def __init__(self, time_attr="time_total_s", reward_attr="episode_reward_mean", perturbation_interval=60.0, hyperparam_mutations={}, resample_probability=0.25, custom_explore_fn=None, log_config=True): if not hyperparam_mutations and not custom_explore_fn: raise TuneError( "You must specify at least one of `hyperparam_mutations` or " "`custom_explore_fn` to use PBT.") FIFOScheduler.__init__(self) self._reward_attr = reward_attr self._time_attr = time_attr self._perturbation_interval = perturbation_interval self._hyperparam_mutations = hyperparam_mutations self._resample_probability = resample_probability self._trial_state = {} self._custom_explore_fn = custom_explore_fn self._log_config = log_config # Metrics self._num_checkpoints = 0 self._num_perturbations = 0
def __init__(self, time_attr="time_total_s", reward_attr=None, metric="episode_reward_mean", mode="max", perturbation_interval=60.0, hyperparam_mutations={}, quantile_fraction=0.25, resample_probability=0.25, custom_explore_fn=None, log_config=True, require_attrs=True): for value in hyperparam_mutations.values(): if not (isinstance(value, (list, dict)) or callable(value)): raise TypeError("`hyperparam_mutation` values must be either " "a List, Dict, or callable.") if not hyperparam_mutations and not custom_explore_fn: raise TuneError( "You must specify at least one of `hyperparam_mutations` or " "`custom_explore_fn` to use PBT.") if quantile_fraction > 0.5 or quantile_fraction < 0: raise TuneError( "You must set `quantile_fraction` to a value between 0 and" "0.5. Current value: '{}'".format(quantile_fraction)) assert mode in ["min", "max"], "`mode` must be 'min' or 'max'!" if reward_attr is not None: mode = "max" metric = reward_attr logger.warning( "`reward_attr` is deprecated and will be removed in a future " "version of Tune. " "Setting `metric={}` and `mode=max`.".format(reward_attr)) FIFOScheduler.__init__(self) self._metric = metric if mode == "max": self._metric_op = 1. elif mode == "min": self._metric_op = -1. self._time_attr = time_attr self._perturbation_interval = perturbation_interval self._hyperparam_mutations = hyperparam_mutations self._quantile_fraction = quantile_fraction self._resample_probability = resample_probability self._trial_state = {} self._custom_explore_fn = custom_explore_fn self._log_config = log_config self._require_attrs = require_attrs # Metrics self._num_checkpoints = 0 self._num_perturbations = 0
def __init__(self, search_alg, scheduler=None, launch_web_server=False, metadata_checkpoint_dir=None, server_port=TuneServer.DEFAULT_PORT, verbose=True, queue_trials=False, reuse_actors=False, trial_executor=None): """Initializes a new TrialRunner. Args: search_alg (SearchAlgorithm): SearchAlgorithm for generating Trial objects. scheduler (TrialScheduler): Defaults to FIFOScheduler. launch_web_server (bool): Flag for starting TuneServer metadata_checkpoint_dir (str): Path where global checkpoints are stored and restored from. server_port (int): Port number for launching TuneServer verbose (bool): Flag for verbosity. If False, trial results will not be output. 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). trial_executor (TrialExecutor): Defaults to RayTrialExecutor. """ self._search_alg = search_alg self._scheduler_alg = scheduler or FIFOScheduler() self.trial_executor = (trial_executor or RayTrialExecutor(queue_trials=queue_trials, reuse_actors=reuse_actors)) # For debugging, it may be useful to halt trials after some time has # elapsed. TODO(ekl) consider exposing this in the API. self._global_time_limit = float( os.environ.get("TRIALRUNNER_WALLTIME_LIMIT", float('inf'))) self._total_time = 0 self._iteration = 0 self._verbose = verbose self._queue_trials = queue_trials self._server = None self._server_port = server_port if launch_web_server: self._server = TuneServer(self, self._server_port) self._trials = [] self._stop_queue = [] self._metadata_checkpoint_dir = metadata_checkpoint_dir self._start_time = time.time() self._session_str = datetime.fromtimestamp( self._start_time).strftime("%Y-%m-%d_%H-%M-%S")
def __init__( self, time_attr="training_iteration", reward_attr=None, metric="episode_reward_mean", mode="max", max_t=100, grace_period=1, reduction_factor=4, brackets=1, ): assert max_t > 0, "Max (time_attr) not valid!" assert max_t >= grace_period, "grace_period must be <= max_t!" assert grace_period > 0, "grace_period must be positive!" assert reduction_factor > 1, "Reduction Factor not valid!" assert brackets > 0, "brackets must be positive!" assert mode in ["min", "max"], "`mode` must be 'min' or 'max'!" if reward_attr is not None: mode = "max" metric = reward_attr logger.warning( "`reward_attr` is deprecated and will be removed in a future " "version of Tune. " "Setting `metric={}` and `mode=max`.".format(reward_attr) ) FIFOScheduler.__init__(self) self._reduction_factor = reduction_factor self._max_t = max_t # Tracks state for new trial add self._brackets = [ _Bracket(grace_period, max_t, reduction_factor, s) for s in range(brackets) ] self._counter = 0 # for self._num_stopped = 0 self._metric = metric if mode == "max": self._metric_op = 1.0 elif mode == "min": self._metric_op = -1.0 self._time_attr = time_attr self._num_paused = 0
def __init__(self, time_attr="time_total_s", reward_attr=None, metric="episode_reward_mean", mode="max", perturbation_interval=60.0, hyperparam_mutations={}, resample_probability=0.25, custom_explore_fn=None, log_config=True): if not hyperparam_mutations and not custom_explore_fn: raise TuneError( "You must specify at least one of `hyperparam_mutations` or " "`custom_explore_fn` to use PBT.") assert mode in ["min", "max"], "`mode` must be 'min' or 'max'!" if reward_attr is not None: mode = "max" metric = reward_attr logger.warning( "`reward_attr` is deprecated and will be removed in a future " "version of Tune. " "Setting `metric={}` and `mode=max`.".format(reward_attr)) FIFOScheduler.__init__(self) self._metric = metric if mode == "max": self._metric_op = 1. elif mode == "min": self._metric_op = -1. self._time_attr = time_attr self._perturbation_interval = perturbation_interval self._hyperparam_mutations = hyperparam_mutations self._resample_probability = resample_probability self._trial_state = {} self._custom_explore_fn = custom_explore_fn self._log_config = log_config # Metrics self._num_checkpoints = 0 self._num_perturbations = 0
def choose_trial_to_run(self, trial_runner): for bracket in self._brackets: for trial in bracket.promotable_trials(): if trial and trial_runner.has_resources(trial.resources): assert trial.status == Trial.PAUSED logger.warning(f"Promoting trial [{trial.config}].") bracket.unpause_trial(trial) return trial trial = FIFOScheduler.choose_trial_to_run(self, trial_runner) if trial: self._brackets[0].unpause_trial(trial) logger.info(f"Choosing trial {trial.config} to run from trialrunner.") return trial
def __init__(self, time_attr="time_total_s", reward_attr="episode_reward_mean", perturbation_interval=60.0, hyperparam_mutations={}, resample_probability=0.25, custom_explore_fn=None): if not hyperparam_mutations and not custom_explore_fn: raise TuneError( "You must specify at least one of `hyperparam_mutations` or " "`custom_explore_fn` to use PBT.") FIFOScheduler.__init__(self) self._reward_attr = reward_attr self._time_attr = time_attr self._perturbation_interval = perturbation_interval self._hyperparam_mutations = hyperparam_mutations self._resample_probability = resample_probability self._trial_state = {} self._custom_explore_fn = custom_explore_fn # Metrics self._num_checkpoints = 0 self._num_perturbations = 0
def main(opt): data_dir = os.path.abspath("../data") device = torch.device("cuda" if torch.cuda.is_available() else "cpu") opt = set_search_space(opt) # scheduler = ASHAScheduler( # metric=opt['metric'], # mode="max", # max_t=opt["epoch"], # grace_period=opt["grace_period"], # reduction_factor=opt["reduction_factor"], # ) scheduler = FIFOScheduler() reporter = CLIReporter(metric_columns=[ "accuracy", "test_acc", "train_acc", "loss", "training_iteration", "forward_nfe", "backward_nfe" ]) # choose a search algorithm from https://docs.ray.io/en/latest/tune/api_docs/suggestion.html # search_alg = AxSearch(metric=opt['metric']) search_alg = None train_fn = train_ray if opt["num_splits"] == 0 else train_ray_rand result = tune.run( partial(train_fn, data_dir=data_dir), name=opt["name"], resources_per_trial={ "cpu": opt["cpus"], "gpu": opt["gpus"] }, search_alg=search_alg, keep_checkpoints_num=3, checkpoint_score_attr=opt['metric'], config=opt, num_samples=opt["num_samples"], scheduler=scheduler, max_failures=2, local_dir="../ray_tune", progress_reporter=reporter, raise_on_failed_trial=False, )
def __init__(self, search_alg, scheduler=None, launch_web_server=False, server_port=TuneServer.DEFAULT_PORT, verbose=True, queue_trials=False, trial_executor=None): """Initializes a new TrialRunner. Args: search_alg (SearchAlgorithm): SearchAlgorithm for generating Trial objects. scheduler (TrialScheduler): Defaults to FIFOScheduler. launch_web_server (bool): Flag for starting TuneServer server_port (int): Port number for launching TuneServer verbose (bool): Flag for verbosity. If False, trial results will not be output. 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): Defaults to RayTrialExecutor. """ self._search_alg = search_alg self._scheduler_alg = scheduler or FIFOScheduler() self._trials = [] self.trial_executor = trial_executor or \ RayTrialExecutor(queue_trials=queue_trials) # For debugging, it may be useful to halt trials after some time has # elapsed. TODO(ekl) consider exposing this in the API. self._global_time_limit = float( os.environ.get("TRIALRUNNER_WALLTIME_LIMIT", float('inf'))) self._total_time = 0 self._server = None if launch_web_server: self._server = TuneServer(self, server_port) self._stop_queue = [] self._verbose = verbose self._queue_trials = queue_trials
def __init__( self, total_atoms, resource_policy="UNIFORM", scaling_dict=SCALING_MAP["LINEAR"], deadline=np.inf, allocation_grid=None, use_pausing=True, grace_period=1, reduction_factor=4, max_t=100, time_attr="training_iteration", metric="episode_reward_mean", mode="max", _no_speculation=False, _ignore_overhead=False, _no_job_limit=False, _assume_linear=False, _fixed_exploration=False, _exploration_ratio=1.0, ): # Arguments for ablative study self._no_speculation = _no_speculation # stored self._ignore_overhead = _ignore_overhead # stored self._no_job_limit = _no_job_limit # stored self._assume_linear = _assume_linear self._fixed_exploration = _fixed_exploration self._exploration_ratio = _exploration_ratio FIFOScheduler.__init__(self) self.use_pausing = use_pausing self._num_paused = 0 self._num_stopped = 0 self._reduction_factor = reduction_factor self._max_t = max_t self._metric = metric self._time_attr = time_attr if mode == "max": self._metric_op = 1.0 elif mode == "min": self._metric_op = -1.0 if self._no_speculation: self._brackets = [ ASHAv2Bracket( min_t=grace_period, max_t=self._max_t, reduction_factor=self._reduction_factor, s=0, ) ] else: self._brackets = [ _DeadlineBracket( self._reduction_factor, max_t=self._max_t, min_t=grace_period, use_pausing=self.use_pausing, ) ] if self._fixed_exploration: logger.warning( f"FIXED EXPLORATION TIME OF {self._exploration_ratio}" ) if self._fixed_exploration: logger.warning( f"FIXED EXPLORATION TIME OF {self._exploration_ratio}" ) self.grace_period = grace_period self.start_time = time.time() self._deadline = deadline self._deadline_time = deadline + time.time() self._longest_duration = -1 check(self._deadline_time > self.start_time) self.total_atoms = total_atoms self.allocator = DynamicAllocator( self.total_atoms, policy=resource_policy, allocation_grid=allocation_grid, recharge_period=5, metric=self._metric, metric_op=self._metric_op, ) if self._assume_linear: logger.warning("ABLATION: ASSUMING LINEAR SCALING.") scaling_dict = SCALING_MAP["LINEAR"] self.scaling_fn = scaling_function_from_dict(scaling_dict) self._startup_times = set() #: Time it takes for a single iteration self._single_atom_iteration_times = []
def __init__(self, search_alg=None, scheduler=None, launch_web_server=False, local_checkpoint_dir=None, remote_checkpoint_dir=None, sync_to_cloud=None, stopper=None, resume=False, server_port=TuneServer.DEFAULT_PORT, fail_fast=False, verbose=True, checkpoint_period=10, trial_executor=None): self._search_alg = search_alg or BasicVariantGenerator() self._scheduler_alg = scheduler or FIFOScheduler() self.trial_executor = trial_executor or RayTrialExecutor() # For debugging, it may be useful to halt trials after some time has # elapsed. TODO(ekl) consider exposing this in the API. self._global_time_limit = float( os.environ.get("TRIALRUNNER_WALLTIME_LIMIT", float("inf"))) self._total_time = 0 self._iteration = 0 self._has_errored = False self._fail_fast = fail_fast self._verbose = verbose self._server = None self._server_port = server_port if launch_web_server: self._server = TuneServer(self, self._server_port) self._trials = [] self._cached_trial_decisions = {} self._stop_queue = [] self._should_stop_experiment = False # used by TuneServer self._local_checkpoint_dir = local_checkpoint_dir if self._local_checkpoint_dir: os.makedirs(self._local_checkpoint_dir, exist_ok=True) self._remote_checkpoint_dir = remote_checkpoint_dir self._syncer = get_cloud_syncer(local_checkpoint_dir, remote_checkpoint_dir, sync_to_cloud) self._stopper = stopper or NoopStopper() self._resumed = False if self._validate_resume(resume_type=resume): try: self.resume() logger.info("Resuming trial.") self._resumed = True except Exception: logger.exception( "Runner restore failed. Restarting experiment.") else: logger.debug("Starting a new experiment.") self._start_time = time.time() self._last_checkpoint_time = -float("inf") self._checkpoint_period = checkpoint_period self._session_str = datetime.fromtimestamp( self._start_time).strftime("%Y-%m-%d_%H-%M-%S") self.checkpoint_file = None if self._local_checkpoint_dir: self.checkpoint_file = os.path.join( self._local_checkpoint_dir, TrialRunner.CKPT_FILE_TMPL.format(self._session_str))
def run_experiments(experiments=None, search_alg=None, scheduler=None, with_server=False, server_port=TuneServer.DEFAULT_PORT, verbose=True, 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. 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. """ 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, launch_web_server=with_server, server_port=server_port, verbose=verbose, queue_trials=queue_trials, trial_executor=trial_executor) logger.info(runner.debug_string(max_debug=99999)) last_debug = 0 while not runner.is_finished(): runner.step() if time.time() - last_debug > DEBUG_PRINT_INTERVAL: logger.info(runner.debug_string()) last_debug = time.time() logger.info(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 __init__(self, 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=False, checkpoint_freq=0, checkpoint_at_end=False, sync_on_checkpoint=True, keep_checkpoints_num=None, checkpoint_score_attr=None, global_checkpoint_period=10, export_formats=None, max_failures=0, fail_fast=True, restore=None, search_alg=None, scheduler=None, with_server=False, server_port=TuneServer.DEFAULT_PORT, verbose=0, progress_reporter=None, resume=False, queue_trials=False, reuse_actors=False, trial_executor=None, raise_on_failed_trial=True, return_trials=False, ray_auto_init=True, shuffle=False): if loggers is None: loggers = [JsonLogger, CSVLogger] config = _transform_config(config) is_trainable = False try: if issubclass(run_or_experiment, Trainable): is_trainable = True except TypeError: pass if not is_trainable: run_or_experiment = wrap_function(run_or_experiment) self.trial_executor = trial_executor or RayTrialExecutor( queue_trials=queue_trials, reuse_actors=reuse_actors, ray_auto_init=ray_auto_init) experiments = [run_or_experiment] self.logger = logging.getLogger(__name__) for i, exp in enumerate(experiments): if not isinstance(exp, Experiment): run_identifier = Experiment.register_if_needed(exp) experiments[i] = 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, sync_on_checkpoint=sync_on_checkpoint, keep_checkpoints_num=keep_checkpoints_num, checkpoint_score_attr=checkpoint_score_attr, export_formats=export_formats, max_failures=max_failures, restore=restore) if fail_fast and max_failures != 0: raise ValueError("max_failures must be 0 if fail_fast=True.") self.runner = TrialRunner( search_alg=search_alg or BasicVariantGenerator(shuffle=shuffle), scheduler=scheduler or FIFOScheduler(), local_checkpoint_dir=experiments[0].checkpoint_dir, remote_checkpoint_dir=experiments[0].remote_checkpoint_dir, sync_to_cloud=sync_to_cloud, stopper=experiments[0].stopper, checkpoint_period=global_checkpoint_period, resume=resume, launch_web_server=with_server, server_port=server_port, verbose=bool(verbose > 1), fail_fast=fail_fast, trial_executor=self.trial_executor) for exp in experiments: self.runner.add_experiment(exp) self._is_worker_stopped = threading.Event() self._worker_exc = None self._worker = threading.Thread(target=self.step_worker, daemon=True) self._worker.start() atexit.register(self.stop)
def __init__(self, search_alg=None, scheduler=None, local_checkpoint_dir=None, remote_checkpoint_dir=None, sync_to_cloud=None, stopper=None, resume=False, server_port=None, fail_fast=False, verbose=True, checkpoint_period=None, trial_executor=None): self._search_alg = search_alg or BasicVariantGenerator() self._scheduler_alg = scheduler or FIFOScheduler() self.trial_executor = trial_executor or RayTrialExecutor() # For debugging, it may be useful to halt trials after some time has # elapsed. TODO(ekl) consider exposing this in the API. self._global_time_limit = float( os.environ.get("TRIALRUNNER_WALLTIME_LIMIT", float("inf"))) self._total_time = 0 self._iteration = 0 self._has_errored = False self._fail_fast = fail_fast if isinstance(self._fail_fast, str): self._fail_fast = self._fail_fast.upper() if self._fail_fast == TrialRunner.RAISE: logger.warning( "fail_fast='raise' detected. Be careful when using this " "mode as resources (such as Ray processes, " "file descriptors, and temporary files) may not be " "cleaned up properly. To use " "a safer mode, use fail_fast=True.") else: raise ValueError("fail_fast must be one of {bool, RAISE}. " f"Got {self._fail_fast}.") self._verbose = verbose self._server = None self._server_port = server_port if server_port is not None: self._server = TuneServer(self, self._server_port) self._trials = [] self._cached_trial_decisions = {} self._stop_queue = [] self._should_stop_experiment = False # used by TuneServer self._local_checkpoint_dir = local_checkpoint_dir if self._local_checkpoint_dir: os.makedirs(self._local_checkpoint_dir, exist_ok=True) self._remote_checkpoint_dir = remote_checkpoint_dir self._syncer = get_cloud_syncer(local_checkpoint_dir, remote_checkpoint_dir, sync_to_cloud) self._stopper = stopper or NoopStopper() self._resumed = False if self._validate_resume(resume_type=resume): errored_only = False if isinstance(resume, str): errored_only = resume.upper() == "ERRORED_ONLY" try: self.resume(run_errored_only=errored_only) self._resumed = True except Exception as e: if self._verbose: logger.error(str(e)) logger.exception("Runner restore failed.") if self._fail_fast: raise logger.info("Restarting experiment.") else: logger.debug("Starting a new experiment.") self._start_time = time.time() self._last_checkpoint_time = -float("inf") if checkpoint_period is None: checkpoint_period = env_integer("TUNE_GLOBAL_CHECKPOINT_S", 10) self._checkpoint_period = checkpoint_period self._session_str = datetime.fromtimestamp( self._start_time).strftime("%Y-%m-%d_%H-%M-%S") self.checkpoint_file = None if self._local_checkpoint_dir: self.checkpoint_file = os.path.join( self._local_checkpoint_dir, TrialRunner.CKPT_FILE_TMPL.format(self._session_str))
def hyper_parameter(task=None, model_name=None, dataset_name=None, config_file=None, space_file=None, scheduler=None, search_alg=None, other_args=None, num_samples=5, max_concurrent=1, cpu_per_trial=1, gpu_per_trial=1): """ Use Ray tune to hyper parameter tune Args: task(str): task name model_name(str): model name dataset_name(str): dataset name config_file(str): config filename used to modify the pipeline's settings. the config file should be json. space_file(str): the file which specifies the parameter search space scheduler(str): the trial sheduler which will be used in ray.tune.run search_alg(str): the search algorithm other_args(dict): the rest parameter args, which will be pass to the Config """ # load config experiment_config = ConfigParser(task, model_name, dataset_name, config_file=config_file, other_args=other_args) # logger logger = get_logger(experiment_config) logger.info(experiment_config.config) # check space_file if space_file is None: logger.error( 'the space_file should not be None when hyperparameter tune.') exit(0) # seed seed = experiment_config.get('seed', 0) set_random_seed(seed) # parse space_file search_sapce = parse_search_space(space_file) # load dataset dataset = get_dataset(experiment_config) # get train valid test data train_data, valid_data, test_data = dataset.get_data() data_feature = dataset.get_data_feature() def train(config, checkpoint_dir=None, experiment_config=None, train_data=None, valid_data=None, data_feature=None): """trainable function which meets ray tune API Args: config (dict): A dict of hyperparameter. """ # modify experiment_config for key in config: if key in experiment_config: experiment_config[key] = config[key] experiment_config['hyper_tune'] = True logger = get_logger(experiment_config) logger.info( 'Begin pipeline, task={}, model_name={}, dataset_name={}'.format( str(task), str(model_name), str(dataset_name))) logger.info('running parameters: ' + str(config)) # load model model = get_model(experiment_config, data_feature) # load executor executor = get_executor(experiment_config, model, data_feature) # checkpoint by ray tune if checkpoint_dir: checkpoint = os.path.join(checkpoint_dir, 'checkpoint') executor.load_model(checkpoint) # train executor.train(train_data, valid_data) # init search algorithm and scheduler if search_alg == 'BasicSearch': algorithm = BasicVariantGenerator() elif search_alg == 'BayesOptSearch': algorithm = BayesOptSearch(metric='loss', mode='min') # add concurrency limit algorithm = ConcurrencyLimiter(algorithm, max_concurrent=max_concurrent) elif search_alg == 'HyperOpt': algorithm = HyperOptSearch(metric='loss', mode='min') # add concurrency limit algorithm = ConcurrencyLimiter(algorithm, max_concurrent=max_concurrent) else: raise ValueError('the search_alg is illegal.') if scheduler == 'FIFO': tune_scheduler = FIFOScheduler() elif scheduler == 'ASHA': tune_scheduler = ASHAScheduler() elif scheduler == 'MedianStoppingRule': tune_scheduler = MedianStoppingRule() else: raise ValueError('the scheduler is illegal') # ray tune run ensure_dir('./libcity/cache/hyper_tune') result = tune.run(tune.with_parameters(train, experiment_config=experiment_config, train_data=train_data, valid_data=valid_data, data_feature=data_feature), resources_per_trial={ 'cpu': cpu_per_trial, 'gpu': gpu_per_trial }, config=search_sapce, metric='loss', mode='min', scheduler=tune_scheduler, search_alg=algorithm, local_dir='./libcity/cache/hyper_tune', num_samples=num_samples) best_trial = result.get_best_trial("loss", "min", "last") logger.info("Best trial config: {}".format(best_trial.config)) logger.info("Best trial final validation loss: {}".format( best_trial.last_result["loss"])) # save best best_path = os.path.join(best_trial.checkpoint.value, "checkpoint") model_state, optimizer_state = torch.load(best_path) model_cache_file = './libcity/cache/model_cache/{}_{}.m'.format( model_name, dataset_name) ensure_dir('./libcity/cache/model_cache') torch.save((model_state, optimizer_state), model_cache_file)
def run(run_or_experiment, name=None, stop=None, config=None, resources_per_trial=None, num_samples=1, local_dir=None, upload_dir=None, trial_name_creator=None, loggers=None, sync_to_cloud=None, sync_to_driver=None, checkpoint_freq=0, checkpoint_at_end=False, export_formats=None, max_failures=3, restore=None, search_alg=None, scheduler=None, with_server=False, server_port=TuneServer.DEFAULT_PORT, verbose=2, resume=False, queue_trials=False, reuse_actors=True, trial_executor=None, raise_on_failed_trial=True, return_trials=True, ray_auto_init=True, sync_function=None): """Executes training. Args: run_or_experiment (function|class|str|Experiment): If function|class|str, this is the algorithm or model to train. This may refer to the name of a built-on algorithm (e.g. RLLib's DQN or PPO), a user-defined trainable function or class, or the string identifier of a trainable function or class registered in the tune registry. If Experiment, then Tune will execute training based on Experiment.spec. name (str): Name of experiment. stop (dict): The stopping criteria. The keys may be any field in the return result of 'train()', whichever is reached first. Defaults to empty dict. config (dict): Algorithm-specific configuration for Tune variant generation (e.g. env, hyperparams). Defaults to empty dict. Custom search algorithms may ignore this. resources_per_trial (dict): Machine resources to allocate per trial, e.g. ``{"cpu": 64, "gpu": 8}``. Note that GPUs will not be assigned unless you specify them here. Defaults to 1 CPU and 0 GPUs in ``Trainable.default_resource_request()``. num_samples (int): Number of times to sample from the hyperparameter space. Defaults to 1. If `grid_search` is provided as an argument, the grid will be repeated `num_samples` of times. local_dir (str): Local dir to save training results to. Defaults to ``~/ray_results``. upload_dir (str): Optional URI to sync training results to (e.g. ``s3://bucket``). trial_name_creator (func): Optional function for generating the trial string representation. loggers (list): List of logger creators to be used with each Trial. If None, defaults to ray.tune.logger.DEFAULT_LOGGERS. See `ray/tune/logger.py`. sync_to_cloud (func|str): Function for syncing the local_dir to and from upload_dir. If string, then it must be a string template that includes `{source}` and `{target}` for the syncer to run. If not provided, the sync command defaults to standard S3 or gsutil sync comamnds. sync_to_driver (func|str): Function for syncing trial logdir from remote node to local. If string, then it must be a string template that includes `{source}` and `{target}` for the syncer to run. If not provided, defaults to using rsync. checkpoint_freq (int): How many training iterations between checkpoints. A value of 0 (default) disables checkpointing. checkpoint_at_end (bool): Whether to checkpoint at the end of the experiment regardless of the checkpoint_freq. Default is False. export_formats (list): List of formats that exported at the end of the experiment. Default is None. max_failures (int): Try to recover a trial from its last checkpoint at least this many times. Only applies if checkpointing is enabled. Setting to -1 will lead to infinite recovery retries. Defaults to 3. restore (str): Path to checkpoint. Only makes sense to set if running 1 trial. Defaults to None. search_alg (SearchAlgorithm): Search Algorithm. Defaults to BasicVariantGenerator. scheduler (TrialScheduler): Scheduler for executing the experiment. Choose among FIFO (default), MedianStopping, AsyncHyperBand, and HyperBand. with_server (bool): Starts a background Tune server. Needed for using the Client API. server_port (int): Port number for launching TuneServer. verbose (int): 0, 1, or 2. Verbosity mode. 0 = silent, 1 = only status updates, 2 = status and trial results. resume (str|bool): One of "LOCAL", "REMOTE", "PROMPT", or bool. LOCAL/True restores the checkpoint from the local_checkpoint_dir. REMOTE restores the checkpoint from remote_checkpoint_dir. PROMPT provides CLI feedback. False forces a new experiment. If resume is set but checkpoint does not exist, ValueError will be thrown. queue_trials (bool): Whether to queue trials when the cluster does not currently have enough resources to launch one. This should be set to True when running on an autoscaling cluster to enable automatic scale-up. reuse_actors (bool): Whether to reuse actors between different trials when possible. This can drastically speed up experiments that start and stop actors often (e.g., PBT in time-multiplexing mode). This requires trials to have the same resource requirements. trial_executor (TrialExecutor): Manage the execution of trials. raise_on_failed_trial (bool): Raise TuneError if there exists failed trial (of ERROR state) when the experiments complete. ray_auto_init (bool): Automatically starts a local Ray cluster if using a RayTrialExecutor (which is the default) and if Ray is not initialized. Defaults to True. sync_function: Deprecated. See `sync_to_cloud` and `sync_to_driver`. Returns: List of Trial objects. Raises: TuneError if any trials failed and `raise_on_failed_trial` is True. Examples: >>> tune.run(mytrainable, scheduler=PopulationBasedTraining()) >>> tune.run(mytrainable, num_samples=5, reuse_actors=True) >>> tune.run( "PG", num_samples=5, config={ "env": "CartPole-v0", "lr": tune.sample_from(lambda _: np.random.rand()) } ) """ trial_executor = trial_executor or RayTrialExecutor( queue_trials=queue_trials, reuse_actors=reuse_actors, ray_auto_init=ray_auto_init) experiment = run_or_experiment if not isinstance(run_or_experiment, Experiment): run_identifier = Experiment._register_if_needed(run_or_experiment) experiment = Experiment( name=name, run=run_identifier, stop=stop, config=config, resources_per_trial=resources_per_trial, num_samples=num_samples, local_dir=local_dir, upload_dir=upload_dir, sync_to_driver=sync_to_driver, trial_name_creator=trial_name_creator, loggers=loggers, checkpoint_freq=checkpoint_freq, checkpoint_at_end=checkpoint_at_end, export_formats=export_formats, max_failures=max_failures, restore=restore, sync_function=sync_function) else: logger.debug("Ignoring some parameters passed into tune.run.") if sync_to_cloud: assert experiment.remote_checkpoint_dir, ( "Need `upload_dir` if `sync_to_cloud` given.") runner = TrialRunner( search_alg=search_alg or BasicVariantGenerator(), scheduler=scheduler or FIFOScheduler(), local_checkpoint_dir=experiment.checkpoint_dir, remote_checkpoint_dir=experiment.remote_checkpoint_dir, sync_to_cloud=sync_to_cloud, resume=resume, launch_web_server=with_server, server_port=server_port, verbose=bool(verbose > 1), trial_executor=trial_executor) runner.add_experiment(experiment) if verbose: print(runner.debug_string(max_debug=99999)) last_debug = 0 while not runner.is_finished(): runner.step() if time.time() - last_debug > DEBUG_PRINT_INTERVAL: if verbose: print(runner.debug_string()) last_debug = time.time() if verbose: print(runner.debug_string(max_debug=99999)) wait_for_sync() errored_trials = [] for trial in runner.get_trials(): if trial.status != Trial.TERMINATED: errored_trials += [trial] if errored_trials: if raise_on_failed_trial: raise TuneError("Trials did not complete", errored_trials) else: logger.error("Trials did not complete: %s", errored_trials) if return_trials: return runner.get_trials() return ExperimentAnalysis(experiment.checkpoint_dir)
def run( run_or_experiment, name=None, metric=None, mode=None, stop=None, time_budget_s=None, config=None, resources_per_trial=None, num_samples=1, local_dir=None, search_alg=None, scheduler=None, keep_checkpoints_num=None, checkpoint_score_attr=None, checkpoint_freq=0, checkpoint_at_end=False, verbose=2, progress_reporter=None, loggers=None, log_to_file=False, trial_name_creator=None, trial_dirname_creator=None, sync_config=None, export_formats=None, max_failures=0, fail_fast=False, restore=None, server_port=None, resume=False, reuse_actors=False, trial_executor=None, raise_on_failed_trial=True, # Deprecated args ray_auto_init=None, run_errored_only=None, queue_trials=None, global_checkpoint_period=None, with_server=None, upload_dir=None, sync_to_cloud=None, sync_to_driver=None, sync_on_checkpoint=None, ): """Executes training. Examples: .. code-block:: python # Run 10 trials (each trial is one instance of a Trainable). Tune runs # in parallel and automatically determines concurrency. tune.run(trainable, num_samples=10) # Run 1 trial, stop when trial has reached 10 iterations tune.run(my_trainable, stop={"training_iteration": 10}) # automatically retry failed trials up to 3 times tune.run(my_trainable, stop={"training_iteration": 10}, max_failures=3) # Run 1 trial, search over hyperparameters, stop after 10 iterations. space = {"lr": tune.uniform(0, 1), "momentum": tune.uniform(0, 1)} tune.run(my_trainable, config=space, stop={"training_iteration": 10}) # Resumes training if a previous machine crashed tune.run(my_trainable, config=space, local_dir=<path/to/dir>, resume=True) # Rerun ONLY failed trials after an experiment is finished. tune.run(my_trainable, config=space, local_dir=<path/to/dir>, resume="ERRORED_ONLY") Args: run_or_experiment (function | class | str | :class:`Experiment`): If function|class|str, this is the algorithm or model to train. This may refer to the name of a built-on algorithm (e.g. RLLib's DQN or PPO), a user-defined trainable function or class, or the string identifier of a trainable function or class registered in the tune registry. If Experiment, then Tune will execute training based on Experiment.spec. If you want to pass in a Python lambda, you will need to first register the function: ``tune.register_trainable("lambda_id", lambda x: ...)``. You can then use ``tune.run("lambda_id")``. metric (str): Metric to optimize. This metric should be reported with `tune.report()`. If set, will be passed to the search algorithm and scheduler. mode (str): Must be one of [min, max]. Determines whether objective is minimizing or maximizing the metric attribute. If set, will be passed to the search algorithm and scheduler. name (str): Name of experiment. stop (dict | callable | :class:`Stopper`): Stopping criteria. If dict, the keys may be any field in the return result of 'train()', whichever is reached first. If function, it must take (trial_id, result) as arguments and return a boolean (True if trial should be stopped, False otherwise). This can also be a subclass of ``ray.tune.Stopper``, which allows users to implement custom experiment-wide stopping (i.e., stopping an entire Tune run based on some time constraint). time_budget_s (int|float|datetime.timedelta): Global time budget in seconds after which all trials are stopped. Can also be a ``datetime.timedelta`` object. config (dict): Algorithm-specific configuration for Tune variant generation (e.g. env, hyperparams). Defaults to empty dict. Custom search algorithms may ignore this. resources_per_trial (dict): Machine resources to allocate per trial, e.g. ``{"cpu": 64, "gpu": 8}``. Note that GPUs will not be assigned unless you specify them here. Defaults to 1 CPU and 0 GPUs in ``Trainable.default_resource_request()``. num_samples (int): Number of times to sample from the hyperparameter space. Defaults to 1. If `grid_search` is provided as an argument, the grid will be repeated `num_samples` of times. local_dir (str): Local dir to save training results to. Defaults to ``~/ray_results``. search_alg (Searcher): Search algorithm for optimization. scheduler (TrialScheduler): Scheduler for executing the experiment. Choose among FIFO (default), MedianStopping, AsyncHyperBand, HyperBand and PopulationBasedTraining. Refer to ray.tune.schedulers for more options. keep_checkpoints_num (int): Number of checkpoints to keep. A value of `None` keeps all checkpoints. Defaults to `None`. If set, need to provide `checkpoint_score_attr`. checkpoint_score_attr (str): Specifies by which attribute to rank the best checkpoint. Default is increasing order. If attribute starts with `min-` it will rank attribute in decreasing order, i.e. `min-validation_loss`. checkpoint_freq (int): How many training iterations between checkpoints. A value of 0 (default) disables checkpointing. This has no effect when using the Functional Training API. checkpoint_at_end (bool): Whether to checkpoint at the end of the experiment regardless of the checkpoint_freq. Default is False. This has no effect when using the Functional Training API. verbose (int): 0, 1, or 2. Verbosity mode. 0 = silent, 1 = only status updates, 2 = status and trial results. progress_reporter (ProgressReporter): Progress reporter for reporting intermediate experiment progress. Defaults to CLIReporter if running in command-line, or JupyterNotebookReporter if running in a Jupyter notebook. loggers (list): List of logger creators to be used with each Trial. If None, defaults to ray.tune.logger.DEFAULT_LOGGERS. See `ray/tune/logger.py`. log_to_file (bool|str|Sequence): Log stdout and stderr to files in Tune's trial directories. If this is `False` (default), no files are written. If `true`, outputs are written to `trialdir/stdout` and `trialdir/stderr`, respectively. If this is a single string, this is interpreted as a file relative to the trialdir, to which both streams are written. If this is a Sequence (e.g. a Tuple), it has to have length 2 and the elements indicate the files to which stdout and stderr are written, respectively. trial_name_creator (Callable[[Trial], str]): Optional function for generating the trial string representation. trial_dirname_creator (Callable[[Trial], str]): Function for generating the trial dirname. This function should take in a Trial object and return a string representing the name of the directory. The return value cannot be a path. sync_config (SyncConfig): Configuration object for syncing. See tune.SyncConfig. export_formats (list): List of formats that exported at the end of the experiment. Default is None. max_failures (int): Try to recover a trial at least this many times. Ray will recover from the latest checkpoint if present. Setting to -1 will lead to infinite recovery retries. Setting to 0 will disable retries. Defaults to 3. fail_fast (bool | str): Whether to fail upon the first error. If fail_fast='raise' provided, Tune will automatically raise the exception received by the Trainable. fail_fast='raise' can easily leak resources and should be used with caution (it is best used with `ray.init(local_mode=True)`). restore (str): Path to checkpoint. Only makes sense to set if running 1 trial. Defaults to None. server_port (int): Port number for launching TuneServer. resume (str|bool): One of "LOCAL", "REMOTE", "PROMPT", "ERRORED_ONLY", or bool. LOCAL/True restores the checkpoint from the local_checkpoint_dir, determined by `name` and `local_dir`. REMOTE restores the checkpoint from remote_checkpoint_dir. PROMPT provides CLI feedback. False forces a new experiment. ERRORED_ONLY resets and reruns ERRORED trials upon resume - previous trial artifacts will be left untouched. If resume is set but checkpoint does not exist, ValueError will be thrown. reuse_actors (bool): Whether to reuse actors between different trials when possible. This can drastically speed up experiments that start and stop actors often (e.g., PBT in time-multiplexing mode). This requires trials to have the same resource requirements. trial_executor (TrialExecutor): Manage the execution of trials. raise_on_failed_trial (bool): Raise TuneError if there exists failed trial (of ERROR state) when the experiments complete. Returns: ExperimentAnalysis: Object for experiment analysis. Raises: TuneError: Any trials failed and `raise_on_failed_trial` is True. """ if global_checkpoint_period: raise ValueError("global_checkpoint_period is deprecated. Set env var " "'TUNE_GLOBAL_CHECKPOINT_S' instead.") if queue_trials: raise ValueError( "queue_trials is deprecated. " "Set env var 'TUNE_DISABLE_QUEUE_TRIALS=1' instead to " "disable queuing behavior.") if ray_auto_init: raise ValueError("ray_auto_init is deprecated. " "Set env var 'TUNE_DISABLE_AUTO_INIT=1' instead or " "call 'ray.init' before calling 'tune.run'.") if with_server: raise ValueError( "with_server is deprecated. It is now enabled by default " "if 'server_port' is not None.") if sync_on_checkpoint or sync_to_cloud or sync_to_driver or upload_dir: raise ValueError( "sync_on_checkpoint / sync_to_cloud / sync_to_driver / " "upload_dir must now be set via `tune.run(" "sync_config=SyncConfig(...)`. See `ray.tune.SyncConfig` for " "more details.") if mode and mode not in ["min", "max"]: raise ValueError( "The `mode` parameter passed to `tune.run()` has to be one of " "['min', 'max']") config = config or {} sync_config = sync_config or SyncConfig() set_sync_periods(sync_config) trial_executor = trial_executor or RayTrialExecutor( reuse_actors=reuse_actors) if isinstance(run_or_experiment, list): experiments = run_or_experiment else: experiments = [run_or_experiment] for i, exp in enumerate(experiments): if not isinstance(exp, Experiment): experiments[i] = Experiment( name=name, run=exp, stop=stop, time_budget_s=time_budget_s, config=config, resources_per_trial=resources_per_trial, num_samples=num_samples, local_dir=local_dir, upload_dir=sync_config.upload_dir, sync_to_driver=sync_config.sync_to_driver, trial_name_creator=trial_name_creator, trial_dirname_creator=trial_dirname_creator, loggers=loggers, log_to_file=log_to_file, checkpoint_freq=checkpoint_freq, checkpoint_at_end=checkpoint_at_end, sync_on_checkpoint=sync_config.sync_on_checkpoint, keep_checkpoints_num=keep_checkpoints_num, checkpoint_score_attr=checkpoint_score_attr, export_formats=export_formats, max_failures=max_failures, restore=restore) else: logger.debug("Ignoring some parameters passed into tune.run.") if sync_config.sync_to_cloud: for exp in experiments: assert exp.remote_checkpoint_dir, ( "Need `upload_dir` if `sync_to_cloud` given.") if fail_fast and max_failures != 0: raise ValueError("max_failures must be 0 if fail_fast=True.") if issubclass(type(search_alg), Searcher): search_alg = SearchGenerator(search_alg) if not search_alg: search_alg = BasicVariantGenerator() if config and not search_alg.set_search_properties(metric, mode, config): if has_unresolved_values(config): raise ValueError( "You passed a `config` parameter to `tune.run()` with " "unresolved parameters, but the search algorithm was already " "instantiated with a search space. Make sure that `config` " "does not contain any more parameter definitions - include " "them in the search algorithm's search space if necessary.") scheduler = scheduler or FIFOScheduler() if not scheduler.set_search_properties(metric, mode): raise ValueError( "You passed a `metric` or `mode` argument to `tune.run()`, but " "the scheduler you are using was already instantiated with their " "own `metric` and `mode` parameters. Either remove the arguments " "from your scheduler or from your call to `tune.run()`") runner = TrialRunner( search_alg=search_alg, scheduler=scheduler, local_checkpoint_dir=experiments[0].checkpoint_dir, remote_checkpoint_dir=experiments[0].remote_checkpoint_dir, sync_to_cloud=sync_config.sync_to_cloud, stopper=experiments[0].stopper, resume=resume, server_port=server_port, verbose=bool(verbose > 1), fail_fast=fail_fast, trial_executor=trial_executor) if not runner.resumed: for exp in experiments: search_alg.add_configurations([exp]) else: logger.info("TrialRunner resumed, ignoring new add_experiment.") if progress_reporter is None: if IS_NOTEBOOK: progress_reporter = JupyterNotebookReporter(overwrite=verbose < 2) else: progress_reporter = CLIReporter() # User Warning for GPUs if trial_executor.has_gpus(): if isinstance(resources_per_trial, dict) and "gpu" in resources_per_trial: # "gpu" is manually set. pass elif _check_default_resources_override(experiments[0].run_identifier): # "default_resources" is manually overriden. pass else: logger.warning("Tune detects GPUs, but no trials are using GPUs. " "To enable trials to use GPUs, set " "tune.run(resources_per_trial={'gpu': 1}...) " "which allows Tune to expose 1 GPU to each trial. " "You can also override " "`Trainable.default_resource_request` if using the " "Trainable API.") while not runner.is_finished(): runner.step() if verbose: _report_progress(runner, progress_reporter) try: runner.checkpoint(force=True) except Exception as e: logger.warning(f"Trial Runner checkpointing failed: {str(e)}") if verbose: _report_progress(runner, progress_reporter, done=True) wait_for_sync() runner.cleanup_trials() incomplete_trials = [] for trial in runner.get_trials(): if trial.status != Trial.TERMINATED: incomplete_trials += [trial] if incomplete_trials: if raise_on_failed_trial: raise TuneError("Trials did not complete", incomplete_trials) else: logger.error("Trials did not complete: %s", incomplete_trials) trials = runner.get_trials() return ExperimentAnalysis(runner.checkpoint_file, trials=trials, default_metric=metric, default_mode=mode)
def run(run_or_experiment, name=None, stop=None, config=None, resources_per_trial=None, num_samples=1, local_dir=None, upload_dir=None, trial_name_creator=None, loggers=None, log_to_file=False, sync_to_cloud=None, sync_to_driver=None, checkpoint_freq=0, checkpoint_at_end=False, sync_on_checkpoint=True, keep_checkpoints_num=None, checkpoint_score_attr=None, global_checkpoint_period=10, export_formats=None, max_failures=0, fail_fast=False, restore=None, search_alg=None, scheduler=None, with_server=False, server_port=TuneServer.DEFAULT_PORT, verbose=2, progress_reporter=None, resume=False, queue_trials=False, reuse_actors=False, trial_executor=None, raise_on_failed_trial=True, return_trials=False, ray_auto_init=True): """Executes training. Args: run_or_experiment (function | class | str | :class:`Experiment`): If function|class|str, this is the algorithm or model to train. This may refer to the name of a built-on algorithm (e.g. RLLib's DQN or PPO), a user-defined trainable function or class, or the string identifier of a trainable function or class registered in the tune registry. If Experiment, then Tune will execute training based on Experiment.spec. If you want to pass in a Python lambda, you will need to first register the function: ``tune.register_trainable("lambda_id", lambda x: ...)``. You can then use ``tune.run("lambda_id")``. name (str): Name of experiment. stop (dict | callable | :class:`Stopper`): Stopping criteria. If dict, the keys may be any field in the return result of 'train()', whichever is reached first. If function, it must take (trial_id, result) as arguments and return a boolean (True if trial should be stopped, False otherwise). This can also be a subclass of ``ray.tune.Stopper``, which allows users to implement custom experiment-wide stopping (i.e., stopping an entire Tune run based on some time constraint). config (dict): Algorithm-specific configuration for Tune variant generation (e.g. env, hyperparams). Defaults to empty dict. Custom search algorithms may ignore this. resources_per_trial (dict): Machine resources to allocate per trial, e.g. ``{"cpu": 64, "gpu": 8}``. Note that GPUs will not be assigned unless you specify them here. Defaults to 1 CPU and 0 GPUs in ``Trainable.default_resource_request()``. num_samples (int): Number of times to sample from the hyperparameter space. Defaults to 1. If `grid_search` is provided as an argument, the grid will be repeated `num_samples` of times. local_dir (str): Local dir to save training results to. Defaults to ``~/ray_results``. upload_dir (str): Optional URI to sync training results and checkpoints to (e.g. ``s3://bucket`` or ``gs://bucket``). trial_name_creator (func): Optional function for generating the trial string representation. loggers (list): List of logger creators to be used with each Trial. If None, defaults to ray.tune.logger.DEFAULT_LOGGERS. See `ray/tune/logger.py`. log_to_file (bool|str|Sequence): Log stdout and stderr to files in Tune's trial directories. If this is `False` (default), no files are written. If `true`, outputs are written to `trialdir/stdout` and `trialdir/stderr`, respectively. If this is a single string, this is interpreted as a file relative to the trialdir, to which both streams are written. If this is a Sequence (e.g. a Tuple), it has to have length 2 and the elements indicate the files to which stdout and stderr are written, respectively. sync_to_cloud (func|str): Function for syncing the local_dir to and from upload_dir. If string, then it must be a string template that includes `{source}` and `{target}` for the syncer to run. If not provided, the sync command defaults to standard S3 or gsutil sync commands. By default local_dir is synced to remote_dir every 300 seconds. To change this, set the TUNE_CLOUD_SYNC_S environment variable in the driver machine. sync_to_driver (func|str|bool): Function for syncing trial logdir from remote node to local. If string, then it must be a string template that includes `{source}` and `{target}` for the syncer to run. If True or not provided, it defaults to using rsync. If False, syncing to driver is disabled. checkpoint_freq (int): How many training iterations between checkpoints. A value of 0 (default) disables checkpointing. This has no effect when using the Functional Training API. checkpoint_at_end (bool): Whether to checkpoint at the end of the experiment regardless of the checkpoint_freq. Default is False. This has no effect when using the Functional Training API. sync_on_checkpoint (bool): Force sync-down of trial checkpoint to driver. If set to False, checkpoint syncing from worker to driver is asynchronous and best-effort. This does not affect persistent storage syncing. Defaults to True. keep_checkpoints_num (int): Number of checkpoints to keep. A value of `None` keeps all checkpoints. Defaults to `None`. If set, need to provide `checkpoint_score_attr`. checkpoint_score_attr (str): Specifies by which attribute to rank the best checkpoint. Default is increasing order. If attribute starts with `min-` it will rank attribute in decreasing order, i.e. `min-validation_loss`. global_checkpoint_period (int): Seconds between global checkpointing. This does not affect `checkpoint_freq`, which specifies frequency for individual trials. export_formats (list): List of formats that exported at the end of the experiment. Default is None. max_failures (int): Try to recover a trial at least this many times. Ray will recover from the latest checkpoint if present. Setting to -1 will lead to infinite recovery retries. Setting to 0 will disable retries. Defaults to 3. fail_fast (bool): Whether to fail upon the first error. restore (str): Path to checkpoint. Only makes sense to set if running 1 trial. Defaults to None. search_alg (Searcher): Search algorithm for optimization. scheduler (TrialScheduler): Scheduler for executing the experiment. Choose among FIFO (default), MedianStopping, AsyncHyperBand, HyperBand and PopulationBasedTraining. Refer to ray.tune.schedulers for more options. with_server (bool): Starts a background Tune server. Needed for using the Client API. server_port (int): Port number for launching TuneServer. verbose (int): 0, 1, or 2. Verbosity mode. 0 = silent, 1 = only status updates, 2 = status and trial results. progress_reporter (ProgressReporter): Progress reporter for reporting intermediate experiment progress. Defaults to CLIReporter if running in command-line, or JupyterNotebookReporter if running in a Jupyter notebook. resume (str|bool): One of "LOCAL", "REMOTE", "PROMPT", or bool. LOCAL/True restores the checkpoint from the local_checkpoint_dir. REMOTE restores the checkpoint from remote_checkpoint_dir. PROMPT provides CLI feedback. False forces a new experiment. If resume is set but checkpoint does not exist, ValueError will be thrown. queue_trials (bool): Whether to queue trials when the cluster does not currently have enough resources to launch one. This should be set to True when running on an autoscaling cluster to enable automatic scale-up. reuse_actors (bool): Whether to reuse actors between different trials when possible. This can drastically speed up experiments that start and stop actors often (e.g., PBT in time-multiplexing mode). This requires trials to have the same resource requirements. trial_executor (TrialExecutor): Manage the execution of trials. raise_on_failed_trial (bool): Raise TuneError if there exists failed trial (of ERROR state) when the experiments complete. ray_auto_init (bool): Automatically starts a local Ray cluster if using a RayTrialExecutor (which is the default) and if Ray is not initialized. Defaults to True. Returns: ExperimentAnalysis: Object for experiment analysis. Raises: TuneError: Any trials failed and `raise_on_failed_trial` is True. Examples: .. code-block:: python # Run 10 trials (each trial is one instance of a Trainable). Tune runs # in parallel and automatically determines concurrency. tune.run(trainable, num_samples=10) # Run 1 trial, stop when trial has reached 10 iterations tune.run(my_trainable, stop={"training_iteration": 10}) # Run 1 trial, search over hyperparameters, stop after 10 iterations. space = {"lr": tune.uniform(0, 1), "momentum": tune.uniform(0, 1)} tune.run(my_trainable, config=space, stop={"training_iteration": 10}) """ config = config or {} trial_executor = trial_executor or RayTrialExecutor( queue_trials=queue_trials, reuse_actors=reuse_actors, ray_auto_init=ray_auto_init) if isinstance(run_or_experiment, list): experiments = run_or_experiment else: experiments = [run_or_experiment] for i, exp in enumerate(experiments): if not isinstance(exp, Experiment): experiments[i] = Experiment( name=name, run=exp, stop=stop, config=config, resources_per_trial=resources_per_trial, num_samples=num_samples, local_dir=local_dir, upload_dir=upload_dir, sync_to_driver=sync_to_driver, trial_name_creator=trial_name_creator, loggers=loggers, log_to_file=log_to_file, checkpoint_freq=checkpoint_freq, checkpoint_at_end=checkpoint_at_end, sync_on_checkpoint=sync_on_checkpoint, keep_checkpoints_num=keep_checkpoints_num, checkpoint_score_attr=checkpoint_score_attr, export_formats=export_formats, max_failures=max_failures, restore=restore) else: logger.debug("Ignoring some parameters passed into tune.run.") if sync_to_cloud: for exp in experiments: assert exp.remote_checkpoint_dir, ( "Need `upload_dir` if `sync_to_cloud` given.") if fail_fast and max_failures != 0: raise ValueError("max_failures must be 0 if fail_fast=True.") if issubclass(type(search_alg), Searcher): search_alg = SearchGenerator(search_alg) if not search_alg: search_alg = BasicVariantGenerator() runner = TrialRunner( search_alg=search_alg, scheduler=scheduler or FIFOScheduler(), local_checkpoint_dir=experiments[0].checkpoint_dir, remote_checkpoint_dir=experiments[0].remote_checkpoint_dir, sync_to_cloud=sync_to_cloud, stopper=experiments[0].stopper, checkpoint_period=global_checkpoint_period, resume=resume, launch_web_server=with_server, server_port=server_port, verbose=bool(verbose > 1), fail_fast=fail_fast, trial_executor=trial_executor) if not runner.resumed: for exp in experiments: search_alg.add_configurations([exp]) else: logger.info("TrialRunner resumed, ignoring new add_experiment.") if progress_reporter is None: if IS_NOTEBOOK: progress_reporter = JupyterNotebookReporter(overwrite=verbose < 2) else: progress_reporter = CLIReporter() # User Warning for GPUs if trial_executor.has_gpus(): if isinstance(resources_per_trial, dict) and "gpu" in resources_per_trial: # "gpu" is manually set. pass elif _check_default_resources_override(experiments[0].run_identifier): # "default_resources" is manually overriden. pass else: logger.warning("Tune detects GPUs, but no trials are using GPUs. " "To enable trials to use GPUs, set " "tune.run(resources_per_trial={'gpu': 1}...) " "which allows Tune to expose 1 GPU to each trial. " "You can also override " "`Trainable.default_resource_request` if using the " "Trainable API.") while not runner.is_finished(): runner.step() if verbose: _report_progress(runner, progress_reporter) try: runner.checkpoint(force=True) except Exception: logger.exception("Trial Runner checkpointing failed.") if verbose: _report_progress(runner, progress_reporter, done=True) wait_for_sync() runner.cleanup_trials() incomplete_trials = [] for trial in runner.get_trials(): if trial.status != Trial.TERMINATED: incomplete_trials += [trial] if incomplete_trials: if raise_on_failed_trial: raise TuneError("Trials did not complete", incomplete_trials) else: logger.error("Trials did not complete: %s", incomplete_trials) trials = runner.get_trials() if return_trials: return trials return ExperimentAnalysis(runner.checkpoint_file, trials=trials)
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, sync_on_checkpoint=True, keep_checkpoints_num=None, checkpoint_score_attr=None, global_checkpoint_period=10, export_formats=None, max_failures=0, restore=None, search_alg=None, scheduler=None, with_server=False, server_port=TuneServer.DEFAULT_PORT, verbose=2, progress_reporter=None, resume=False, queue_trials=False, reuse_actors=False, trial_executor=None, raise_on_failed_trial=True, return_trials=False, ray_auto_init=True, 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|func): The 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). config (dict): Algorithm-specific configuration for Tune variant generation (e.g. env, hyperparams). Defaults to empty dict. Custom search algorithms may ignore this. resources_per_trial (dict): Machine resources to allocate per trial, e.g. ``{"cpu": 64, "gpu": 8}``. Note that GPUs will not be assigned unless you specify them here. Defaults to 1 CPU and 0 GPUs in ``Trainable.default_resource_request()``. num_samples (int): Number of times to sample from the hyperparameter space. Defaults to 1. If `grid_search` is provided as an argument, the grid will be repeated `num_samples` of times. local_dir (str): Local dir to save training results to. Defaults to ``~/ray_results``. upload_dir (str): Optional URI to sync training results and checkpoints to (e.g. ``s3://bucket`` or ``gs://bucket``). trial_name_creator (func): Optional function for generating the trial string representation. loggers (list): List of logger creators to be used with each Trial. If None, defaults to ray.tune.logger.DEFAULT_LOGGERS. See `ray/tune/logger.py`. sync_to_cloud (func|str): Function for syncing the local_dir to and from upload_dir. If string, then it must be a string template that includes `{source}` and `{target}` for the syncer to run. If not provided, the sync command defaults to standard S3 or gsutil sync commands. sync_to_driver (func|str|bool): Function for syncing trial logdir from remote node to local. If string, then it must be a string template that includes `{source}` and `{target}` for the syncer to run. If True or not provided, it defaults to using rsync. If False, syncing to driver is disabled. checkpoint_freq (int): How many training iterations between checkpoints. A value of 0 (default) disables checkpointing. checkpoint_at_end (bool): Whether to checkpoint at the end of the experiment regardless of the checkpoint_freq. Default is False. sync_on_checkpoint (bool): Force sync-down of trial checkpoint to driver. If set to False, checkpoint syncing from worker to driver is asynchronous and best-effort. This does not affect persistent storage syncing. Defaults to True. keep_checkpoints_num (int): Number of checkpoints to keep. A value of `None` keeps all checkpoints. Defaults to `None`. If set, need to provide `checkpoint_score_attr`. checkpoint_score_attr (str): Specifies by which attribute to rank the best checkpoint. Default is increasing order. If attribute starts with `min-` it will rank attribute in decreasing order, i.e. `min-validation_loss`. global_checkpoint_period (int): Seconds between global checkpointing. This does not affect `checkpoint_freq`, which specifies frequency for individual trials. export_formats (list): List of formats that exported at the end of the experiment. Default is None. max_failures (int): Try to recover a trial at least this many times. Ray will recover from the latest checkpoint if present. Setting to -1 will lead to infinite recovery retries. Setting to 0 will disable retries. Defaults to 3. 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, HyperBand and PopulationBasedTraining. Refer to ray.tune.schedulers for more options. with_server (bool): Starts a background Tune server. Needed for using the Client API. server_port (int): Port number for launching TuneServer. verbose (int): 0, 1, or 2. Verbosity mode. 0 = silent, 1 = only status updates, 2 = status and trial results. progress_reporter (ProgressReporter): Progress reporter for reporting intermediate experiment progress. Defaults to CLIReporter if running in command-line, or JupyterNotebookReporter if running in a Jupyter notebook. resume (str|bool): One of "LOCAL", "REMOTE", "PROMPT", or bool. LOCAL/True restores the checkpoint from the local_checkpoint_dir. REMOTE restores the checkpoint from remote_checkpoint_dir. PROMPT provides CLI feedback. False forces a new experiment. If resume is set but checkpoint does not exist, ValueError will be thrown. queue_trials (bool): Whether to queue trials when the cluster does not currently have enough resources to launch one. This should be set to True when running on an autoscaling cluster to enable automatic scale-up. reuse_actors (bool): Whether to reuse actors between different trials when possible. This can drastically speed up experiments that start and stop actors often (e.g., PBT in time-multiplexing mode). This requires trials to have the same resource requirements. trial_executor (TrialExecutor): Manage the execution of trials. raise_on_failed_trial (bool): Raise TuneError if there exists failed trial (of ERROR state) when the experiments complete. ray_auto_init (bool): Automatically starts a local Ray cluster if using a RayTrialExecutor (which is the default) and if Ray is not initialized. Defaults to True. 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) if isinstance(run_or_experiment, list): experiments = run_or_experiment else: experiments = [run_or_experiment] if len(experiments) > 1: logger.info( "Running multiple concurrent experiments is experimental and may " "not work with certain features.") for i, exp in enumerate(experiments): if not isinstance(exp, Experiment): run_identifier = Experiment.register_if_needed(exp) experiments[i] = 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, sync_on_checkpoint=sync_on_checkpoint, keep_checkpoints_num=keep_checkpoints_num, checkpoint_score_attr=checkpoint_score_attr, export_formats=export_formats, max_failures=max_failures, restore=restore, sync_function=sync_function) else: logger.debug("Ignoring some parameters passed into tune.run.") if sync_to_cloud: for exp in experiments: assert exp.remote_checkpoint_dir, ( "Need `upload_dir` if `sync_to_cloud` given.") runner = TrialRunner( search_alg=search_alg or BasicVariantGenerator(), scheduler=scheduler or FIFOScheduler(), local_checkpoint_dir=experiments[0].checkpoint_dir, remote_checkpoint_dir=experiments[0].remote_checkpoint_dir, sync_to_cloud=sync_to_cloud, checkpoint_period=global_checkpoint_period, resume=resume, launch_web_server=with_server, server_port=server_port, verbose=bool(verbose > 1), trial_executor=trial_executor) for exp in experiments: runner.add_experiment(exp) if progress_reporter is None: if IS_NOTEBOOK: progress_reporter = JupyterNotebookReporter(overwrite=verbose < 2) else: progress_reporter = CLIReporter() # User Warning for GPUs if trial_executor.has_gpus(): if isinstance(resources_per_trial, dict) and "gpu" in resources_per_trial: # "gpu" is manually set. pass elif _check_default_resources_override(experiments[0].run_identifier): # "default_resources" is manually overriden. pass else: logger.warning("Tune detects GPUs, but no trials are using GPUs. " "To enable trials to use GPUs, set " "tune.run(resources_per_trial={'gpu': 1}...) " "which allows Tune to expose 1 GPU to each trial. " "You can also override " "`Trainable.default_resource_request` if using the " "Trainable API.") while not runner.is_finished(): runner.step() if verbose: _report_progress(runner, progress_reporter) try: runner.checkpoint(force=True) except Exception: logger.exception("Trial Runner checkpointing failed.") if verbose: _report_progress(runner, progress_reporter, done=True) wait_for_sync() 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) trials = runner.get_trials() if return_trials: return trials logger.info("Returning an analysis object by default. You can call " "`analysis.trials` to retrieve a list of trials. " "This message will be removed in future versions of Tune.") return ExperimentAnalysis(runner.checkpoint_file, trials=trials)
def __init__(self, time_attr="time_total_s", reward_attr=None, metric="episode_reward_mean", mode="max", perturbation_interval=60.0, hyperparam_mutations={}, quantile_fraction=0.25, resample_probability=0.25, custom_explore_fn=None, log_config=True, require_attrs=True, synch=False): for value in hyperparam_mutations.values(): if not (isinstance(value, (list, dict, sample_from)) or callable(value)): raise TypeError("`hyperparam_mutation` values must be either " "a List, Dict, a tune search space object, or " "callable.") if type(value) is sample_from: raise ValueError("arbitrary tune.sample_from objects are not " "supported for `hyperparam_mutation` values." "You must use other built in primitives like" "tune.uniform, tune.loguniform, etc.") if not hyperparam_mutations and not custom_explore_fn: raise TuneError( "You must specify at least one of `hyperparam_mutations` or " "`custom_explore_fn` to use PBT.") if quantile_fraction > 0.5 or quantile_fraction < 0: raise ValueError( "You must set `quantile_fraction` to a value between 0 and" "0.5. Current value: '{}'".format(quantile_fraction)) if perturbation_interval <= 0: raise ValueError( "perturbation_interval must be a positive number greater " "than 0. Current value: '{}'".format(perturbation_interval)) assert mode in ["min", "max"], "`mode` must be 'min' or 'max'!" if reward_attr is not None: mode = "max" metric = reward_attr logger.warning( "`reward_attr` is deprecated and will be removed in a future " "version of Tune. " "Setting `metric={}` and `mode=max`.".format(reward_attr)) FIFOScheduler.__init__(self) self._metric = metric if mode == "max": self._metric_op = 1. elif mode == "min": self._metric_op = -1. self._time_attr = time_attr self._perturbation_interval = perturbation_interval self._hyperparam_mutations = hyperparam_mutations self._quantile_fraction = quantile_fraction self._resample_probability = resample_probability self._trial_state = {} self._custom_explore_fn = custom_explore_fn self._log_config = log_config self._require_attrs = require_attrs self._synch = synch self._next_perturbation_sync = self._perturbation_interval # Metrics self._num_checkpoints = 0 self._num_perturbations = 0
def run( run_or_experiment: Union[str, Callable, Type], name: Optional[str] = None, metric: Optional[str] = None, mode: Optional[str] = None, stop: Union[None, Mapping, Stopper, Callable[[str, Mapping], bool]] = None, time_budget_s: Union[None, int, float, datetime.timedelta] = None, config: Optional[Dict[str, Any]] = None, resources_per_trial: Union[None, Mapping[str, Union[ float, int, Mapping]], PlacementGroupFactory] = None, num_samples: int = 1, local_dir: Optional[str] = None, search_alg: Optional[Union[Searcher, SearchAlgorithm]] = None, scheduler: Optional[TrialScheduler] = None, keep_checkpoints_num: Optional[int] = None, checkpoint_score_attr: Optional[str] = None, checkpoint_freq: int = 0, checkpoint_at_end: bool = False, verbose: Union[int, Verbosity] = Verbosity.V3_TRIAL_DETAILS, progress_reporter: Optional[ProgressReporter] = None, log_to_file: bool = False, trial_name_creator: Optional[Callable[[Trial], str]] = None, trial_dirname_creator: Optional[Callable[[Trial], str]] = None, sync_config: Optional[SyncConfig] = None, export_formats: Optional[Sequence] = None, max_failures: int = 0, fail_fast: bool = False, restore: Optional[str] = None, server_port: Optional[int] = None, resume: bool = False, queue_trials: bool = False, reuse_actors: bool = False, trial_executor: Optional[RayTrialExecutor] = None, raise_on_failed_trial: bool = True, callbacks: Optional[Sequence[Callback]] = None, # Deprecated args loggers: Optional[Sequence[Type[Logger]]] = None, ray_auto_init: Optional = None, run_errored_only: Optional = None, global_checkpoint_period: Optional = None, with_server: Optional = None, upload_dir: Optional = None, sync_to_cloud: Optional = None, sync_to_driver: Optional = None, sync_on_checkpoint: Optional = None, _remote: bool = None, ) -> ExperimentAnalysis: """Executes training. When a SIGINT signal is received (e.g. through Ctrl+C), the tuning run will gracefully shut down and checkpoint the latest experiment state. Sending SIGINT again (or SIGKILL/SIGTERM instead) will skip this step. Examples: .. code-block:: python # Run 10 trials (each trial is one instance of a Trainable). Tune runs # in parallel and automatically determines concurrency. tune.run(trainable, num_samples=10) # Run 1 trial, stop when trial has reached 10 iterations tune.run(my_trainable, stop={"training_iteration": 10}) # automatically retry failed trials up to 3 times tune.run(my_trainable, stop={"training_iteration": 10}, max_failures=3) # Run 1 trial, search over hyperparameters, stop after 10 iterations. space = {"lr": tune.uniform(0, 1), "momentum": tune.uniform(0, 1)} tune.run(my_trainable, config=space, stop={"training_iteration": 10}) # Resumes training if a previous machine crashed tune.run(my_trainable, config=space, local_dir=<path/to/dir>, resume=True) # Rerun ONLY failed trials after an experiment is finished. tune.run(my_trainable, config=space, local_dir=<path/to/dir>, resume="ERRORED_ONLY") Args: run_or_experiment (function | class | str | :class:`Experiment`): If function|class|str, this is the algorithm or model to train. This may refer to the name of a built-on algorithm (e.g. RLLib's DQN or PPO), a user-defined trainable function or class, or the string identifier of a trainable function or class registered in the tune registry. If Experiment, then Tune will execute training based on Experiment.spec. If you want to pass in a Python lambda, you will need to first register the function: ``tune.register_trainable("lambda_id", lambda x: ...)``. You can then use ``tune.run("lambda_id")``. metric (str): Metric to optimize. This metric should be reported with `tune.report()`. If set, will be passed to the search algorithm and scheduler. mode (str): Must be one of [min, max]. Determines whether objective is minimizing or maximizing the metric attribute. If set, will be passed to the search algorithm and scheduler. name (str): Name of experiment. stop (dict | callable | :class:`Stopper`): Stopping criteria. If dict, the keys may be any field in the return result of 'train()', whichever is reached first. If function, it must take (trial_id, result) as arguments and return a boolean (True if trial should be stopped, False otherwise). This can also be a subclass of ``ray.tune.Stopper``, which allows users to implement custom experiment-wide stopping (i.e., stopping an entire Tune run based on some time constraint). time_budget_s (int|float|datetime.timedelta): Global time budget in seconds after which all trials are stopped. Can also be a ``datetime.timedelta`` object. config (dict): Algorithm-specific configuration for Tune variant generation (e.g. env, hyperparams). Defaults to empty dict. Custom search algorithms may ignore this. resources_per_trial (dict|PlacementGroupFactory): Machine resources to allocate per trial, e.g. ``{"cpu": 64, "gpu": 8}``. Note that GPUs will not be assigned unless you specify them here. Defaults to 1 CPU and 0 GPUs in ``Trainable.default_resource_request()``. This can also be a PlacementGroupFactory object wrapping arguments to create a per-trial placement group. num_samples (int): Number of times to sample from the hyperparameter space. Defaults to 1. If `grid_search` is provided as an argument, the grid will be repeated `num_samples` of times. If this is -1, (virtually) infinite samples are generated until a stopping condition is met. local_dir (str): Local dir to save training results to. Defaults to ``~/ray_results``. search_alg (Searcher|SearchAlgorithm): Search algorithm for optimization. scheduler (TrialScheduler): Scheduler for executing the experiment. Choose among FIFO (default), MedianStopping, AsyncHyperBand, HyperBand and PopulationBasedTraining. Refer to ray.tune.schedulers for more options. keep_checkpoints_num (int): Number of checkpoints to keep. A value of `None` keeps all checkpoints. Defaults to `None`. If set, need to provide `checkpoint_score_attr`. checkpoint_score_attr (str): Specifies by which attribute to rank the best checkpoint. Default is increasing order. If attribute starts with `min-` it will rank attribute in decreasing order, i.e. `min-validation_loss`. checkpoint_freq (int): How many training iterations between checkpoints. A value of 0 (default) disables checkpointing. This has no effect when using the Functional Training API. checkpoint_at_end (bool): Whether to checkpoint at the end of the experiment regardless of the checkpoint_freq. Default is False. This has no effect when using the Functional Training API. verbose (Union[int, Verbosity]): 0, 1, 2, or 3. Verbosity mode. 0 = silent, 1 = only status updates, 2 = status and brief trial results, 3 = status and detailed trial results. Defaults to 3. progress_reporter (ProgressReporter): Progress reporter for reporting intermediate experiment progress. Defaults to CLIReporter if running in command-line, or JupyterNotebookReporter if running in a Jupyter notebook. log_to_file (bool|str|Sequence): Log stdout and stderr to files in Tune's trial directories. If this is `False` (default), no files are written. If `true`, outputs are written to `trialdir/stdout` and `trialdir/stderr`, respectively. If this is a single string, this is interpreted as a file relative to the trialdir, to which both streams are written. If this is a Sequence (e.g. a Tuple), it has to have length 2 and the elements indicate the files to which stdout and stderr are written, respectively. trial_name_creator (Callable[[Trial], str]): Optional function for generating the trial string representation. trial_dirname_creator (Callable[[Trial], str]): Function for generating the trial dirname. This function should take in a Trial object and return a string representing the name of the directory. The return value cannot be a path. sync_config (SyncConfig): Configuration object for syncing. See tune.SyncConfig. export_formats (list): List of formats that exported at the end of the experiment. Default is None. max_failures (int): Try to recover a trial at least this many times. Ray will recover from the latest checkpoint if present. Setting to -1 will lead to infinite recovery retries. Setting to 0 will disable retries. Defaults to 0. fail_fast (bool | str): Whether to fail upon the first error. If fail_fast='raise' provided, Tune will automatically raise the exception received by the Trainable. fail_fast='raise' can easily leak resources and should be used with caution (it is best used with `ray.init(local_mode=True)`). restore (str): Path to checkpoint. Only makes sense to set if running 1 trial. Defaults to None. server_port (int): Port number for launching TuneServer. resume (str|bool): One of "LOCAL", "REMOTE", "PROMPT", "ERRORED_ONLY", or bool. LOCAL/True restores the checkpoint from the local_checkpoint_dir, determined by `name` and `local_dir`. REMOTE restores the checkpoint from remote_checkpoint_dir. PROMPT provides CLI feedback. False forces a new experiment. ERRORED_ONLY resets and reruns ERRORED trials upon resume - previous trial artifacts will be left untouched. If resume is set but checkpoint does not exist, ValueError will be thrown. queue_trials (bool): Whether to queue trials when the cluster does not currently have enough resources to launch one. This should be set to True when running on an autoscaling cluster to enable automatic scale-up. reuse_actors (bool): Whether to reuse actors between different trials when possible. This can drastically speed up experiments that start and stop actors often (e.g., PBT in time-multiplexing mode). This requires trials to have the same resource requirements. trial_executor (TrialExecutor): Manage the execution of trials. raise_on_failed_trial (bool): Raise TuneError if there exists failed trial (of ERROR state) when the experiments complete. callbacks (list): List of callbacks that will be called at different times in the training loop. Must be instances of the ``ray.tune.callback.Callback`` class. If not passed, `LoggerCallback` and `SyncerCallback` callbacks are automatically added. _remote (bool): Whether to run the Tune driver in a remote function. This is disabled automatically if a custom trial executor is passed in. This is enabled by default in Ray client mode. Returns: ExperimentAnalysis: Object for experiment analysis. Raises: TuneError: Any trials failed and `raise_on_failed_trial` is True. """ if _remote is None: _remote = ray.util.client.ray.is_connected() if _remote is True and trial_executor: raise ValueError("cannot use custom trial executor") if not trial_executor or isinstance(trial_executor, RayTrialExecutor): _ray_auto_init() if _remote: return ray.get( ray.remote(num_cpus=0)(run).remote( run_or_experiment, name, metric, mode, stop, time_budget_s, config, resources_per_trial, num_samples, local_dir, search_alg, scheduler, keep_checkpoints_num, checkpoint_score_attr, checkpoint_freq, checkpoint_at_end, verbose, progress_reporter, log_to_file, trial_name_creator, trial_dirname_creator, sync_config, export_formats, max_failures, fail_fast, restore, server_port, resume, queue_trials, reuse_actors, trial_executor, raise_on_failed_trial, callbacks, # Deprecated args loggers, ray_auto_init, run_errored_only, global_checkpoint_period, with_server, upload_dir, sync_to_cloud, sync_to_driver, sync_on_checkpoint, _remote=False)) all_start = time.time() if global_checkpoint_period: raise ValueError("global_checkpoint_period is deprecated. Set env var " "'TUNE_GLOBAL_CHECKPOINT_S' instead.") if ray_auto_init: raise ValueError("ray_auto_init is deprecated. " "Set env var 'TUNE_DISABLE_AUTO_INIT=1' instead or " "call 'ray.init' before calling 'tune.run'.") if with_server: raise ValueError( "with_server is deprecated. It is now enabled by default " "if 'server_port' is not None.") if sync_on_checkpoint or sync_to_cloud or sync_to_driver or upload_dir: raise ValueError( "sync_on_checkpoint / sync_to_cloud / sync_to_driver / " "upload_dir must now be set via `tune.run(" "sync_config=SyncConfig(...)`. See `ray.tune.SyncConfig` for " "more details.") if mode and mode not in ["min", "max"]: raise ValueError( "The `mode` parameter passed to `tune.run()` has to be one of " "['min', 'max']") set_verbosity(verbose) config = config or {} sync_config = sync_config or SyncConfig() set_sync_periods(sync_config) if num_samples == -1: num_samples = sys.maxsize trial_executor = trial_executor or RayTrialExecutor( reuse_actors=reuse_actors, queue_trials=queue_trials) if isinstance(run_or_experiment, list): experiments = run_or_experiment else: experiments = [run_or_experiment] for i, exp in enumerate(experiments): if not isinstance(exp, Experiment): experiments[i] = Experiment( name=name, run=exp, stop=stop, time_budget_s=time_budget_s, config=config, resources_per_trial=resources_per_trial, num_samples=num_samples, local_dir=local_dir, upload_dir=sync_config.upload_dir, sync_to_driver=sync_config.sync_to_driver, trial_name_creator=trial_name_creator, trial_dirname_creator=trial_dirname_creator, log_to_file=log_to_file, checkpoint_freq=checkpoint_freq, checkpoint_at_end=checkpoint_at_end, sync_on_checkpoint=sync_config.sync_on_checkpoint, keep_checkpoints_num=keep_checkpoints_num, checkpoint_score_attr=checkpoint_score_attr, export_formats=export_formats, max_failures=max_failures, restore=restore) else: logger.debug("Ignoring some parameters passed into tune.run.") if sync_config.sync_to_cloud: for exp in experiments: assert exp.remote_checkpoint_dir, ( "Need `upload_dir` if `sync_to_cloud` given.") if fail_fast and max_failures != 0: raise ValueError("max_failures must be 0 if fail_fast=True.") if issubclass(type(search_alg), Searcher): search_alg = SearchGenerator(search_alg) if not search_alg: search_alg = BasicVariantGenerator() if config and not search_alg.set_search_properties(metric, mode, config): if has_unresolved_values(config): raise ValueError( "You passed a `config` parameter to `tune.run()` with " "unresolved parameters, but the search algorithm was already " "instantiated with a search space. Make sure that `config` " "does not contain any more parameter definitions - include " "them in the search algorithm's search space if necessary.") scheduler = scheduler or FIFOScheduler() if not scheduler.set_search_properties(metric, mode): raise ValueError( "You passed a `metric` or `mode` argument to `tune.run()`, but " "the scheduler you are using was already instantiated with their " "own `metric` and `mode` parameters. Either remove the arguments " "from your scheduler or from your call to `tune.run()`") # Create syncer callbacks callbacks = create_default_callbacks( callbacks, sync_config, metric=metric, loggers=loggers) runner = TrialRunner( search_alg=search_alg, scheduler=scheduler, local_checkpoint_dir=experiments[0].checkpoint_dir, remote_checkpoint_dir=experiments[0].remote_checkpoint_dir, sync_to_cloud=sync_config.sync_to_cloud, stopper=experiments[0].stopper, resume=resume, server_port=server_port, fail_fast=fail_fast, trial_executor=trial_executor, callbacks=callbacks, metric=metric) if not runner.resumed: for exp in experiments: search_alg.add_configurations([exp]) else: logger.info("TrialRunner resumed, ignoring new add_experiment.") if progress_reporter is None: if IS_NOTEBOOK: progress_reporter = JupyterNotebookReporter( overwrite=not has_verbosity(Verbosity.V2_TRIAL_NORM)) else: progress_reporter = CLIReporter() if not progress_reporter.set_search_properties(metric, mode): raise ValueError( "You passed a `metric` or `mode` argument to `tune.run()`, but " "the reporter you are using was already instantiated with their " "own `metric` and `mode` parameters. Either remove the arguments " "from your reporter or from your call to `tune.run()`") progress_reporter.set_total_samples(search_alg.total_samples) # User Warning for GPUs if trial_executor.has_gpus(): if isinstance(resources_per_trial, dict) and "gpu" in resources_per_trial: # "gpu" is manually set. pass elif _check_default_resources_override(experiments[0].run_identifier): # "default_resources" is manually overridden. pass else: logger.warning("Tune detects GPUs, but no trials are using GPUs. " "To enable trials to use GPUs, set " "tune.run(resources_per_trial={'gpu': 1}...) " "which allows Tune to expose 1 GPU to each trial. " "You can also override " "`Trainable.default_resource_request` if using the " "Trainable API.") original_handler = signal.getsignal(signal.SIGINT) state = {signal.SIGINT: False} def sigint_handler(sig, frame): logger.warning( "SIGINT received (e.g. via Ctrl+C), ending Ray Tune run. " "This will try to checkpoint the experiment state one last time. " "Press CTRL+C one more time (or send SIGINT/SIGKILL/SIGTERM) " "to skip. ") state[signal.SIGINT] = True # Restore original signal handler to react to future SIGINT signals signal.signal(signal.SIGINT, original_handler) if not int(os.getenv("TUNE_DISABLE_SIGINT_HANDLER", "0")): signal.signal(signal.SIGINT, sigint_handler) tune_start = time.time() while not runner.is_finished() and not state[signal.SIGINT]: runner.step() if has_verbosity(Verbosity.V1_EXPERIMENT): _report_progress(runner, progress_reporter) tune_taken = time.time() - tune_start try: runner.checkpoint(force=True) except Exception as e: logger.warning(f"Trial Runner checkpointing failed: {str(e)}") if has_verbosity(Verbosity.V1_EXPERIMENT): _report_progress(runner, progress_reporter, done=True) wait_for_sync() runner.cleanup_trials() incomplete_trials = [] for trial in runner.get_trials(): if trial.status != Trial.TERMINATED: incomplete_trials += [trial] if incomplete_trials: if raise_on_failed_trial and not state[signal.SIGINT]: raise TuneError("Trials did not complete", incomplete_trials) else: logger.error("Trials did not complete: %s", incomplete_trials) all_taken = time.time() - all_start if has_verbosity(Verbosity.V1_EXPERIMENT): logger.info(f"Total run time: {all_taken:.2f} seconds " f"({tune_taken:.2f} seconds for the tuning loop).") if state[signal.SIGINT]: logger.warning( "Experiment has been interrupted, but the most recent state was " "saved. You can continue running this experiment by passing " "`resume=True` to `tune.run()`") trials = runner.get_trials() return ExperimentAnalysis( runner.checkpoint_file, trials=trials, default_metric=metric, default_mode=mode)
def 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_function=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=False, trial_executor=None, raise_on_failed_trial=True, early_stop_all_trials=False): """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_function (func|str): Function for syncing the local_dir to upload_dir. If string, then it must be a string template for syncer to run. If not provided, the sync command defaults to standard S3 or gsutil sync comamnds. 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 (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. reuse_actors (bool): Whether to reuse actors between different trials when possible. This can drastically speed up experiments that start and stop actors often (e.g., PBT in time-multiplexing mode). This requires trials to have the same resource requirements. trial_executor (TrialExecutor): Manage the execution of trials. raise_on_failed_trial (bool): Raise TuneError if there exists failed trial (of ERROR state) when the experiments complete. Returns: 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()) } ) """ experiment = run_or_experiment if not isinstance(run_or_experiment, Experiment): experiment = Experiment(name, run_or_experiment, stop, config, resources_per_trial, num_samples, local_dir, upload_dir, trial_name_creator, loggers, sync_function, checkpoint_freq, checkpoint_at_end, export_formats, max_failures, restore) else: logger.debug("Ignoring some parameters passed into tune.run.") checkpoint_dir = _find_checkpoint_dir(experiment) should_restore = _prompt_restore(checkpoint_dir, resume) runner = None if should_restore: try: runner = TrialRunner.restore(checkpoint_dir, search_alg, scheduler, trial_executor) except Exception: logger.exception("Runner restore failed. Restarting experiment.") else: logger.info("Starting a new experiment.") if not runner: scheduler = scheduler or FIFOScheduler() search_alg = search_alg or BasicVariantGenerator() search_alg.add_configurations([experiment]) 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, reuse_actors=reuse_actors, 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 early_stop_all_trials: # Check if any trial has good validation loss, in which case we stop all trials should_stop = False for trial in runner.get_trials(): try: result = trial.last_result if any(result[criteria] >= stop_value for criteria, stop_value in trial.stopping_criterion.items()): should_stop = True break except Exception: pass if should_stop: # Checkpoint all trials for trial in runner.get_trials(): if hasattr(trial, "runner") and trial.runner: runner.trial_executor.save(trial, storage=Checkpoint.DISK) runner.stop_trial(trial) break 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 __init__(self, search_alg=None, scheduler=None, local_checkpoint_dir=None, remote_checkpoint_dir=None, sync_to_cloud=None, stopper=None, resume=False, server_port=None, fail_fast=False, checkpoint_period=None, trial_executor=None, callbacks=None, metric=None): self._search_alg = search_alg or BasicVariantGenerator() self._scheduler_alg = scheduler or FIFOScheduler() self.trial_executor = trial_executor or RayTrialExecutor() self._pending_trial_queue_times = {} # Setting this to 0 still allows adding one new (pending) trial, # but it will prevent us from trying to fill the trial list self._max_pending_trials = 0 # Can be updated in `self.add_trial()` self._metric = metric if "TRIALRUNNER_WALLTIME_LIMIT" in os.environ: raise ValueError( "The TRIALRUNNER_WALLTIME_LIMIT environment variable is " "deprecated. " "Use `tune.run(time_budget_s=limit)` instead.") self._total_time = 0 self._iteration = 0 self._has_errored = False self._fail_fast = fail_fast if isinstance(self._fail_fast, str): self._fail_fast = self._fail_fast.upper() if self._fail_fast == TrialRunner.RAISE: warnings.warn( "fail_fast='raise' detected. Be careful when using this " "mode as resources (such as Ray processes, " "file descriptors, and temporary files) may not be " "cleaned up properly. To use " "a safer mode, use fail_fast=True.") else: raise ValueError("fail_fast must be one of {bool, RAISE}. " f"Got {self._fail_fast}.") self._server = None self._server_port = server_port if server_port is not None: self._server = TuneServer(self, self._server_port) self._trials = [] self._cached_trial_decisions = {} self._queued_trial_decisions = {} self._updated_queue = False self._stop_queue = [] self._should_stop_experiment = False # used by TuneServer self._local_checkpoint_dir = local_checkpoint_dir if self._local_checkpoint_dir: os.makedirs(self._local_checkpoint_dir, exist_ok=True) self._remote_checkpoint_dir = remote_checkpoint_dir self._syncer = get_cloud_syncer(local_checkpoint_dir, remote_checkpoint_dir, sync_to_cloud) self._stopper = stopper or NoopStopper() self._resumed = False if self._validate_resume(resume_type=resume): errored_only = False if isinstance(resume, str): errored_only = resume.upper() == "ERRORED_ONLY" try: self.resume(run_errored_only=errored_only) self._resumed = True except Exception as e: if has_verbosity(Verbosity.V3_TRIAL_DETAILS): logger.error(str(e)) logger.exception("Runner restore failed.") if self._fail_fast: raise logger.info("Restarting experiment.") else: logger.debug("Starting a new experiment.") self._start_time = time.time() self._last_checkpoint_time = -float("inf") self._session_str = datetime.fromtimestamp( self._start_time).strftime("%Y-%m-%d_%H-%M-%S") self.checkpoint_file = None if self._local_checkpoint_dir: self.checkpoint_file = os.path.join( self._local_checkpoint_dir, TrialRunner.CKPT_FILE_TMPL.format(self._session_str)) self._callbacks = CallbackList(callbacks or []) self._callbacks.setup() if checkpoint_period is None: checkpoint_period = os.getenv("TUNE_GLOBAL_CHECKPOINT_S", "auto") self._checkpoint_period = checkpoint_period self._checkpoint_manager = self._create_checkpoint_manager()
"scale": 0.2 }, "type": "double" }], metrics=[ dict(name="std", objective="minimize", strategy="optimize"), dict(name="average", strategy="store") ], observation_budget=samples, parallel_bandwidth=1) config = {"num_samples": samples, "config": {}} algo = SigOptSearch(connection=conn, experiment_id=experiment.id, name="SigOpt Example Existing Experiment", max_concurrent=1, metric=["average", "std"], mode=["obs", "min"]) scheduler = FIFOScheduler() tune.run(easy_objective, name="my_exp", search_alg=algo, scheduler=scheduler, **config)
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 compile( self, input_df, model_create_func, search_space, recipe, feature_transformers=None, # model=None, future_seq_len=1, validation_df=None, mc=False, metric="mse", metric_mode="min"): """ Do necessary preparations for the engine :param input_df: :param search_space: :param num_samples: :param stop: :param search_algorithm: :param search_algorithm_params: :param fixed_params: :param feature_transformers: :param model: :param validation_df: :param metric: :return: """ # prepare parameters for search engine runtime_params = recipe.runtime_params() num_samples = runtime_params['num_samples'] stop = dict(runtime_params) search_algorithm_params = recipe.search_algorithm_params() search_algorithm = recipe.search_algorithm() fixed_params = recipe.fixed_params() schedule_algorithm = recipe.scheduler_algorithm() del stop['num_samples'] self.search_space = self._prepare_tune_config(search_space) self.stop_criteria = stop self.num_samples = num_samples if schedule_algorithm == 'AsyncHyperBand': from ray.tune.schedulers import AsyncHyperBandScheduler self.sched = AsyncHyperBandScheduler( time_attr="training_iteration", metric="reward_metric", mode="max", max_t=50, grace_period=1, reduction_factor=3, brackets=3, ) else: from ray.tune.schedulers import FIFOScheduler self.sched = FIFOScheduler() if search_algorithm == 'BayesOpt': self.search_algorithm = BayesOptSearch( self.search_space, metric="reward_metric", mode="max", utility_kwargs=search_algorithm_params["utility_kwargs"]) elif search_algorithm == 'SkOpt': from skopt import Optimizer from ray.tune.suggest.skopt import SkOptSearch opt_params = recipe.opt_params() optimizer = Optimizer(opt_params) self.search_algorithm = SkOptSearch( optimizer, list(self.search_space.keys()), metric="reward_metric", mode="max", ) else: self.search_algorithm = None self.fixed_params = fixed_params self.train_func = self._prepare_train_func( input_df=input_df, model_create_func=model_create_func, feature_transformers=feature_transformers, future_seq_len=future_seq_len, validation_df=validation_df, metric=metric, metric_mode=metric_mode, mc=mc, remote_dir=self.remote_dir)
for epoch in range(1, args.epochs + 1): train(epoch) test() if __name__ == "__main__": args = parser.parse_args() import ray from ray import tune from ray.tune.schedulers import AsyncHyperBandScheduler, FIFOScheduler ray.init() if args.scheduler == 'fifo': sched = FIFOScheduler() elif args.scheduler == 'asynchyperband': sched = AsyncHyperBandScheduler(time_attr="training_iteration", reward_attr="neg_mean_loss", max_t=400, grace_period=60) else: raise NotImplementedError tune.register_trainable( "TRAIN_FN", lambda config, reporter: train_indoor(args, config, reporter)) tune.run( "TRAIN_FN", name=args.expname, verbose=2, scheduler=sched,
def __init__(self, search_alg=None, scheduler=None, launch_web_server=False, local_checkpoint_dir=None, remote_checkpoint_dir=None, sync_to_cloud=None, resume=False, server_port=TuneServer.DEFAULT_PORT, verbose=True, checkpoint_period=10, trial_executor=None): """Initializes a new TrialRunner. Args: search_alg (SearchAlgorithm): SearchAlgorithm for generating Trial objects. scheduler (TrialScheduler): Defaults to FIFOScheduler. launch_web_server (bool): Flag for starting TuneServer local_checkpoint_dir (str): Path where global checkpoints are stored and restored from. remote_checkpoint_dir (str): Remote path where global checkpoints are stored and restored from. Used if `resume` == REMOTE. resume (str|False): see `tune.py:run`. sync_to_cloud (func|str): see `tune.py:run`. server_port (int): Port number for launching TuneServer. verbose (bool): Flag for verbosity. If False, trial results will not be output. trial_executor (TrialExecutor): Defaults to RayTrialExecutor. """ self._search_alg = search_alg or BasicVariantGenerator() self._scheduler_alg = scheduler or FIFOScheduler() self.trial_executor = trial_executor or RayTrialExecutor() # For debugging, it may be useful to halt trials after some time has # elapsed. TODO(ekl) consider exposing this in the API. self._global_time_limit = float( os.environ.get("TRIALRUNNER_WALLTIME_LIMIT", float("inf"))) self._total_time = 0 self._iteration = 0 self._verbose = verbose self._server = None self._server_port = server_port if launch_web_server: self._server = TuneServer(self, self._server_port) self._trials = [] self._stop_queue = [] self._local_checkpoint_dir = local_checkpoint_dir if self._local_checkpoint_dir and not os.path.exists( self._local_checkpoint_dir): os.makedirs(self._local_checkpoint_dir) self._remote_checkpoint_dir = remote_checkpoint_dir self._syncer = get_syncer(local_checkpoint_dir, remote_checkpoint_dir, sync_to_cloud) self._resumed = False if self._validate_resume(resume_type=resume): try: self.resume() logger.info("Resuming trial.") self._resumed = True except Exception: logger.exception( "Runner restore failed. Restarting experiment.") else: logger.info("Starting a new experiment.") self._start_time = time.time() self._last_checkpoint_time = -float("inf") self._checkpoint_period = checkpoint_period self._session_str = datetime.fromtimestamp( self._start_time).strftime("%Y-%m-%d_%H-%M-%S")
def __init__( self, time_attr: str = "time_total_s", metric: Optional[str] = None, mode: Optional[str] = None, perturbation_interval: float = 60.0, burn_in_period: float = 0.0, hyperparam_mutations: Dict = None, quantile_fraction: float = 0.25, resample_probability: float = 0.25, custom_explore_fn: Optional[Callable] = None, log_config: bool = True, require_attrs: bool = True, synch: bool = False, ): hyperparam_mutations = hyperparam_mutations or {} for value in hyperparam_mutations.values(): if not (isinstance(value, (list, dict, Domain)) or callable(value)): raise TypeError("`hyperparam_mutation` values must be either " "a List, Dict, a tune search space object, or " "a callable.") if isinstance(value, Function): raise ValueError("arbitrary tune.sample_from objects are not " "supported for `hyperparam_mutation` values." "You must use other built in primitives like" "tune.uniform, tune.loguniform, etc.") if not hyperparam_mutations and not custom_explore_fn: raise TuneError( "You must specify at least one of `hyperparam_mutations` " "or `custom_explore_fn` to use PBT.") if quantile_fraction > 0.5 or quantile_fraction < 0: raise ValueError( "You must set `quantile_fraction` to a value between 0 and" "0.5. Current value: '{}'".format(quantile_fraction)) if perturbation_interval <= 0: raise ValueError( "perturbation_interval must be a positive number greater " "than 0. Current value: '{}'".format(perturbation_interval)) if mode: assert mode in ["min", "max"], "`mode` must be 'min' or 'max'." FIFOScheduler.__init__(self) self._metric = metric self._mode = mode self._metric_op = None if self._mode == "max": self._metric_op = 1.0 elif self._mode == "min": self._metric_op = -1.0 self._time_attr = time_attr self._perturbation_interval = perturbation_interval self._burn_in_period = burn_in_period self._hyperparam_mutations = hyperparam_mutations self._quantile_fraction = quantile_fraction self._resample_probability = resample_probability self._trial_state = {} self._custom_explore_fn = custom_explore_fn self._log_config = log_config self._require_attrs = require_attrs self._synch = synch self._next_perturbation_sync = max( self._perturbation_interval, self._burn_in_period, ) # Metrics self._num_checkpoints = 0 self._num_perturbations = 0
def __init__(self, time_attr="time_total_s", reward_attr=None, metric="episode_reward_mean", mode="max", perturbation_interval=60.0, hyperparam_mutations={}, quantile_fraction=0.25, resample_probability=0.25, custom_explore_fn=None, log_config=True): if not hyperparam_mutations and not custom_explore_fn: raise TuneError( "You must specify at least one of `hyperparam_mutations` or " "`custom_explore_fn` to use PBT.") if quantile_fraction > 0.5 or quantile_fraction < 0: raise TuneError( "You must set `quantile_fraction` to a value between 0 and" "0.5. Current value: '{}'".format(quantile_fraction)) assert mode in ["min", "max"], "`mode` must be 'min' or 'max'!" if reward_attr is not None: mode = "max" metric = reward_attr logger.warning( "`reward_attr` is deprecated and will be removed in a future " "version of Tune. " "Setting `metric={}` and `mode=max`.".format(reward_attr)) FIFOScheduler.__init__(self) self._metric = metric if mode == "max": self._metric_op = 1. elif mode == "min": self._metric_op = -1. self._time_attr = time_attr self._perturbation_interval = perturbation_interval self._hyperparam_mutations = hyperparam_mutations self._quantile_fraction = quantile_fraction self._resample_probability = resample_probability self._trial_state = {} self._custom_explore_fn = custom_explore_fn self._log_config = log_config self.meta = { 'timesteps': [], 'lengthscales': [], 'closest': [], 'meandist': [] } self.latest = 0 # when we last did bayesopt self.data = pd.DataFrame() self.bounds = {} for key, distribution in self._hyperparam_mutations.items(): self.bounds[key] = [ np.min([distribution() for _ in range(999999)]), np.max([distribution() for _ in range(999999)]) ] # Metrics self._num_checkpoints = 0 self._num_perturbations = 0