Contents

Ray Train User Guide

Tip

Get in touch with us if you’re using or considering using Ray Train!

Ray Train provides solutions for training machine learning models in a distributed manner on Ray. As of Ray 1.8, support for Deep Learning is available in ray.train (formerly Ray SGD). For other model types, distributed training support is available through other libraries:

In this guide, we cover examples for the following use cases:

Backends

Ray Train provides a thin API around different backend frameworks for distributed deep learning. At the moment, Ray Train allows you to perform training with:

  • PyTorch: Ray Train initializes your distributed process group, allowing you to run your DistributedDataParallel training script. See PyTorch Distributed Overview for more information.

  • TensorFlow: Ray Train configures TF_CONFIG for you, allowing you to run your MultiWorkerMirroredStrategy training script. See Distributed training with TensorFlow for more information.

  • Horovod: Ray Train configures the Horovod environment and Rendezvous server for you, allowing you to run your DistributedOptimizer training script. See Horovod documentation for more information.

Porting code to Ray Train

The following instructions assume you have a training function that can already be run on a single worker for one of the supported backend frameworks.

Update training function

First, you’ll want to update your training function to support distributed training.

Ray Train will set up your distributed process group for you and also provides utility methods to automatically prepare your model and data for distributed training.

First, use the prepare_model function to automatically move your model to the right device and wrap it in DistributedDataParallel

import torch
from torch.nn.parallel import DistributedDataParallel
+from ray import train


def train_func():
-   device = torch.device(f"cuda:{train.local_rank()}" if
-         torch.cuda.is_available() else "cpu")
-   torch.cuda.set_device(device)

    # Create model.
    model = NeuralNetwork()

-   model = model.to(device)
-   model = DistributedDataParallel(model,
-       device_ids=[train.local_rank()] if torch.cuda.is_available() else None)

+   model = train.torch.prepare_model(model)

    ...

Then, use the prepare_data_loader function to automatically add a DistributedSampler to your DataLoader and move the batches to the right device.

import torch
from torch.utils.data import DataLoader, DistributedSampler
+from ray import train


def train_func():
-   device = torch.device(f"cuda:{train.local_rank()}" if
-          torch.cuda.is_available() else "cpu")
-   torch.cuda.set_device(device)

    ...

-   data_loader = DataLoader(my_dataset, sampler=DistributedSampler(dataset))

+   data_loader = DataLoader(my_dataset)
+   data_loader = train.torch.prepare_data_loader(data_loader)

    for X, y in data_loader:
-       X = X.to_device(device)
-       y = y.to_device(device)

Create Ray Train Trainer

The Trainer is the primary Ray Train class that is used to manage state and execute training. You can create a simple Trainer for the backend of choice with one of the following:

from ray.train import Trainer
trainer = Trainer(backend="torch", num_workers=2)

To customize the backend setup, you can replace the string argument with a Backend Configurations object.

from ray.train import Trainer
from ray.train.torch import TorchConfig

trainer = Trainer(backend=TorchConfig(...), num_workers=2)

For more configurability, please reference the Trainer API.

Run training function

With a distributed training function and a Ray Train Trainer, you are now ready to start training!

trainer.start() # set up resources
trainer.run(train_func)
trainer.shutdown() # clean up resources

Configuring Training

With Ray Train, you can execute a training function (train_func) in a distributed manner by calling trainer.run(train_func). To pass arguments into the training function, you can expose a single config dictionary parameter:

-def train_func():
+def train_func(config):

Then, you can pass in the config dictionary as an argument to Trainer.run:

-trainer.run(train_func)
+config = {} # This should be populated.
+trainer.run(train_func, config=config)

Putting this all together, you can run your training function with different configurations. As an example:

from ray.train import Trainer

def train_func(config):
    results = []
    for i in range(config["num_epochs"]):
        results.append(i)
    return results

trainer = Trainer(backend="torch", num_workers=2)
trainer.start()
print(trainer.run(train_func, config={"num_epochs": 2}))
# [[0, 1], [0, 1]]
print(trainer.run(train_func, config={"num_epochs": 5}))
# [[0, 1, 2, 3, 4], [0, 1, 2, 3, 4]]
trainer.shutdown()

A primary use-case for config is to try different hyperparameters. To perform hyperparameter tuning with Ray Train, please refer to the Ray Tune integration.

Log Directory Structure

Each Trainer will have a local directory created for logs, and each call to Trainer.run will create its own sub-directory of logs.

By default, the logdir will be created at ~/ray_results/train_<datestring>. This can be overridden in the Trainer constructor to an absolute path or a path relative to ~/ray_results.

Log directories are exposed through the following attributes:

Attribute

Example

trainer.logdir

/home/ray_results/train_2021-09-01_12-00-00

trainer.latest_run_dir

/home/ray_results/train_2021-09-01_12-00-00/run_001

Logs will be written by:

  1. Logging Callbacks

  2. Checkpoints

Logging, Monitoring, and Callbacks

Reporting intermediate results

Ray Train provides a train.report(**kwargs) API for reporting intermediate results from the training function (run on distributed workers) up to the Trainer (where your python script is executed).

Using Trainer.run, these results can be processed through Callbacks with a handle_result method defined.

The primary use-case for reporting is for metrics (accuracy, loss, etc.) at the end of each training epoch.

def train_func():
    ...
    for i in range(num_epochs):
        results = model.train(...)
        train.report(results)
    return model

For custom handling, the lower-level Trainer.run_iterator API produces a TrainingIterator which will iterate over the reported results.

Autofilled metrics

In addition to user defined metrics, a few fields are automatically populated:

# Unix epoch time in seconds when the data is reported.
_timestamp
# Time in seconds between iterations.
_time_this_iter_s
# The iteration ID, where each iteration is defined by one call to train.report().
# This is a 1-indexed incrementing integer ID.
_training_iteration

For debugging purposes, a more extensive set of metrics can be included in any run by setting the TRAIN_RESULT_ENABLE_DETAILED_AUTOFILLED_METRICS environment variable to 1.

# The local date string when the data is reported.
_date
# The worker hostname (platform.node()).
_hostname
# The worker IP address.
_node_ip
# The worker process ID (os.getpid()).
_pid
# The cumulative training time of all iterations so far.
_time_total_s

Callbacks

You may want to plug in your training code with your favorite experiment management framework. Ray Train provides an interface to fetch intermediate results and callbacks to process/log your intermediate results.

You can plug all of these into Ray Train with the following interface:

from ray import train
from ray.train import Trainer, TrainingCallback
from typing import List, Dict

class PrintingCallback(TrainingCallback):
    def handle_result(self, results: List[Dict], **info):
        print(results)

def train_func():
    for i in range(3):
        train.report(epoch=i)

trainer = Trainer(backend="torch", num_workers=2)
trainer.start()
result = trainer.run(
    train_func,
    callbacks=[PrintingCallback()]
)
# [{'epoch': 0, '_timestamp': 1630471763, '_time_this_iter_s': 0.0020279884338378906, '_training_iteration': 1}, {'epoch': 0, '_timestamp': 1630471763, '_time_this_iter_s': 0.0014922618865966797, '_training_iteration': 1}]
# [{'epoch': 1, '_timestamp': 1630471763, '_time_this_iter_s': 0.0008401870727539062, '_training_iteration': 2}, {'epoch': 1, '_timestamp': 1630471763, '_time_this_iter_s': 0.0007486343383789062, '_training_iteration': 2}]
# [{'epoch': 2, '_timestamp': 1630471763, '_time_this_iter_s': 0.0014500617980957031, '_training_iteration': 3}, {'epoch': 2, '_timestamp': 1630471763, '_time_this_iter_s': 0.0015292167663574219, '_training_iteration': 3}]
trainer.shutdown()

Logging Callbacks

The following TrainingCallbacks are available and will write to a file within the log directory of each training run.

  1. JsonLoggerCallback

  2. TBXLoggerCallback

Custom Callbacks

If the provided callbacks do not cover your desired integrations or use-cases, you may always implement a custom callback by subclassing TrainingCallback. If the callback is general enough, please feel welcome to add it to the ray repository.

A simple example for creating a callback that will print out results:

from ray.train import TrainingCallback

class PrintingCallback(TrainingCallback):
    def handle_result(self, results: List[Dict], **info):
        print(results)

Example: PyTorch Distributed metrics

In real applications, you may want to calculate optimization metrics besides accuracy and loss: recall, precision, Fbeta, etc.

Ray Train natively supports TorchMetrics, which provides a collection of machine learning metrics for distributed, scalable Pytorch models.

Here is an example:

from ray import train
from train.train import Trainer, TrainingCallback
from typing import List, Dict

import torch
import torchmetrics

class PrintingCallback(TrainingCallback):
    def handle_result(self, results: List[Dict], **info):
        print(results)

def train_func(config):
    preds = torch.randn(10, 5).softmax(dim=-1)
    target = torch.randint(5, (10,))
    accuracy = torchmetrics.functional.accuracy(preds, target).item()
    train.report(accuracy=accuracy)

trainer = Trainer(backend="torch", num_workers=2)
trainer.start()
result = trainer.run(
    train_func,
    callbacks=[PrintingCallback()]
)
# [{'accuracy': 0.20000000298023224, '_timestamp': 1630716913, '_time_this_iter_s': 0.0039408206939697266, '_training_iteration': 1},
#  {'accuracy': 0.10000000149011612, '_timestamp': 1630716913, '_time_this_iter_s': 0.0030548572540283203, '_training_iteration': 1}]
trainer.shutdown()

Checkpointing

Ray Train provides a way to save state during the training process. This is useful for:

  1. Integration with Ray Tune to use certain Ray Tune schedulers.

  2. Running a long-running training job on a cluster of pre-emptible machines/pods.

  3. Persisting trained model state to later use for serving/inference.

  4. In general, storing any model artifacts.

Saving checkpoints

Checkpoints can be saved by calling train.save_checkpoint(**kwargs) in the training function. This will cause the checkpoint state from the distributed workers to be saved on the Trainer (where your python script is executed).

The latest saved checkpoint can be accessed through the Trainer’s latest_checkpoint attribute.

As a trivial example, let’s create a “model” which is an integer value that is equivalent to the sum of the number of epochs.

from ray import train
from ray.train import Trainer

def train_func(config):
    model = 0 # This should be replaced with a real model.
    for epoch in range(config["num_epochs"]):
        model += epoch # This should be replaced with training logic.
        train.save_checkpoint(epoch=epoch, model=model)

trainer = Trainer(backend="torch", num_workers=2)
trainer.start()

trainer.run(train_func, config={"num_epochs": 1})
print(trainer.latest_checkpoint)
# {'epoch': 0, 'model': 0}

trainer.run(train_func, config={"num_epochs": 5})
print(trainer.latest_checkpoint)
# {'epoch': 4, 'model': 10}

trainer.run(train_func, config={"num_epochs": 3})
print(trainer.latest_checkpoint)
# {'epoch': 2, 'model': 3}

trainer.shutdown()

By default, checkpoints will be persisted to local disk in the log directory of each run.

print(trainer.latest_checkpoint_dir)
# /home/ray_results/train_2021-09-01_12-00-00/run_001/checkpoints

# By default, the "best" checkpoint path will refer to the most recent one.
# This can be configured by defining a CheckpointStrategy.
print(trainer.best_checkpoint_path)
# /home/ray_results/train_2021-09-01_12-00-00/run_001/checkpoints/checkpoint_000005

Note

Persisting checkpoints to durable storage (e.g. S3) is not yet supported.

Configuring checkpoints

For more configurability of checkpointing behavior (specifically saving checkpoints to disk), a CheckpointStrategy can be passed into Trainer.run.

As an example, to completely disable writing checkpoints to disk:

from ray import train
from ray.train import CheckpointStrategy, Trainer

def train_func():
    for epoch in range(3):
        train.save_checkpoint(epoch=epoch)

checkpoint_strategy = CheckpointStrategy(num_to_keep=0)

trainer = Trainer(backend="torch", num_workers=2)
trainer.start()
trainer.run(train_func, checkpoint_strategy=checkpoint_strategy)
trainer.shutdown()

You may also config CheckpointStrategy to keep the “N best” checkpoints persisted to disk. The following example shows how you could keep the 2 checkpoints with the lowest “loss” value:

from ray import train
from ray.train import CheckpointStrategy, Trainer


def train_func():
    # first checkpoint
    train.save_checkpoint(loss=2)
    # second checkpoint
    train.save_checkpoint(loss=4)
    # third checkpoint
    train.save_checkpoint(loss=1)
    # fourth checkpoint
    train.save_checkpoint(loss=3)

# Keep the 2 checkpoints with the smallest "loss" value.
checkpoint_strategy = CheckpointStrategy(num_to_keep=2,
                                         checkpoint_score_attribute="loss",
                                         checkpoint_score_order="min")

trainer = Trainer(backend="torch", num_workers=2)
trainer.start()
trainer.run(train_func, checkpoint_strategy=checkpoint_strategy)
print(trainer.best_checkpoint_path)
# /home/ray_results/train_2021-09-01_12-00-00/run_001/checkpoints/checkpoint_000003
print(trainer.latest_checkpoint_dir)
# /home/ray_results/train_2021-09-01_12-00-00/run_001/checkpoints
print([checkpoint_path for checkpoint_path in trainer.latest_checkpoint_dir.iterdir()])
# [PosixPath('/home/ray_results/train_2021-09-01_12-00-00/run_001/checkpoints/checkpoint_000003'),
# PosixPath('/home/ray_results/train_2021-09-01_12-00-00/run_001/checkpoints/checkpoint_000001')]
trainer.shutdown()

Loading checkpoints

Checkpoints can be loaded into the training function in 2 steps:

  1. From the training function, train.load_checkpoint() can be used to access the most recently saved checkpoint. This is useful to continue training even if there’s a worker failure.

  2. The checkpoint to start training with can be bootstrapped by passing in a checkpoint to trainer.run().

from ray import train
from ray.train import Trainer

def train_func(config):
    checkpoint = train.load_checkpoint() or {}
    # This should be replaced with a real model.
    model = checkpoint.get("model", 0)
    start_epoch = checkpoint.get("epoch", -1) + 1
    for epoch in range(start_epoch, config["num_epochs"]):
        model += epoch
        train.save_checkpoint(epoch=epoch, model=model)

trainer = Trainer(backend="torch", num_workers=2)
trainer.start()
trainer.run(train_func, config={"num_epochs": 5},
            checkpoint={"epoch": 2, "model": 3})
trainer.shutdown()

print(trainer.latest_checkpoint)
# {'epoch': 4, 'model': 10}

Fault Tolerance & Elastic Training

Ray Train has built-in fault tolerance to recover from worker failures (i.e. RayActorErrors). When a failure is detected, the workers will be shut down and new workers will be added in. The training function will be restarted, but progress from the previous execution can be resumed through checkpointing.

Warning

In order to retain progress when recovery, your training function must implement logic for both saving and loading checkpoints.

Each instance of recovery from a worker failure is considered a retry. The number of retries is configurable through the max_retries argument of the Trainer constructor.

Note

Elastic Training is not yet supported.

Distributed Data Ingest (Ray Datasets)

Ray Train provides native support for Ray Datasets to support the following use cases:

  1. Large Datasets: With Ray Datasets, you can easily work with datasets that are too big to fit on a single node. Ray Datasets will distribute the dataset across the Ray Cluster and allow you to perform dataset operations (map, filter, etc.) on the distributed dataset.

  2. Automatic locality-aware sharding: If provided a Ray Dataset, Ray Train will automatically shard the dataset and assign each shard to a training worker while minimizing cross-node data transfer. Unlike with standard Torch or TensorFlow datasets, each training worker will only load its assigned shard into memory rather than the entire Dataset.

  3. Pipelined Execution: Ray Datasets also supports pipelining, meaning that data processing operations can be run concurrently with training. Training is no longer blocked on expensive data processing operations (such as global shuffling) and this minimizes the amount of time your GPUs are idle. See Dataset Pipelines for more information.

To get started, pass in a Ray Dataset (or multiple) into Trainer.run. Underneath the hood, Ray Train will automatically shard the given dataset.

Warning

If you are doing distributed training with TensorFlow, you will need to disable TensorFlow’s built-in autosharding as the data on each worker is already sharded.

def train_func():
    ...
    tf_dataset = ray.train.get_dataset_shard().to_tf()
    options = tf.data.Options()
    options.experimental_distribute.auto_shard_policy = \
        tf.data.experimental.AutoShardPolicy.OFF
    tf_dataset = tf_dataset.with_options(options)

Simple Dataset Example

def train_func(config):
    # Create your model here.
    model = NeuralNetwork()

    batch_size = config["worker_batch_size"]

    train_data_shard = ray.train.get_dataset_shard("train")
    train_torch_dataset = train_data_shard.to_torch(label_column="label",
                                              batch_size=batch_size)

    validation_data_shard = ray.train.get_dataset_shard("validation")
    validation_torch_dataset = validation_data_shard.to_torch(label_column="label",
                                                              batch_size=batch_size)

    for epoch in config["num_epochs"]:
        for X, y in train_torch_dataset:
            model.train()
            output = model(X)
            # Train on one batch.
        for X, y in validation_torch_dataset:
            model.eval()
            output = model(X)
            # Validate one batch.
    return model

trainer = Trainer(num_workers=8, backend="torch")
dataset = ray.data.read_csv("...")

# Random split dataset into 80% training data and 20% validation data.
split_index = int(dataset.count() * 0.8)
train_dataset, validation_dataset = \
    dataset.random_shuffle().split_at_indices([split_index])

result = trainer.run(
    train_func,
    config={"worker_batch_size": 64, "num_epochs": 2},
    dataset={
        "train": train_dataset,
        "validation": validation_dataset
    })

Pipelined Execution

For pipelined execution, you just need to convert your Dataset into a DatasetPipeline. All operations after this conversion will be executed in a pipelined fashion.

See Dataset Pipelines for more semantics on pipelining.

Example: Per-Epoch Shuffle Pipeline

A common use case is to have a training pipeline that globally shuffles the dataset before every epoch.

This is very simple to do with Ray Datasets + Ray Train.

def train_func():
    # This is a dummy train function just iterating over the dataset.
    # You should replace this with your training logic.
    dataset_pipeline_shard = ray.train.get_dataset_shard()
    # Infinitely long iterator of randomly shuffled dataset shards.
    dataset_iterator = train_dataset_pipeline_shard.iter_epochs()
    for _ in range(config["num_epochs"]):
        # Single randomly shuffled dataset shard.
        train_dataset = next(dataset_iterator)
        # Convert shard to native Torch Dataset.
        train_torch_dataset = train_dataset.to_torch(label_column="label",
                                                     batch_size=batch_size)
        # Train on your Torch Dataset here!

# Create a pipeline that loops over its source dataset indefinitely,
# with each repeat of the dataset randomly shuffled.
dataset_pipeline: DatasetPipeline = ray.data \
    .read_parquet(...) \
    .repeat() \
    .random_shuffle_each_window()

# Pass in the pipeline to the Trainer.
# The Trainer will automatically split the DatasetPipeline for you.
trainer = Trainer(num_workers=8, backend="torch")
result = trainer.run(
    train_func,
    config={"worker_batch_size": 64, "num_epochs": 2},
    dataset=dataset_pipeline)

You can easily set the working set size for the global shuffle by specifying the window size of the DatasetPipeline.

# Create a pipeline that loops over its source dataset indefinitely.
pipe: DatasetPipeline = ray.data \
    .read_parquet(...) \
    .window(blocks_per_window=10) \
    .repeat() \
    .random_shuffle_each_window()

See Example: Per-Epoch Shuffle Pipeline for more info.

Hyperparameter tuning (Ray Tune)

Hyperparameter tuning with Ray Tune is natively supported with Ray Train. Specifically, you can take an existing training function and follow these steps:

Step 1: Convert to Tune Trainable

Instantiate your Trainer and call trainer.to_tune_trainable, which will produce an object (“Trainable”) that will be passed to Ray Tune.

from ray import train
from ray.train import Trainer

def train_func(config):
    # In this example, nothing is expected to change over epochs,
    # and the output metric is equivalent to the input value.
    for _ in range(config["num_epochs"]):
        train.report(output=config["input"])

trainer = Trainer(backend="torch", num_workers=2)
trainable = trainer.to_tune_trainable(train_func)

Step 2: Call tune.run

Call tune.run on the created Trainable to start multiple Tune “trials”, each running a Ray Train job and each with a unique hyperparameter configuration.

from ray import tune
analysis = tune.run(trainable, config={
    "num_epochs": 2,
    "input": tune.grid_search([1, 2, 3])
})
print(analysis.get_best_config(metric="output", mode="max"))
# {'num_epochs': 2, 'input': 3}

A couple caveats:

  • Tune will ignore the return value of train_func. To save your best trained model, you will need to use the train.save_checkpoint API.

  • You should not call tune.report or tune.checkpoint_dir in your training function. Functional parity is achieved through train.report, train.save_checkpoint, and train.load_checkpoint. This allows you to go from Ray Train to Ray Train+RayTune without changing any code in the training function.

from ray import train, tune
from ray.train import Trainer

def train_func(config):
    # In this example, nothing is expected to change over epochs,
    # and the output metric is equivalent to the input value.
    for _ in range(config["num_epochs"]):
        train.report(output=config["input"])

trainer = Trainer(backend="torch", num_workers=2)
trainable = trainer.to_tune_trainable(train_func)
analysis = tune.run(trainable, config={
    "num_epochs": 2,
    "input": tune.grid_search([1, 2, 3])
})
print(analysis.get_best_config(metric="output", mode="max"))
# {'num_epochs': 2, 'input': 3}

Backwards Compatibility with Ray SGD

If you are currently using RaySGD, you can migrate to Ray Train by following: Migrating from Ray SGD to Ray Train.

Ray Train: Distributed Deep Learning Ray Train Examples