
return MockNodeSyncer(local_dir, remote_dir, client)
mock_get_node_syncer.side_effect = mock_get_syncer_fn
def mock_find_dir_fn(checkpoint_path):
"""Converts back to local path first."""
checkpoint_path = checkpoint_path[len(MOCK_REMOTE_DIR):]
checkpoint_path = os.path.join("/", checkpoint_path)
return find_checkpoint_dir(checkpoint_path)
# __fake_remote trainables save to a separate "remote" directory.
# TrainableUtil will not check this path unless we mock it.
mock_find_dir.side_effect = mock_find_dir_fn
# Test recovery of trial that has been checkpointed
t1 = Trial(trainable_id, **kwargs)
runner.add_trial(t1)
# Start trial, process result (x2), process save
for _ in range(4):
runner.step()
assert t1.has_checkpoint()
cluster.add_node(num_cpus=1)
cluster.remove_node(node)
cluster.wait_for_nodes()
shutil.rmtree(os.path.dirname(t1.checkpoint.value))
runner.step() # Collect result 3, kick off + fail result 4
runner.step() # Dispatch restore
runner.step() # Process restore + step 4
for _ in range(3):
if t1.status != Trial.TERMINATED:
runner.step()
assert t1.status == Trial.TERMINATED, runner.debug_string()
