Each hyperparameter configuration evaluation is called a *trial*, and multiple trials are run in parallel. Configurations are either generated by Tune or drawn from a user-specified **search algorithm**. The trials are scheduled and managed by a **trial scheduler**.
More information about Tune's `search algorithms can be found here <tune-searchalg.html>`__. More information about Tune's `trial schedulers can be found here <tune-schedulers.html>`__. You can check out our `examples page <tune-examples.html>`__ for more code examples.
The class-based API will require users to subclass ``ray.tune.Trainable``. The Trainable interface `can be found here <tune-package-ref.html#ray.tune.Trainable>`__.
User-defined functions will need to have following signature and call ``tune.track.log``, which will allow you to report metrics used for scheduling, search, or early stopping:
Tune will run this function on a separate thread in a Ray actor process. Note that this API is not checkpointable, since the thread will never return control back to its caller. ``tune.track`` documentation can be `found here <tune-package-ref.html#module-ray.tune.track>`__.
If you have a lambda function that you want to train, you will need to first register the function: ``tune.register_trainable("lambda_id", lambda x: ...)``. You can then use ``lambda_id`` in place of ``my_trainable``.
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.
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``:
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 <tune-usage#resource-allocation-using-gpus>`_, which provides more details about GPU usage and trials that are distributed:
..code-block:: python
# If you have 4 CPUs on your machine and 1 GPU, this will run 1 trial at a time.
You can use ``tune.grid_search`` to specify an axis of a grid search. By default, Tune also supports sampling parameters from user-specified lambda functions, which can be used independently or in combination with grid search.
If you specify an explicit Search Algorithm such as any SuggestionAlgorithm, you may not be able to specify lambdas or grid search with this interface, as the search algorithm may require a different search space declaration.
Use ``tune.sample_from(<func>)`` to sample a value for a hyperparameter. The ``func`` should take in a ``spec`` object, which has a ``config`` namespace from which you can access other hyperparameters. This is useful for conditional distributions:
Tune provides a couple helper functions for common parameter distributions, wrapping numpy random utilities such as ``np.random.uniform``, ``np.random.choice``, and ``np.random.randn``. See the `Package Reference <tune-package-ref.html#ray.tune.uniform>`_ for more details.
The following shows grid search over two nested parameters combined with random sampling from two lambda functions, generating 9 different trials. Note that the value of ``beta`` depends on the value of ``alpha``, which is represented by referencing ``spec.config.alpha`` in the lambda function. This lets you specify conditional parameter distributions.
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.
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``.
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 <https://github.com/ray-project/ray/blob/master/python/ray/tune/examples/tune_mnist_keras.py>`__.
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``.
The ``Trainable`` also provides the ``default_resource_requests`` interface to automatically declare the ``resources_per_trial`` based on the given configuration.
To enable checkpointing, you must implement a `Trainable class <tune-usage.html#trainable-api>`__ (Trainable functions are not checkpointable, since they never return control back to their caller). The easiest way to do this is to subclass the pre-defined ``Trainable`` class and implement ``_save``, and ``_restore`` abstract methods, as seen in `this example <https://github.com/ray-project/ray/blob/master/python/ray/tune/examples/hyperband_example.py>`__.
For PyTorch model training, this would look something like this `PyTorch example <https://github.com/ray-project/ray/blob/master/python/ray/tune/examples/mnist_pytorch_trainable.py>`__:
Checkpoints will be saved by training iteration to ``local_dir/exp_name/trial_name/checkpoint_<iter>``. You can restore a single trial checkpoint by using ``tune.run(restore=<checkpoint_dir>)``.
Tune also generates temporary checkpoints for pausing and switching between trials. For this purpose, it is important not to depend on absolute paths in the implementation of ``save``. See the below reference:
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:
**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.:
**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``:
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:
..code-block:: python
# 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
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 <tune-distributed.html#example-for-using-spot-instances-aws>`_.
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.
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.
This feature is still experimental, so any provided Trial Scheduler or Search Algorithm will not be preserved. Only ``FIFOScheduler`` and ``BasicVariantGenerator`` will be supported.
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``.
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.
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).
You can use this to stop all trials after the criteria is fulfilled by any individual trial:
..code-block:: python
class Stopper:
def __init__(self):
self.should_stop = False
def stop(self, trial_id, result):
if not self.should_stop and result['foo'] > 10:
self.should_stop = True
return self.should_stop
stopper = Stopper()
tune.run(my_trainable, stop=stopper.stop)
Note that in the above example all trials will not stop immediately, but will do so once their current iterations are complete.
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.
To visualize learning in tensorboard, install TensorFlow:
..code-block:: bash
$ pip install tensorflow
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>``:
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:
These loggers will be called along with the default Tune loggers. All loggers must inherit the `Logger interface <tune-package-ref.html#ray.tune.logger.Logger>`__. Tune enables default loggers for Tensorboard, CSV, and JSON formats. You can also check out `logger.py <https://github.com/ray-project/ray/blob/master/python/ray/tune/logger.py>`__ for implementation details. An example can be found in `logging_example.py <https://github.com/ray-project/ray/blob/master/python/ray/tune/examples/logging_example.py>`__.
..warning:: If you run into issues for TensorBoard logging, consider using the TensorBoardX Logger (``from ray.tune.logger import TBXLogger``)
TBXLogger (TensorboardX)
~~~~~~~~~~~~~~~~~~~~~~~~
Tune provides a logger using `TensorBoardX <https://github.com/lanpa/tensorboardX>`_. You can install tensorboardX via ``pip install tensorboardX``. This logger automatically outputs loggers similar to the default TensorFlow logging format but is nice if you are undergoing a TF1 to TF2 transition. By default, it will log any scalar value provided via the result dictionary along with HParams information.
Tune also provides a default logger for `MLFlow <https://mlflow.org>`_. You can install MLFlow via ``pip install mlflow``. An example can be found `mlflow_example.py <https://github.com/ray-project/ray/blob/master/python/ray/tune/examples/mlflow_example.py>`__. Note that this currently does not include artifact logging support. For this, you can use the native MLFlow APIs inside your Trainable definition.
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 <autoscaling.html>`__.
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:
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 <https://github.com/ray-project/ray/blob/master/python/ray/tune/syncer.py>`__ for implementation details.
You can interact with an ongoing experiment with the Tune Client API. The Tune Client API is organized around REST, which includes resource-oriented URLs, accepts form-encoded requests, returns JSON-encoded responses, and uses standard HTTP protocol.
The easiest way to use the Tune Client API is with the built-in TuneClient. To use TuneClient, verify that you have the ``requests`` library installed:
..code-block:: bash
$ pip install requests
Then, on the client side, you can use the following class. If on a cluster, you may want to forward this port (e.g. ``ssh -L <local_port>:localhost:<remote_port> <address>``) so that you can use the Client on your local machine.
For an example notebook for using the Client API, see the `Client API Example <https://github.com/ray-project/ray/tree/master/python/ray/tune/TuneClient.ipynb>`__.
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``.
..code-block:: python
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.
``tune`` has an easy-to-use command line interface (CLI) to manage and monitor your experiments on Ray. To do this, verify that you have the ``tabulate`` library installed:
-``tune list-trials``: List tabular information about trials within an experiment. Empty columns will be dropped by default. Add the ``--sort`` flag to sort the output by specific columns. Add the ``--filter`` flag to filter the output in the format ``"<column> <operator> <value>"``. Add the ``--output`` flag to write the trial information to a specific file (CSV or Pickle). Add the ``--columns`` and ``--result-columns`` flags to select specific columns to display.
