Tune User Guide

The basic Tune API [tune.run(Trainable)] has two main parts: a Training API and tune.run.

Training API

Training can be done with either a Class API (tune.Trainable) or function-based API (track.log). Here is an example tune.Trainable that you can use to dry-run Tune:

from ray import tune

class trainable(tune.Trainable):
    def _setup(self, config):
        if config["print_me"]:
            print(config["print_me"])

    def _train(self):
        # run one step of training code.
        # important: this method is called repeatedly!
        result_dict = {"accuracy": 0.5, "f1": 0.1, ...}
        return result_dict

tune.run(trainable, config={"print_me": "hello-world"}, stop={"training_iteration": 200})

The function-based API is for fast prototyping but has limited functionality. Here is a function-based API example:

from ray import tune
import time

def trainable(config):
    if config["print_me"]:
        print(config["print_me"])

    for i in range(200):
        time.sleep(1)
        result_dict = {"accuracy": 0.5, "f1": 0.1, ...}
        tune.track.log(**result_dict)

tune.run(trainable, config={"print_me": "hello-world"})

To read more, check out the Trainable API docs.

Running Tune

Use tune.run to generate and execute your hyperparameter sweep:

tune.run(trainable)

# Run a total of 10 evaluations of the Trainable. Tune runs in
# parallel and automatically determines concurrency.
tune.run(trainable, num_samples=10)

This function will report status on the command line until all Trials stop:

== Status ==
Memory usage on this node: 11.4/16.0 GiB
Using FIFO scheduling algorithm.
Resources requested: 4/12 CPUs, 0/0 GPUs, 0.0/3.17 GiB heap, 0.0/1.07 GiB objects
Result logdir: /Users/foo/ray_results/myexp
Number of trials: 4 (4 RUNNING)
+----------------------+----------+---------------------+-----------+--------+--------+--------+--------+------------------+-------+
| Trial name           | status   | loc                 |    param1 | param2 | param3 |    acc |   loss |   total time (s) |  iter |
|----------------------+----------+---------------------+-----------+--------+--------+--------+--------+------------------+-------|
| MyTrainable_a826033a | RUNNING  | 10.234.98.164:31115 | 0.303706  | 0.0761 | 0.4328 | 0.1289 | 1.8572 |          7.54952 |    15 |
| MyTrainable_a8263fc6 | RUNNING  | 10.234.98.164:31117 | 0.929276  | 0.158  | 0.3417 | 0.4865 | 1.6307 |          7.0501  |    14 |
| MyTrainable_a8267914 | RUNNING  | 10.234.98.164:31111 | 0.068426  | 0.0319 | 0.1147 | 0.9585 | 1.9603 |          7.0477  |    14 |
| MyTrainable_a826b7bc | RUNNING  | 10.234.98.164:31112 | 0.729127  | 0.0748 | 0.1784 | 0.1797 | 1.7161 |          7.05715 |    14 |
+----------------------+----------+---------------------+-----------+--------+--------+--------+--------+------------------+-------+

All results reported by the trainable will be logged locally to a unique directory per experiment, e.g. ~/ray_results/example-experiment in the above example. On a cluster, incremental results will be synced to local disk on the head node. All results will have autofilled metrics in addition to your own user-defined metrics.

Trial Parallelism

Tune automatically runs N concurrent trials, where N is the number of CPUs (cores) on your machine. By default, Tune assumes that each trial will only require 1 CPU. You can override this with resources_per_trial:

# If you have 4 CPUs on your machine, this will run 4 concurrent trials at a time.
tune.run(trainable, num_samples=10)

# If you have 4 CPUs on your machine, this will run 2 concurrent trials at a time.
tune.run(trainable, num_samples=10, resources_per_trial={"cpu": 2})

# If you have 4 CPUs on your machine, this will run 1 trial at a time.
tune.run(trainable, num_samples=10, resources_per_trial={"cpu": 4})

To leverage GPUs, you can set gpu in resources_per_trial. A trial will only be executed if there are resources available. See the section on resource allocation, which provides more details about GPU usage and trials that are distributed:

# If you have 4 CPUs on your machine and 1 GPU, this will run 1 trial at a time.
tune.run(trainable, num_samples=10, resources_per_trial={"cpu": 2, "gpu": 1})

To attach to a Ray cluster or use ray.init manual resource overrides, simply run ray.init before tune.run:

# Setup a local ray cluster and override resources. This will run 50 trials in parallel:
ray.init(num_cpus=100)
tune.run(trainable, num_samples=100, resources_per_trial={"cpu": 2})

# Connect to an existing distributed Ray cluster
ray.init(address=<ray_redis_address>)
tune.run(trainable, num_samples=100, resources_per_trial={"cpu": 2, "gpu": 1})

Tip

To run everything sequentially, use Ray Local Mode.

Analyzing Results

Tune provides an ExperimentAnalysis object for analyzing results from tune.run.

analysis = tune.run(
    trainable,
    name="example-experiment",
    num_samples=10,
)

You can use the ExperimentAnalysis object to obtain the best configuration of the experiment:

>>> print("Best config is", analysis.get_best_config(metric="mean_accuracy"))
Best config is: {'lr': 0.011537575723482687, 'momentum': 0.8921971713692662}

See the full documentation for the Analysis object: ExperimentAnalysis.

Custom Trial Names

To specify custom trial names, you can pass use the trial_name_creator argument to tune.run. This takes a function with the following signature:

def trial_name_string(trial):
    """
    Args:
        trial (Trial): A generated trial object.

    Returns:
        trial_name (str): String representation of Trial.
    """
    return str(trial)

tune.run(
    MyTrainableClass,
    name="example-experiment",
    num_samples=1,
    trial_name_creator=trial_name_string
)

An example can be found in logging_example.py.

Sampling Multiple Times

By default, each random variable and grid search point is sampled once. To take multiple random samples, add num_samples: N to the experiment config. If grid_search is provided as an argument, the grid will be repeated num_samples of times.

 tune.run(
     my_trainable,
     name="my_trainable",
     config={
         "alpha": tune.sample_from(lambda spec: np.random.uniform(100)),
         "beta": tune.sample_from(lambda spec: spec.config.alpha * np.random.normal()),
         "nn_layers": [
             tune.grid_search([16, 64, 256]),
             tune.grid_search([16, 64, 256]),
         ],
     },
     num_samples=10
 )

E.g. in the above, num_samples=10 repeats the 3x3 grid search 10 times, for a total of 90 trials, each with randomly sampled values of alpha and beta.

Resource Allocation (Using GPUs)

Tune will allocate the specified GPU and CPU resources_per_trial to each individual trial (defaulting to 1 CPU per trial). Under the hood, Tune runs each trial as a Ray actor, using Ray’s resource handling to allocate resources and place actors. A trial will not be scheduled unless at least that amount of resources is available in the cluster, preventing the cluster from being overloaded.

Fractional values are also supported, (i.e., "gpu": 0.2). You can find an example of this in the Keras MNIST example.

If GPU resources are not requested, the CUDA_VISIBLE_DEVICES environment variable will be set as empty, disallowing GPU access. Otherwise, it will be set to the GPUs in the list (this is managed by Ray).

Advanced Resource Allocation

Trainables can themselves be distributed. If your trainable function / class creates further Ray actors or tasks that also consume CPU / GPU resources, you will also want to set extra_cpu or extra_gpu to reserve extra resource slots for the actors you will create. For example, if a trainable class requires 1 GPU itself, but will launch 4 actors each using another GPU, then it should set "gpu": 1, "extra_gpu": 4.

 tune.run(
     my_trainable,
     name="my_trainable",
     resources_per_trial={
         "cpu": 1,
         "gpu": 1,
         "extra_gpu": 4
     }
 )

The Trainable also provides the default_resource_requests interface to automatically declare the resources_per_trial based on the given configuration.

classmethod Trainable.default_resource_request(config)[source]

Provides a static resource requirement for the given configuration.

This can be overriden by sub-classes to set the correct trial resource allocation, so the user does not need to.

@classmethod
def default_resource_request(cls, config):
    return Resources(
        cpu=0,
        gpu=0,
        extra_cpu=config["workers"],
        extra_gpu=int(config["use_gpu"]) * config["workers"])
Returns

A Resources object consumed by Tune for queueing.

Return type

Resources

Trainable (Trial) Checkpointing

When running a hyperparameter search, Tune can automatically and periodically save/checkpoint your model. Checkpointing is used for

  • saving a model at the end of training

  • modifying a model in the middle of training

  • fault-tolerance in experiments with pre-emptible machines.

  • enables certain Trial Schedulers such as HyperBand and PBT.

To enable checkpointing, you must implement a Trainable class (Trainable functions are not checkpointable, since they never return control back to their caller).

Checkpointing assumes that the model state will be saved to disk on whichever node the Trainable is running on. You can checkpoint with three different mechanisms: manually, periodically, and at termination.

Manual Checkpointing: A custom Trainable can manually trigger checkpointing by returning should_checkpoint: True (or tune.result.SHOULD_CHECKPOINT: True) in the result dictionary of _train. This can be especially helpful in spot instances:

def _train(self):
    # training code
    result = {"mean_accuracy": accuracy}
    if detect_instance_preemption():
        result.update(should_checkpoint=True)
    return result

Periodic Checkpointing: periodic checkpointing can be used to provide fault-tolerance for experiments. This can be enabled by setting checkpoint_freq=<int> and max_failures=<int> to checkpoint trials every N iterations and recover from up to M crashes per trial, e.g.:

tune.run(
    my_trainable,
    checkpoint_freq=10,
    max_failures=5,
)

Checkpointing at Termination: The checkpoint_freq may not coincide with the exact end of an experiment. If you want a checkpoint to be created at the end of a trial, you can additionally set the checkpoint_at_end=True:

 tune.run(
     my_trainable,
     checkpoint_freq=10,
     checkpoint_at_end=True,
     max_failures=5,
 )

The checkpoint will be saved at a path that looks like local_dir/exp_name/trial_name/checkpoint_x/, where the x is the number of iterations so far when the checkpoint is saved. To restore the checkpoint, you can use the restore argument and specify a checkpoint file. By doing this, you can change whatever experiments’ configuration such as the experiment’s name, the training iteration or so:

# Restored previous trial from the given checkpoint
tune.run(
    "PG",
    name="RestoredExp", # The name can be different.
    stop={"training_iteration": 10}, # train 5 more iterations than previous
    restore="~/ray_results/Original/PG_<xxx>/checkpoint_5/checkpoint-5",
    config={"env": "CartPole-v0"},
)

Fault Tolerance

Tune will automatically restart trials in case of trial failures/error (if max_failures != 0), both in the single node and distributed setting.

Tune will restore trials from the latest checkpoint, where available. In the distributed setting, if using the autoscaler with rsync enabled, Tune will automatically sync the trial folder with the driver. For example, if a node is lost while a trial (specifically, the corresponding Trainable actor of the trial) is still executing on that node and a checkpoint of the trial exists, Tune will wait until available resources are available to begin executing the trial again.

If the trial/actor is placed on a different node, Tune will automatically push the previous checkpoint file to that node and restore the remote trial actor state, allowing the trial to resume from the latest checkpoint even after failure.

Take a look at an example: Pre-emptible Instances (Cloud).

Recovering From Failures

Tune automatically persists the progress of your entire experiment (a tune.run session), so if an experiment crashes or is otherwise cancelled, it can be resumed by passing one of True, False, “LOCAL”, “REMOTE”, or “PROMPT” to tune.run(resume=...). Note that this only works if trial checkpoints are detected, whether it be by manual or periodic checkpointing.

Settings:

  • The default setting of resume=False creates a new experiment.

  • resume="LOCAL" and resume=True restore the experiment from local_dir/[experiment_name].

  • resume="REMOTE" syncs the upload dir down to the local dir and then restores the experiment from local_dir/experiment_name.

  • resume="PROMPT" will cause Tune to prompt you for whether you want to resume. You can always force a new experiment to be created by changing the experiment name.

Note that trials will be restored to their last checkpoint. If trial checkpointing is not enabled, unfinished trials will be restarted from scratch.

E.g.:

tune.run(
    my_trainable,
    checkpoint_freq=10,
    local_dir="~/path/to/results",
    resume=True
)

Upon a second run, this will restore the entire experiment state from ~/path/to/results/my_experiment_name. Importantly, any changes to the experiment specification upon resume will be ignored. For example, if the previous experiment has reached its termination, then resuming it with a new stop criterion makes no effect: the new experiment will terminate immediately after initialization. If you want to change the configuration, such as training more iterations, you can do so restore the checkpoint by setting restore=<path-to-checkpoint> - note that this only works for a single trial.

Warning

This feature is still experimental, so any provided Trial Scheduler or Search Algorithm will not be preserved. Only FIFOScheduler and BasicVariantGenerator will be supported.

Handling Large Datasets

You often will want to compute a large object (e.g., training data, model weights) on the driver and use that object within each trial. Tune provides a pin_in_object_store utility function that can be used to broadcast such large objects. Objects pinned in this way will never be evicted from the Ray object store while the driver process is running, and can be efficiently retrieved from any task via get_pinned_object.

import ray
from ray import tune
from ray.tune.utils import pin_in_object_store, get_pinned_object

import numpy as np

ray.init()

# X_id can be referenced in closures
X_id = pin_in_object_store(np.random.random(size=100000000))

def f(config, reporter):
    X = get_pinned_object(X_id)
    # use X

tune.run(f)

Custom Stopping Criteria

You can control when trials are stopped early by passing the stop argument to tune.run. This argument takes either a dictionary or a function.

If a dictionary is passed in, the keys may be any field in the return result of tune.track.log in the Function API or train() (including the results from _train and auto-filled metrics).

In the example below, each trial will be stopped either when it completes 10 iterations OR when it reaches a mean accuracy of 0.98. Note that training_iteration is an auto-filled metric by Tune.

tune.run(
    my_trainable,
    stop={"training_iteration": 10, "mean_accuracy": 0.98}
)

For more flexibility, you can pass in a function instead. If a function is passed in, it must take (trial_id, result) as arguments and return a boolean (True if trial should be stopped and False otherwise).

def stopper(trial_id, result):
    return result["mean_accuracy"] / result["training_iteration"] > 5

tune.run(my_trainable, stop=stopper)

Finally, you can implement the Stopper abstract class for stopping entire experiments. For example, the following example stops all trials after the criteria is fulfilled by any individual trial, and prevents new ones from starting:

from ray.tune import Stopper

class CustomStopper(Stopper):
    def __init__(self):
        self.should_stop = False

    def __call__(self, trial_id, result):
        if not self.should_stop and result['foo'] > 10:
            self.should_stop = True
        return self.should_stop

    def stop_all(self):
        """Returns whether to stop trials and prevent new ones from starting."""
        return self.should_stop

stopper = CustomStopper()
tune.run(my_trainable, stop=stopper)

Note that in the above example the currently running trials will not stop immediately but will do so once their current iterations are complete.

Auto-Filled Results

During training, Tune will automatically fill certain fields if not already provided. All of these can be used as stopping conditions or in the Scheduler/Search Algorithm specification.

# (Optional/Auto-filled) training is terminated. Filled only if not provided.
DONE = "done"

# (Optional) Enum for user controlled checkpoint
SHOULD_CHECKPOINT = "should_checkpoint"

# (Auto-filled) The hostname of the machine hosting the training process.
HOSTNAME = "hostname"

# (Auto-filled) The auto-assigned id of the trial.
TRIAL_ID = "trial_id"

# (Auto-filled) The auto-assigned id of the trial.
EXPERIMENT_TAG = "experiment_tag"

# (Auto-filled) The node ip of the machine hosting the training process.
NODE_IP = "node_ip"

# (Auto-filled) The pid of the training process.
PID = "pid"

# (Optional) Mean reward for current training iteration
EPISODE_REWARD_MEAN = "episode_reward_mean"

# (Optional) Mean loss for training iteration
MEAN_LOSS = "mean_loss"

# (Optional) Mean accuracy for training iteration
MEAN_ACCURACY = "mean_accuracy"

# Number of episodes in this iteration.
EPISODES_THIS_ITER = "episodes_this_iter"

# (Optional/Auto-filled) Accumulated number of episodes for this trial.
EPISODES_TOTAL = "episodes_total"

# Number of timesteps in this iteration.
TIMESTEPS_THIS_ITER = "timesteps_this_iter"

# (Auto-filled) Accumulated number of timesteps for this entire trial.
TIMESTEPS_TOTAL = "timesteps_total"

# (Auto-filled) Time in seconds this iteration took to run.
# This may be overriden to override the system-computed time difference.
TIME_THIS_ITER_S = "time_this_iter_s"

# (Auto-filled) Accumulated time in seconds for this entire trial.
TIME_TOTAL_S = "time_total_s"

# (Auto-filled) The index of this training iteration.
TRAINING_ITERATION = "training_iteration"

The following fields will automatically show up on the console output, if provided:

  1. episode_reward_mean

  2. mean_loss

  3. mean_accuracy

  4. timesteps_this_iter (aggregated into timesteps_total).

TensorBoard

To visualize learning in tensorboard, install tensorboardX:

$ pip install tensorboardX

Then, after you run a experiment, you can visualize your experiment with TensorBoard by specifying the output directory of your results. Note that if you running Ray on a remote cluster, you can forward the tensorboard port to your local machine through SSH using ssh -L 6006:localhost:6006 <address>:

$ tensorboard --logdir=~/ray_results/my_experiment

If you are running Ray on a remote multi-user cluster where you do not have sudo access, you can run the following commands to make sure tensorboard is able to write to the tmp directory:

$ export TMPDIR=/tmp/$USER; mkdir -p $TMPDIR; tensorboard --logdir=~/ray_results
_images/ray-tune-tensorboard.png

If using TF2, Tune also automatically generates TensorBoard HParams output, as shown below:

tune.run(
    ...,
    config={
        "lr": tune.grid_search([1e-5, 1e-4]),
        "momentum": tune.grid_search([0, 0.9])
    }
)
_images/tune-hparams.png

Logging

You can pass in your own logging mechanisms to output logs in custom formats as follows:

from ray.tune.logger import DEFAULT_LOGGERS

tune.run(
    MyTrainableClass,
    name="experiment_name",
    loggers=DEFAULT_LOGGERS + (CustomLogger1, CustomLogger2)
)

These loggers will be called along with the default Tune loggers. All loggers must inherit the Logger interface (Logger). Tune enables default loggers for Tensorboard, CSV, and JSON formats. You can also check out logger.py for implementation details. An example can be found in logging_example.py. See the Logging API.

Uploading/Syncing

Tune automatically syncs the trial folder on remote nodes back to the head node. This requires the ray cluster to be started with the autoscaler. By default, local syncing requires rsync to be installed. You can customize the sync command with the sync_to_driver argument in tune.run by providing either a function or a string.

If a string is provided, then it must include replacement fields {source} and {target}, like rsync -savz -e "ssh -i ssh_key.pem" {source} {target}. Alternatively, a function can be provided with the following signature:

def custom_sync_func(source, target):
    sync_cmd = "rsync {source} {target}".format(
        source=source,
        target=target)
    sync_process = subprocess.Popen(sync_cmd, shell=True)
    sync_process.wait()

tune.run(
    MyTrainableClass,
    name="experiment_name",
    sync_to_driver=custom_sync_func,
)

When syncing results back to the driver, the source would be a path similar to ubuntu@192.0.0.1:/home/ubuntu/ray_results/trial1, and the target would be a local path. This custom sync command would be also be used in node failures, where the source argument would be the path to the trial directory and the target would be a remote path. The sync_to_driver would be invoked to push a checkpoint to new node for a queued trial to resume.

If an upload directory is provided, Tune will automatically sync results to the given directory, natively supporting standard S3/gsutil commands. You can customize this to specify arbitrary storages with the sync_to_cloud argument. This argument is similar to sync_to_cloud in that it supports strings with the same replacement fields and arbitrary functions. See syncer.py for implementation details.

tune.run(
    MyTrainableClass,
    name="experiment_name",
    sync_to_cloud=custom_sync_func,
)

Debugging

By default, Tune will run hyperparameter evaluations on multiple processes. However, if you need to debug your training process, it may be easier to do everything on a single process. You can force all Ray functions to occur on a single process with local_mode by calling the following before tune.run.

ray.init(local_mode=True)

Note that some behavior such as writing to files by depending on the current working directory in a Trainable and setting global process variables may not work as expected. Local mode with multiple configuration evaluations will interleave computation, so it is most naturally used when running a single configuration evaluation.

Further Questions or Issues?

You can post questions or issues or feedback through the following channels:

  1. ray-dev@googlegroups.com: For discussions about development or any general questions and feedback.

  2. StackOverflow: For questions about how to use Ray.

  3. GitHub Issues: For bug reports and feature requests.