[tune] Trainable functions cannot execute Ray methods since Ray is not thread safe #1851
Description
Describe the problem
The following snippet will crash with
ray.tune.error.TuneError: Error running trial: The Ray methods are not thread safe and must be called from the main thread. This method was called from thread Thread-2.
import ray
from ray.tune import run_experiments, register_trainable
ray.init()
def my_train(config, reporter): # add the reporter parameter
ray.put("HI")
ray.tune.register_trainable("my_train", my_train)
ray.tune.run_experiments({
"my_experiment": {
"run": "my_train",
}
})
This is since Tune has to run long-running functions off of the main thread. This prevents tune users from interacting with Ray when using the trainable functions interface.
One workaround is to define the trainable as a Trainable class, which tune does run on the main thread since train() yields control back to the tune periodically for returning results.
Metadata
Metadata
Assignees
Labels
No labels