def tearDown(self):
print("Tearing down....")
try:
self.runner._server.shutdown()
self.runner = None
except Exception as e:
print(e)
ray.shutdown()
_register_all()
def _start_new_cluster():
cluster = Cluster(
initialize_head=True,
connect=True,
head_node_args={
"num_cpus": 1,
"_internal_config": json.dumps({
"num_heartbeats_timeout": 10
})
})
# Pytest doesn't play nicely with imports
_register_all()
return cluster
def start_connected_emptyhead_cluster():
"""Starts head with no resources."""
cluster = Cluster(
initialize_head=True,
connect=True,
head_node_args={
"num_cpus": 0,
"_internal_config": json.dumps({
"num_heartbeats_timeout": 10
})
})
# Pytest doesn't play nicely with imports
_register_all()
yield cluster
# The code after the yield will run as teardown code.
ray.shutdown()
cluster.shutdown()
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 tearDown(self):
ray.shutdown()
_register_all()
def tearDown(self):
shutil.rmtree(self.absolute_local_dir, ignore_errors=True)
self.absolute_local_dir = None
ray.shutdown()
# Without this line, test_tune_server.testAddTrial would fail.
_register_all()
def do_test(self, alg: str, config: dict, fn=None):
fn = fn or self._do_test_fault_ignore
try:
fn(alg, config)
finally:
_register_all() # re-register the evicted objects
from __future__ import print_function
import random
import unittest
import numpy as np
from ray.tune.hyperband import HyperBandScheduler
from ray.tune.async_hyperband import AsyncHyperBandScheduler
from ray.tune.pbt import PopulationBasedTraining, explore
from ray.tune.median_stopping_rule import MedianStoppingRule
from ray.tune.result import TrainingResult
from ray.tune.trial import Trial, Resources
from ray.tune.trial_scheduler import TrialScheduler
from ray.rllib import _register_all
_register_all()
def result(t, rew):
return TrainingResult(time_total_s=t,
episode_reward_mean=rew,
training_iteration=int(t))
class EarlyStoppingSuite(unittest.TestCase):
def basicSetup(self, rule):
t1 = Trial("PPO") # mean is 450, max 900, t_max=10
t2 = Trial("PPO") # mean is 450, max 450, t_max=5
for i in range(10):
self.assertEqual(
rule.on_trial_result(None, t1, result(i, i * 100)),
def tearDown(self):
ray.shutdown()
_register_all() # re-register the evicted objects
def start_ray():
ray.init()
_register_all()
yield
ray.shutdown()
def tearDown(self):
ray.worker.cleanup()
_register_all() # re-register the evicted objects
def setUp(self):
_register_all() # re-register the evicted objects
def tearDownClass(cls):
ray.shutdown()
_register_all()
def setUp(self):
# Wait up to five seconds for placement groups when starting a trial
os.environ["TUNE_PLACEMENT_GROUP_WAIT_S"] = "5"
# Block for results even when placement groups are pending
os.environ["TUNE_TRIAL_STARTUP_GRACE_PERIOD"] = "0"
_register_all() # re-register the evicted objects
#!/usr/bin/env python
import os
import sys
if __name__ == "__main__":
# Do not import tf for testing purposes.
os.environ["RLLIB_TEST_NO_TF_IMPORT"] = "1"
# Test registering (includes importing) all Trainers.
from ray.rllib import _register_all
# This should surface any dependency on tf, e.g. inside function
# signatures/typehints.
_register_all()
from ray.rllib.agents.a3c import A2CTrainer
assert ("tensorflow" not in sys.modules
), "`tensorflow` initially present, when it shouldn't!"
# Note: No ray.init(), to test it works without Ray
trainer = A2CTrainer(env="CartPole-v0",
config={
"framework": "torch",
"num_workers": 0
})
trainer.train()
assert (
"tensorflow" not in sys.modules
def setUp(self):
self.trial_executor = RayTrialExecutor(queue_trials=False)
ray.init()
_register_all() # Needed for flaky tests
def start_ray():
ray.init(log_to_driver=False, local_mode=True)
_register_all()
yield
ray.shutdown()
def tearDown(self):
ray.shutdown()
_register_all() # re-register the evicted objects
def setUp(self):
self.trial_executor = RayTrialExecutor()
ray.init(num_cpus=2, ignore_reinit_error=True)
_register_all() # Needed for flaky tests
def tearDown(self):
shutil.rmtree(self.logdir)
ray.shutdown()
_register_all()
def tearDown(self):
ray.shutdown()
_register_all() # re-register the evicted objects
shutil.rmtree(self.tmpdir)
def start_ray():
ray.init()
_register_all()
yield
ray.shutdown()
def execute(self):
timesteps = 0
best_period_value = None
if self.pr.agent.name() == "A2C":
trainer = A2CTrainer(config=self.rllib_config,
logger_creator=rllib_logger_creator)
elif self.pr.agent.name() == "PPO":
trainer = PPOTrainer(config=self.rllib_config,
logger_creator=rllib_logger_creator)
# import pdb; pdb.set_trace()
else:
raise ValueError('There is no rllib trainer with name ' +
self.pr.agent.name())
tf_writer = SummaryWriter(
self.pr.save_logs_to) if self.pr.save_logs_to else None
reward_metric = Metric(short_name='rews',
long_name='trajectory reward',
formatting_string='{:5.1f}',
higher_is_better=True)
time_step_metric = Metric(short_name='steps',
long_name='total number of steps',
formatting_string='{:5.1f}',
higher_is_better=True)
metrics = [reward_metric, time_step_metric]
if self.pr.train:
start_time = time.time()
policy_save_tag = 0
while timesteps < self.pr.total_steps:
result = trainer.train()
timesteps = result["timesteps_total"]
reward_metric.log(result['evaluation']['episode_reward_mean'])
time_step_metric.log(result['evaluation']['episode_len_mean'])
# import pdb; pdb.set_trace()
# # Get a metric list from each environment.
# if hasattr(trainer, "evaluation_workers"):
# metric_lists = sum(trainer.evaluation_workers.foreach_worker(lambda w: w.foreach_env(lambda e: e.metrics)), [])
# else:
# metric_lists = sum(trainer.workers.foreach_worker(lambda w: w.foreach_env(lambda e: e.metrics)), [])
# metrics = metric_lists[0]
# # Aggregate metrics from all other environments.
# for metric_list in metric_lists[1:]:
# for i, metric in enumerate(metric_list):
# metrics[i]._values.extend(metric._values)
save_logs_to = self.pr.save_logs_to
model_save_paths_dict = self.pr.model_save_paths_dict
# Consider whether to save a model.
saved = False
if model_save_paths_dict is not None and metrics[
0].currently_optimal:
# trainer.get_policy().model.save(model_save_paths_dict)
policy_save_tag += 1
trainer.get_policy().model.save_model_in_progress(
model_save_paths_dict, policy_save_tag)
saved = True
# Write the metrics for this reporting period.
total_seconds = time.time() - start_time
logger.write_and_condense_metrics(total_seconds, 'iters',
timesteps, saved, metrics,
tf_writer)
# Clear the metrics, both those maintained by the training workers and by the evaluation ones.
condense_fn = lambda environment: [
m.condense_values() for m in environment.metrics
]
trainer.workers.foreach_worker(
lambda w: w.foreach_env(condense_fn))
if hasattr(trainer, "evaluation_workers"):
trainer.evaluation_workers.foreach_worker(
lambda w: w.foreach_env(condense_fn))
else:
start_time = time.time()
env = trainer.workers.local_worker().env
metrics = env.metrics
worker = trainer.workers.local_worker()
steps = steps_since_report = 0
while True:
batch = worker.sample()
current_steps = len(batch["obs"])
steps += current_steps
steps_since_report += current_steps
if steps_since_report >= self.pr.reporting_interval:
total_seconds = time.time() - start_time
# Write the metrics for this reporting period.
logger.write_and_condense_metrics(total_seconds, 'iters',
steps, False, metrics,
tf_writer)
steps_since_report = 0
if steps >= self.pr.total_steps:
break
env.close()
# Get a summary metric for the entire stage, based on the environment's first metric.
summary_metric = logger.summarize_stage(metrics[0])
# Temporary workaround for https://github.com/ray-project/ray/issues/8205
ray.shutdown()
_register_all()
return summary_metric