Python tune.run_experiments函数代码示例

    def launch(self):
        """Actual entry point into the class instance where everything happens.
        Lots of delegating to classes that are in subclass or can be over-ridden.
        """
        self.register_env_creator()

        # All worker nodes will block at this step during training
        ray_cluster_config = self.ray_init_config()
        if not self.is_master_node:
            return

        # Start the driver on master node
        ray.init(**ray_cluster_config)
        experiment_config = self.get_experiment_config()
        experiment_config = self.customize_experiment_config(experiment_config)
        print("Running experiment with config %s" % json.dumps(experiment_config, indent=2))
        run_experiments(experiment_config)
        all_wokers_host_names = self.get_all_host_names()[1:]
        # If distributed job, send TERMINATION_SIGNAL to all workers.
        if len(all_wokers_host_names) > 0:
            self.sage_cluster_communicator.create_s3_signal(TERMINATION_SIGNAL)

        algo = experiment_config["training"]["run"]
        env_string = experiment_config["training"]["config"]["env"]
        config = experiment_config["training"]["config"]
        self.save_checkpoint_and_serving_model(algorithm=algo,
                                               env_string=env_string,
                                               config=config)

 def run():
     run_experiments(
         {
             "foo": {
                 "run": MyResettableClass,
                 "max_failures": 1,
                 "num_samples": 4,
                 "config": {
                     "fake_reset_not_supported": True
                 },
             }
         },
         reuse_actors=True,
         scheduler=FrequentPausesScheduler())

def test_cluster_rllib_restore(start_connected_cluster, tmpdir):
    cluster = start_connected_cluster
    dirpath = str(tmpdir)
    script = """
import time
import ray
from ray import tune

ray.init(redis_address="{redis_address}")

kwargs = dict(
    run="PG",
    env="CartPole-v1",
    stop=dict(training_iteration=10),
    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)
    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(100):
        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"):
                break
        time.sleep(0.3)

    if not TrialRunner.checkpoint_exists(metadata_checkpoint_dir):
        raise RuntimeError("Checkpoint file didn't appear.")

    ray.shutdown()
    cluster.shutdown()
    cluster = _start_new_cluster()
    cluster.wait_for_nodes()

    # Restore properly from checkpoint
    trials2 = tune.run_experiments(
        {
            "experiment": {
                "run": "PG",
                "checkpoint_freq": 1,
                "local_dir": dirpath
            }
        },
        resume=True)
    assert all(t.status == Trial.TERMINATED for t in trials2)
    cluster.shutdown()

def test_ls(start_ray, capsys, tmpdir):
    """This test captures output of list_trials."""
    experiment_name = "test_ls"
    experiment_path = os.path.join(str(tmpdir), experiment_name)
    num_samples = 2
    with capsys.disabled():
        tune.run_experiments({
            experiment_name: {
                "run": "__fake",
                "stop": {
                    "training_iteration": 1
                },
                "num_samples": num_samples,
                "local_dir": str(tmpdir)
            }
        })

    commands.list_trials(experiment_path, info_keys=("status", ))
    captured = capsys.readouterr().out.strip()
    lines = captured.split("\n")
    assert sum("TERMINATED" in line for line in lines) == num_samples

 def testTrialReuseEnabled(self):
     trials = run_experiments(
         {
             "foo": {
                 "run": MyResettableClass,
                 "num_samples": 4,
                 "config": {},
             }
         },
         reuse_actors=True,
         scheduler=FrequentPausesScheduler())
     self.assertEqual([t.last_result["num_resets"] for t in trials],
                      [1, 2, 3, 4])

def test_lsx(start_ray, capsys, tmpdir):
    """This test captures output of list_experiments."""
    project_path = str(tmpdir)
    num_experiments = 3
    for i in range(num_experiments):
        experiment_name = "test_lsx{}".format(i)
        with capsys.disabled():
            tune.run_experiments({
                experiment_name: {
                    "run": "__fake",
                    "stop": {
                        "training_iteration": 1
                    },
                    "num_samples": 1,
                    "local_dir": project_path
                }
            })

    commands.list_experiments(project_path, info_keys=("total_trials", ))
    captured = capsys.readouterr().out.strip()
    lines = captured.split("\n")
    assert sum("1" in line for line in lines) >= 3

def test_cluster_down_full(start_connected_cluster, tmpdir):
    """Tests that run_experiment restoring works on cluster shutdown."""
    cluster = start_connected_cluster
    dirpath = str(tmpdir)

    exp1_args = dict(
        run="__fake",
        stop=dict(training_iteration=3),
        local_dir=dirpath,
        checkpoint_freq=1)
    exp2_args = dict(run="__fake", stop=dict(training_iteration=3))
    exp3_args = dict(
        run="__fake",
        stop=dict(training_iteration=3),
        config=dict(mock_error=True))
    exp4_args = dict(
        run="__fake",
        stop=dict(training_iteration=3),
        config=dict(mock_error=True),
        checkpoint_freq=1)
    all_experiments = {
        "exp1": exp1_args,
        "exp2": exp2_args,
        "exp3": exp3_args,
        "exp4": exp4_args
    }

    tune.run_experiments(all_experiments, raise_on_failed_trial=False)

    ray.shutdown()
    cluster.shutdown()
    cluster = _start_new_cluster()

    trials = tune.run_experiments(
        all_experiments, resume=True, raise_on_failed_trial=False)
    assert len(trials) == 4
    assert all(t.status in [Trial.TERMINATED, Trial.ERROR] for t in trials)
    cluster.shutdown()

#!/usr/bin/env python

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import sys

import ray
from ray.tune import register_trainable, run_experiments


def f(config, reporter):
    reporter(timesteps_total=1)


if __name__ == "__main__":
    ray.init()
    register_trainable("my_class", f)
    run_experiments({
        "test": {
            "run": "my_class",
            "stop": {
                "training_iteration": 1
            }
        }
    })
    assert 'ray.rllib' not in sys.modules, "RLlib should not be imported"

    args, _ = parser.parse_known_args()

    mnist_spec = {
        'run': train,
        'num_samples': 10,
        'stop': {
            'mean_accuracy': 0.99,
            'timesteps_total': 600,
        },
        'config': {
            'activation': grid_search(['relu', 'elu', 'tanh']),
        },
    }

    if args.smoke_test:
        mnist_spec['stop']['training_iteration'] = 2
        mnist_spec['num_samples'] = 1

    ray.init()

    from ray.tune.schedulers import AsyncHyperBandScheduler
    run_experiments(
        {
            'tune_mnist_test': mnist_spec
        },
        scheduler=AsyncHyperBandScheduler(
            time_attr="timesteps_total",
            reward_attr="mean_accuracy",
            max_t=600,
        ))

    parser = argparse.ArgumentParser()
    parser.add_argument(
        '--smoke-test', action='store_true', help='Finish quickly for testing')
    args, _ = parser.parse_known_args()

    register_trainable("my_class", TrainMNIST)
    mnist_spec = {
        'run': 'my_class',
        'stop': {
          'mean_accuracy': 0.99,
          'time_total_s': 600,
        },
        'config': {
            'learning_rate': lambda spec:  10 ** np.random.uniform(-5, -3),
            'activation': grid_search(['relu', 'elu', 'tanh']),
        },
        "repeat": 10,
    }

    if args.smoke_test:
        mnist_spec['stop']['training_iteration'] = 2
        mnist_spec['repeat'] = 2

    ray.init()
    hyperband = HyperBandScheduler(
        time_attr="timesteps_total", reward_attr="mean_accuracy",
        max_t=100)

    run_experiments(
        {'mnist_hyperband_test': mnist_spec}, scheduler=hyperband)

    for i in range(args.num_iters):
        if i % 10 == 0:
            start = time.time()
            loss = sgd.step(fetch_stats=True)["loss"]
            metrics = sgd.foreach_model(lambda model: model.get_metrics())
            acc = [m["accuracy"] for m in metrics]
            print("Iter", i, "loss", loss, "accuracy", acc)
            print("Time per iteration", time.time() - start)
            assert len(set(acc)) == 1, ("Models out of sync", acc)
            reporter(timesteps_total=i, mean_loss=loss, mean_accuracy=acc[0])
        else:
            sgd.step()


if __name__ == "__main__":
    args = parser.parse_args()
    ray.init(redis_address=args.redis_address)

    if args.tune:
        run_experiments({
            "mnist_sgd": {
                "run": train_mnist,
                "config": {
                    "args": args,
                },
            },
        })
    else:
        train_mnist({"args": args}, lambda **kw: None)

    result = subprocess.check_output(
        "ps aux | grep '{}' | grep -v grep || true".format(UNIQUE_CMD),
        shell=True)
    return result


if __name__ == "__main__":
    register_env("subproc", lambda config: EnvWithSubprocess(config))
    ray.init()
    assert os.path.exists(UNIQUE_FILE_0)
    assert os.path.exists(UNIQUE_FILE_1)
    assert not leaked_processes()
    run_experiments({
        "demo": {
            "run": "PG",
            "env": "subproc",
            "num_samples": 1,
            "config": {
                "num_workers": 1,
            },
            "stop": {
                "training_iteration": 1
            },
        },
    })
    leaked = leaked_processes()
    assert not leaked, "LEAKED PROCESSES: {}".format(leaked)
    assert not os.path.exists(UNIQUE_FILE_0), "atexit handler not called"
    assert not os.path.exists(UNIQUE_FILE_1), "atexit handler not called"
    print("OK")

        },
        {
            'name': 'height',
            'type': 'int',
            'bounds': {
                'min': -100,
                'max': 100
            },
        },
    ]

    config = {
        "my_exp": {
            "run": "exp",
            "num_samples": 10 if args.smoke_test else 1000,
            "config": {
                "iterations": 100,
            },
            "stop": {
                "timesteps_total": 100
            },
        }
    }
    algo = SigOptSearch(
        space,
        name="SigOpt Example Experiment",
        max_concurrent=1,
        reward_attr="neg_mean_loss")
    scheduler = AsyncHyperBandScheduler(reward_attr="neg_mean_loss")
    run_experiments(config, search_alg=algo, scheduler=scheduler)

        num_cpus=10,
        num_gpus=0,
        resources={str(i): 2},
        object_store_memory=object_store_memory,
        redis_max_memory=redis_max_memory)
ray.init(redis_address=cluster.redis_address)

# Run the workload.

pbt = PopulationBasedTraining(
    time_attr="training_iteration",
    reward_attr="episode_reward_mean",
    perturbation_interval=10,
    hyperparam_mutations={
        "lr": [0.1, 0.01, 0.001, 0.0001],
    })

run_experiments(
    {
        "pbt_test": {
            "run": "PG",
            "env": "CartPole-v0",
            "num_samples": 8,
            "config": {
                "lr": 0.01,
            },
        }
    },
    scheduler=pbt,
    verbose=False)

run_experiments({
    "carla-dqn": {
        "run": "DQN",
        "env": "carla_env",
        "resources": {"cpu": 4, "gpu": 1},
        "config": {
            "env_config": env_config,
            "model": {
                "custom_model": "carla",
                "custom_options": {
                    "image_shape": [
                        80, 80,
                        lambda spec: spec.config.env_config.framestack * (
                            spec.config.env_config.use_depth_camera and 1 or 3
                        ),
                    ],
                },
                "conv_filters": [
                    [16, [8, 8], 4],
                    [32, [4, 4], 2],
                    [512, [10, 10], 1],
                ],
            },
            "timesteps_per_iteration": 100,
            "learning_starts": 1000,
            "schedule_max_timesteps": 100000,
            "gamma": 0.8,
            "tf_session_args": {
              "gpu_options": {"allow_growth": True},
            },
        },
    },
})