Ray.tune: отслеживаемый актер не управляется ⇐ Python

[Anonymous]

Код: Выделить всё

from ray import tune
from ray.tune.search.optuna import OptunaSearch
from ray.tune.integration.pytorch_lightning import TuneReportCheckpointCallback
from pytorch_lightning import Trainer
from pytorch_lightning.callbacks import EarlyStopping
from data_module import DataModule
from fcnn_regressor import FCNNRegressor
from typing import Any

dm = DataModule(X_train, X_test, y_train, y_test,
batch_size=2028)

def train_model(config: dict[str, Any]) -> None:
model = FCNNRegressor(
input_dim=X_train.shape[1],
hidden_dim1=config["hidden_dim1"],
hidden_dim2=config["hidden_dim2"],
dropout_prob=config["dropout_prob"],
lr=config["lr"],
weight_decay=config["weight_decay"],
)

trainer = Trainer(
max_epochs=100,
callbacks=[
EarlyStopping(monitor="val_loss", patience=5, mode="min"),
TuneReportCheckpointCallback({"val_loss": "val_loss"},
on="validation_end")
],
)
trainer.fit(model, datamodule=dm)

search_space = {
"hidden_dim1": tune.randint(64, 257),
"hidden_dim2": tune.sample_from(
lambda cfg: np.random.randint(16, cfg["hidden_dim1"] // 2 + 1)),
"dropout_prob": tune.uniform(0.1, 0.4),
"lr": tune.loguniform(1e-5, 1e-2),
"weight_decay": tune.loguniform(1e-6, 1e-2),
}

optuna = OptunaSearch(metric="val_loss", mode="min")

tuner = tune.Tuner(
train_model,
param_space=search_space,
tune_config=tune.TuneConfig(num_samples=20, search_alg=optuna),
)

results = tuner.fit()
print(results.get_best_result("val_loss", "min").metrics)
< /code>
Почему существует следующая ошибка: < /p>
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
File ~/pyenvs/base/lib/python3.12/site-packages/ray/tune/tune.py:994, in run(run_or_experiment, name, metric, mode, stop, time_budget_s, config, resources_per_trial, num_samples, storage_path, storage_filesystem, search_alg, scheduler, checkpoint_config, verbose, progress_reporter, log_to_file, trial_name_creator, trial_dirname_creator, sync_config, export_formats, max_failures, fail_fast, restore, resume, resume_config, reuse_actors, raise_on_failed_trial, callbacks, max_concurrent_trials, keep_checkpoints_num, checkpoint_score_attr, checkpoint_freq, checkpoint_at_end, chdir_to_trial_dir, local_dir, _remote, _remote_string_queue, _entrypoint)
993 while not runner.is_finished() and not experiment_interrupted_event.is_set():
--> 994     runner.step()
995     if has_verbosity(Verbosity.V1_EXPERIMENT):

File ~/pyenvs/base/lib/python3.12/site-packages/ray/tune/execution/tune_controller.py:685, in TuneController.step(self)
684 # Handle one event
--> 685 if not self._actor_manager.next(timeout=0.1):
686     # If there are no actors running, warn about potentially
687     # insufficient resources
688     if not self._actor_manager.num_live_actors:

File ~/pyenvs/base/lib/python3.12/site-packages/ray/air/execution/_internal/actor_manager.py:225, in RayActorManager.next(self, timeout)
224 else:
--> 225     self._handle_ready_resource_future()
226     # Ready resource futures don't count as one event as they don't trigger
227     # any callbacks. So we repeat until we hit anything that is not a resource
228     # future.

File ~/pyenvs/base/lib/python3.12/site-packages/ray/air/execution/_internal/actor_manager.py:310, in RayActorManager._handle_ready_resource_future(self)
309 # We handle resource futures one by one, so only try to start 1 actor at a time
--> 310 self._try_start_actors(max_actors=1)
...
732     self._pending_actors_to_enqueued_actor_tasks[tracked_actor].append(
733         (tracked_actor_task, method_name, args, kwargs)
734     )

ValueError: Tracked actor is not managed by this event manager: 
Output is truncated. View as a scrollable element or open in a text editor.  Adjust cell output settings...
< /code>
Если это необходимо, вы можете проверить код модулей: < /p>
[list]
[*]fcnn_regressor

- https://pastebin.com/ngqcvnji
[*]

Код: Выделить всё

data_module

- https://pastebin.com/9x167drn
[/list]
Я попытался использовать функцию ray.put и положить DataModule intar_model . Первый путь приводит к ошибке, связанной с «недопустимым типом объекта Refs, , дается» или «неверный тип объекта, , дается». Второй путь не разрешил проблему
Данные являются Torch.tensor размера [463715, 90] (только числовые столбцы)

Подробнее здесь: https://stackoverflow.com/questions/796 ... ot-managed