Distributed training with TensorFlow

View on TensorFlow.org Run in Google Colab View source on GitHub Download notebook

Overview

tf.distribute.Strategy is a TensorFlow API to distribute training across multiple GPUs, multiple machines or TPUs. Using this API, you can distribute your existing models and training code with minimal code changes.

tf.distribute.Strategy has been designed with these key goals in mind:

  • Easy to use and support multiple user segments, including researchers, ML engineers, etc.
  • Provide good performance out of the box.
  • Easy switching between strategies.

tf.distribute.Strategy can be used with a high-level API like Keras, and can also be used to distribute custom training loops (and, in general, any computation using TensorFlow).

In TensorFlow 2.x, you can execute your programs eagerly, or in a graph using tf.function. tf.distribute.Strategy intends to support both these modes of execution, but works best with tf.function. Eager mode is only recommended for debugging purpose and not supported for TPUStrategy. Although training is the focus of this guide, this API can also be used for distributing evaluation and prediction on different platforms.

You can use tf.distribute.Strategy with very few changes to your code, because the underlying components of TensorFlow have been changed to become strategy-aware. This includes variables, layers, models, optimizers, metrics, summaries, and checkpoints.

In this guide, you will learn about various types of strategies and how you can use them in different situations. To learn how to debug performance issues, see the Optimize TensorFlow GPU performance guide.

# Import TensorFlow
import tensorflow as tf
2021-08-02 22:18:25.188129: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudart.so.11.0

Types of strategies

tf.distribute.Strategy intends to cover a number of use cases along different axes. Some of these combinations are currently supported and others will be added in the future. Some of these axes are:

  • Synchronous vs asynchronous training: These are two common ways of distributing training with data parallelism. In sync training, all workers train over different slices of input data in sync, and aggregating gradients at each step. In async training, all workers are independently training over the input data and updating variables asynchronously. Typically sync training is supported via all-reduce and async through parameter server architecture.
  • Hardware platform: You may want to scale your training onto multiple GPUs on one machine, or multiple machines in a network (with 0 or more GPUs each), or on Cloud TPUs.

In order to support these use cases, there are 6 strategies available. The next section explains which of these are supported in which scenarios in TensorFlow. Here is a quick overview:

Training API MirroredStrategy TPUStrategy MultiWorkerMirroredStrategy CentralStorageStrategy ParameterServerStrategy
Keras API Supported Supported Supported Experimental support Supported planned post 2.4
Custom training loop Supported Supported Supported Experimental support Experimental support
Estimator API Limited Support Not supported Limited Support Limited Support Limited Support

MirroredStrategy

tf.distribute.MirroredStrategy supports synchronous distributed training on multiple GPUs on one machine. It creates one replica per GPU device. Each variable in the model is mirrored across all the replicas. Together, these variables form a single conceptual variable called MirroredVariable. These variables are kept in sync with each other by applying identical updates.

Efficient all-reduce algorithms are used to communicate the variable updates across the devices. All-reduce aggregates tensors across all the devices by adding them up, and makes them available on each device. It’s a fused algorithm that is very efficient and can reduce the overhead of synchronization significantly. There are many all-reduce algorithms and implementations available, depending on the type of communication available between devices. By default, it uses the the NVIDIA Collective Communication Library (NCCL) as the all-reduce implementation. You can choose from a few other options or write your own.

Here is the simplest way of creating MirroredStrategy:

mirrored_strategy = tf.distribute.MirroredStrategy()
WARNING:tensorflow:Collective ops is not configured at program startup. Some performance features may not be enabled.
2021-08-02 22:18:26.445902: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcuda.so.1
2021-08-02 22:18:27.104337: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-02 22:18:27.105253: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1733] Found device 0 with properties: 
pciBusID: 0000:00:05.0 name: Tesla V100-SXM2-16GB computeCapability: 7.0
coreClock: 1.53GHz coreCount: 80 deviceMemorySize: 15.78GiB deviceMemoryBandwidth: 836.37GiB/s
2021-08-02 22:18:27.105285: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudart.so.11.0
2021-08-02 22:18:27.109261: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcublas.so.11
2021-08-02 22:18:27.109352: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcublasLt.so.11
2021-08-02 22:18:27.110507: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcufft.so.10
2021-08-02 22:18:27.110851: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcurand.so.10
2021-08-02 22:18:27.111886: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcusolver.so.11
2021-08-02 22:18:27.112752: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcusparse.so.11
2021-08-02 22:18:27.112924: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudnn.so.8
2021-08-02 22:18:27.113017: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-02 22:18:27.113930: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-02 22:18:27.114804: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1871] Adding visible gpu devices: 0
2021-08-02 22:18:27.115378: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 AVX512F FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2021-08-02 22:18:27.116008: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-02 22:18:27.116993: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1733] Found device 0 with properties: 
pciBusID: 0000:00:05.0 name: Tesla V100-SXM2-16GB computeCapability: 7.0
coreClock: 1.53GHz coreCount: 80 deviceMemorySize: 15.78GiB deviceMemoryBandwidth: 836.37GiB/s
2021-08-02 22:18:27.117116: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-02 22:18:27.118046: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-02 22:18:27.119015: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1871] Adding visible gpu devices: 0
2021-08-02 22:18:27.119068: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudart.so.11.0
2021-08-02 22:18:27.737604: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1258] Device interconnect StreamExecutor with strength 1 edge matrix:
2021-08-02 22:18:27.737637: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1264]      0 
2021-08-02 22:18:27.737645: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1277] 0:   N 
2021-08-02 22:18:27.737840: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-02 22:18:27.738963: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-02 22:18:27.739906: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-02 22:18:27.740808: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1418] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 14646 MB memory) -> physical GPU (device: 0, name: Tesla V100-SXM2-16GB, pci bus id: 0000:00:05.0, compute capability: 7.0)
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)

This will create a MirroredStrategy instance, which will use all the GPUs that are visible to TensorFlow, and NCCL—as the cross-device communication.

If you wish to use only some of the GPUs on your machine, you can do so like this:

mirrored_strategy = tf.distribute.MirroredStrategy(devices=["/gpu:0", "/gpu:1"])
WARNING:tensorflow:Some requested devices in `tf.distribute.Strategy` are not visible to TensorFlow: /job:localhost/replica:0/task:0/device:GPU:0,/job:localhost/replica:0/task:0/device:GPU:1
WARNING:tensorflow:Collective ops is not configured at program startup. Some performance features may not be enabled.
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1')

If you wish to override the cross device communication, you can do so using the cross_device_ops argument by supplying an instance of tf.distribute.CrossDeviceOps. Currently, tf.distribute.HierarchicalCopyAllReduce and tf.distribute.ReductionToOneDevice are two options other than tf.distribute.NcclAllReduce, which is the default.

mirrored_strategy = tf.distribute.MirroredStrategy(
    cross_device_ops=tf.distribute.HierarchicalCopyAllReduce())
WARNING:tensorflow:Collective ops is not configured at program startup. Some performance features may not be enabled.
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)

TPUStrategy

tf.distribute.TPUStrategy lets you run your TensorFlow training on Tensor Processing Units (TPUs). TPUs are Google's specialized ASICs designed to dramatically accelerate machine learning workloads. They are available on Google Colab, the TPU Research Cloud, and Cloud TPU.

In terms of distributed training architecture, TPUStrategy is the same MirroredStrategy—it implements synchronous distributed training. TPUs provide their own implementation of efficient all-reduce and other collective operations across multiple TPU cores, which are used in TPUStrategy.

Here is how you would instantiate TPUStrategy:

cluster_resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
    tpu=tpu_address)
tf.config.experimental_connect_to_cluster(cluster_resolver)
tf.tpu.experimental.initialize_tpu_system(cluster_resolver)
tpu_strategy = tf.distribute.TPUStrategy(cluster_resolver)

The TPUClusterResolver instance helps locate the TPUs. In Colab, you don't need to specify any arguments to it.

If you want to use this for Cloud TPUs:

  • You must specify the name of your TPU resource in the tpu argument.
  • You must initialize the tpu system explicitly at the start of the program. This is required before TPUs can be used for computation. Initializing the tpu system also wipes out the TPU memory, so it's important to complete this step first in order to avoid losing state.

MultiWorkerMirroredStrategy

tf.distribute.MultiWorkerMirroredStrategy is very similar to MirroredStrategy. It implements synchronous distributed training across multiple workers, each with potentially multiple GPUs. Similar to tf.distribute.MirroredStrategy, it creates copies of all variables in the model on each device across all workers.

Here is the simplest way of creating MultiWorkerMirroredStrategy:

strategy = tf.distribute.MultiWorkerMirroredStrategy()
WARNING:tensorflow:Collective ops is not configured at program startup. Some performance features may not be enabled.
INFO:tensorflow:Single-worker MultiWorkerMirroredStrategy with local_devices = ('/device:GPU:0',), communication = CommunicationImplementation.AUTO

MultiWorkerMirroredStrategy has two implementations for cross-device communications. CommunicationImplementation.RING is RPC-based and supports both CPUs and GPUs. CommunicationImplementation.NCCL uses NCCL and provides state-of-art performance on GPUs but it doesn't support CPUs. CollectiveCommunication.AUTO defers the choice to Tensorflow. You can specify them in the following way:

communication_options = tf.distribute.experimental.CommunicationOptions(
    implementation=tf.distribute.experimental.CommunicationImplementation.NCCL)
strategy = tf.distribute.MultiWorkerMirroredStrategy(
    communication_options=communication_options)
WARNING:tensorflow:Collective ops is not configured at program startup. Some performance features may not be enabled.
INFO:tensorflow:Single-worker MultiWorkerMirroredStrategy with local_devices = ('/device:GPU:0',), communication = CommunicationImplementation.NCCL

One of the key differences to get multi worker training going, as compared to multi-GPU training, is the multi-worker setup. The TF_CONFIG environment variable is the standard way in TensorFlow to specify the cluster configuration to each worker that is part of the cluster. Learn more about setting up TF_CONFIG.

ParameterServerStrategy

Parameter server training is a common data-parallel method to scale up model training on multiple machines. A parameter server training cluster consists of workers and parameter servers. Variables are created on parameter servers and they are read and updated by workers in each step. Check out the Parameter server training tutorial for details.

In TensorFlow 2, parameter server training uses a central coordinator-based architecture via the tf.distribute.experimental.coordinator.ClusterCoordinator class.

In this implementation, the worker and parameter server tasks run tf.distribute.Servers that listen for tasks from the coordinator. The coordinator creates resources, dispatches training tasks, writes checkpoints, and deals with task failures.

In the programming running on the coordinator, you will use a ParameterServerStrategy object to define a training step and use a ClusterCoordinator to dispatch training steps to remote workers. Here is the simplest way to create them:

strategy = tf.distribute.experimental.ParameterServerStrategy(
    tf.distribute.cluster_resolver.TFConfigClusterResolver(),
    variable_partitioner=variable_partitioner)
coordinator = tf.distribute.experimental.coordinator.ClusterCoordinator(
    strategy)

In TensorFlow 1, ParameterServerStrategy is available only with an Estimator via tf.compat.v1.distribute.experimental.ParameterServerStrategy symbol.

CentralStorageStrategy

tf.distribute.experimental.CentralStorageStrategy does synchronous training as well. Variables are not mirrored, instead they are placed on the CPU and operations are replicated across all local GPUs. If there is only one GPU, all variables and operations will be placed on that GPU.

Create an instance of CentralStorageStrategy by:

central_storage_strategy = tf.distribute.experimental.CentralStorageStrategy()
INFO:tensorflow:ParameterServerStrategy (CentralStorageStrategy if you are using a single machine) with compute_devices = ['/job:localhost/replica:0/task:0/device:GPU:0'], variable_device = '/job:localhost/replica:0/task:0/device:GPU:0'

This will create a CentralStorageStrategy instance which will use all visible GPUs and CPU. Update to variables on replicas will be aggregated before being applied to variables.

Other strategies

In addition to the above strategies, there are two other strategies which might be useful for prototyping and debugging when using tf.distribute APIs.

Default Strategy

The Default Strategy is a distribution strategy which is present when no explicit distribution strategy is in scope. It implements the tf.distribute.Strategy interface but is a pass-through and provides no actual distribution. For instance, strategy.run(fn) will simply call fn. Code written using this strategy should behave exactly as code written without any strategy. You can think of it as a "no-op" strategy.

The Default Strategy is a singleton—and one cannot create more instances of it. It can be obtained using tf.distribute.get_strategy outside any explicit strategy's scope (the same API that can be used to get the current strategy inside an explicit strategy's scope).

default_strategy = tf.distribute.get_strategy()

This strategy serves two main purposes:

  • It allows writing distribution aware library code unconditionally. For example, in tf.optimizers can use tf.distribute.get_strategy and use that strategy for reducing gradients—it will always return a strategy object on which you can call the reduce API.
# In optimizer or other library code
# Get currently active strategy
strategy = tf.distribute.get_strategy()
strategy.reduce("SUM", 1., axis=None)  # reduce some values
1.0
  • Similar to library code, it can be used to write end users' programs to work with and without distribution strategy, without requiring conditional logic. Here's a sample code snippet illustrating this:
if tf.config.list_physical_devices('GPU'):
  strategy = tf.distribute.MirroredStrategy()
else:  # Use the Default Strategy
  strategy = tf.distribute.get_strategy() 

with strategy.scope():
  # Do something interesting
  print(tf.Variable(1.))
WARNING:tensorflow:Collective ops is not configured at program startup. Some performance features may not be enabled.
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)
MirroredVariable:{
  0: <tf.Variable 'Variable:0' shape=() dtype=float32, numpy=1.0>
}

OneDeviceStrategy

tf.distribute.OneDeviceStrategy is a strategy to place all variables and computation on a single specified device.

strategy = tf.distribute.OneDeviceStrategy(device="/gpu:0")

This strategy is distinct from the Default Strategy in a number of ways. In the Default Strategy, the variable placement logic remains unchanged when compared to running TensorFlow without any distribution strategy. But when using OneDeviceStrategy, all variables created in its scope are explicitly placed on the specified device. Moreover, any functions called via OneDeviceStrategy.run will also be placed on the specified device.

Input distributed through this strategy will be prefetched to the specified device. In the Default Strategy, there is no input distribution.

Similar to the Default Strategy, this strategy could also be used to test your code before switching to other strategies which actually distribute to multiple devices/machines. This will exercise the distribution strategy machinery somewhat more than default strategy, but not to the full extent as using MirroredStrategy or TPUStrategy etc. If you want code that behaves as if no strategy, then use default strategy.

So far you've seen the different strategies available and how you can instantiate them. The next few sections show the different ways in which you can use them to distribute your training.

Using tf.distribute.Strategy with tf.keras.Model.fit

tf.distribute.Strategy is integrated into tf.keras, which is TensorFlow's implementation of the Keras API specification. tf.keras is a high-level API to build and train models. By integrating into tf.keras backend, it's seamless for you to distribute your training written in the Keras training framework using Model.fit.

Here's what you need to change in your code:

  1. Create an instance of the appropriate tf.distribute.Strategy.
  2. Move the creation of Keras model, optimizer and metrics inside strategy.scope.

TensorFlow distribution strategies support all types of Keras models—Sequential, Functional, and subclassed.

Here is a snippet of code to do this for a very simple Keras model with one dense layer:

mirrored_strategy = tf.distribute.MirroredStrategy()

with mirrored_strategy.scope():
  model = tf.keras.Sequential([tf.keras.layers.Dense(1, input_shape=(1,))])

model.compile(loss='mse', optimizer='sgd')
WARNING:tensorflow:Collective ops is not configured at program startup. Some performance features may not be enabled.
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).

This example uses MirroredStrategy, so you can run this on a machine with multiple GPUs. strategy.scope() indicates to Keras which strategy to use to distribute the training. Creating models/optimizers/metrics inside this scope allows you to create distributed variables instead of regular variables. Once this is set up, you can fit your model like you would normally. MirroredStrategy takes care of replicating the model's training on the available GPUs, aggregating gradients, and more.

dataset = tf.data.Dataset.from_tensors(([1.], [1.])).repeat(100).batch(10)
model.fit(dataset, epochs=2)
model.evaluate(dataset)
Epoch 1/2
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
2021-08-02 22:18:28.409827: W tensorflow/core/grappler/optimizers/data/auto_shard.cc:695] AUTO sharding policy will apply DATA sharding policy as it failed to apply FILE sharding policy because of the following reason: Found an unshardable source dataset: name: "TensorDataset/_2"
op: "TensorDataset"
input: "Placeholder/_0"
input: "Placeholder/_1"
attr {
  key: "Toutput_types"
  value {
    list {
      type: DT_FLOAT
      type: DT_FLOAT
    }
  }
}
attr {
  key: "output_shapes"
  value {
    list {
      shape {
        dim {
          size: 1
        }
      }
      shape {
        dim {
          size: 1
        }
      }
    }
  }
}

2021-08-02 22:18:28.441658: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:176] None of the MLIR Optimization Passes are enabled (registered 2)
2021-08-02 22:18:28.442018: I tensorflow/core/platform/profile_utils/cpu_utils.cc:114] CPU Frequency: 2000194999 Hz
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
2021-08-02 22:18:29.772644: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcublas.so.11
10/10 [==============================] - 2s 2ms/step - loss: 5.1070
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
Epoch 2/2
10/10 [==============================] - 0s 2ms/step - loss: 2.2573
2021-08-02 22:18:30.195273: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcublasLt.so.11
2021-08-02 22:18:30.291277: W tensorflow/core/grappler/optimizers/data/auto_shard.cc:695] AUTO sharding policy will apply DATA sharding policy as it failed to apply FILE sharding policy because of the following reason: Found an unshardable source dataset: name: "TensorDataset/_2"
op: "TensorDataset"
input: "Placeholder/_0"
input: "Placeholder/_1"
attr {
  key: "Toutput_types"
  value {
    list {
      type: DT_FLOAT
      type: DT_FLOAT
    }
  }
}
attr {
  key: "output_shapes"
  value {
    list {
      shape {
        dim {
          size: 1
        }
      }
      shape {
        dim {
          size: 1
        }
      }
    }
  }
}
10/10 [==============================] - 1s 2ms/step - loss: 1.4018
1.4018447399139404

Here a tf.data.Dataset provides the training and eval input. You can also use NumPy arrays:

import numpy as np

inputs, targets = np.ones((100, 1)), np.ones((100, 1))
model.fit(inputs, targets, epochs=2, batch_size=10)
Epoch 1/2
2021-08-02 22:18:31.536786: W tensorflow/core/grappler/optimizers/data/auto_shard.cc:695] AUTO sharding policy will apply DATA sharding policy as it failed to apply FILE sharding policy because of the following reason: Did not find a shardable source, walked to a node which is not a dataset: name: "FlatMapDataset/_9"
op: "FlatMapDataset"
input: "PrefetchDataset/_8"
attr {
  key: "Targuments"
  value {
    list {
    }
  }
}
attr {
  key: "f"
  value {
    func {
      name: "__inference_Dataset_flat_map_slice_batch_indices_1006"
    }
  }
}
attr {
  key: "output_shapes"
  value {
    list {
      shape {
        dim {
          size: 10
        }
      }
    }
  }
}
attr {
  key: "output_types"
  value {
    list {
      type: DT_INT64
    }
  }
}
. Consider either turning off auto-sharding or switching the auto_shard_policy to DATA to shard this dataset. You can do this by creating a new `tf.data.Options()` object then setting `options.experimental_distribute.auto_shard_policy = AutoShardPolicy.DATA` before applying the options object to the dataset via `dataset.with_options(options)`.
10/10 [==============================] - 1s 2ms/step - loss: 0.9977
Epoch 2/2
10/10 [==============================] - 0s 2ms/step - loss: 0.4410
<tensorflow.python.keras.callbacks.History at 0x7fae8c430750>

In both cases—with Dataset or Numpy—each batch of the given input is divided equally among the multiple replicas. For instance, if you are using the MirroredStrategy with 2 GPUs, each batch of size 10 will get divided among the 2 GPUs, with each receiving 5 input examples in each step. Each epoch will then train faster as you add more GPUs. Typically, you would want to increase your batch size as you add more accelerators, so as to make effective use of the extra computing power. You will also need to re-tune your learning rate, depending on the model. You can use strategy.num_replicas_in_sync to get the number of replicas.

# Compute global batch size using number of replicas.
BATCH_SIZE_PER_REPLICA = 5
global_batch_size = (BATCH_SIZE_PER_REPLICA *
                     mirrored_strategy.num_replicas_in_sync)
dataset = tf.data.Dataset.from_tensors(([1.], [1.])).repeat(100)
dataset = dataset.batch(global_batch_size)

LEARNING_RATES_BY_BATCH_SIZE = {5: 0.1, 10: 0.15}
learning_rate = LEARNING_RATES_BY_BATCH_SIZE[global_batch_size]

What's supported now?

Training API MirroredStrategy TPUStrategy MultiWorkerMirroredStrategy ParameterServerStrategy CentralStorageStrategy
Keras APIs Supported Supported Experimental support Experimental support Experimental support

Examples and tutorials

Here is a list of tutorials and examples that illustrate the above integration end-to-end with Keras:

  1. Tutorial to train MNIST with MirroredStrategy.
  2. Tutorial to train MNIST using MultiWorkerMirroredStrategy.
  3. Guide on training MNIST using TPUStrategy.
  4. Tutorial for parameter server training in TensorFlow 2 with ParameterServerStrategy.
  5. TensorFlow Model Garden repository containing collections of state-of-the-art models implemented using various strategies.

Using tf.distribute.Strategy with custom training loops

As you've seen, using tf.distribute.Strategy with Keras model.fit requires changing only a couple lines of your code. With a little more effort, you can also use tf.distribute.Strategy with custom training loops.

If you need more flexibility and control over your training loops than is possible with Estimator or Keras, you can write custom training loops. For instance, when using a GAN, you may want to take a different number of generator or discriminator steps each round. Similarly, the high level frameworks are not very suitable for Reinforcement Learning training.

The tf.distribute.Strategy classes provide a core set of methods through to support custom training loops. Using these may require minor restructuring of the code initially, but once that is done, you should be able to switch between GPUs, TPUs, and multiple machines simply by changing the strategy instance.

Here, you will see a brief snippet illustrating this use case for a simple training example using the same Keras model as before.

First, create the model and optimizer inside the strategy's scope. This ensures that any variables created with the model and optimizer are mirrored variables.

with mirrored_strategy.scope():
  model = tf.keras.Sequential([tf.keras.layers.Dense(1, input_shape=(1,))])
  optimizer = tf.keras.optimizers.SGD()

Next, create the input dataset and call tf.distribute.Strategy.experimental_distribute_dataset to distribute the dataset based on the strategy.

dataset = tf.data.Dataset.from_tensors(([1.], [1.])).repeat(100).batch(
    global_batch_size)
dist_dataset = mirrored_strategy.experimental_distribute_dataset(dataset)
2021-08-02 22:18:32.763360: W tensorflow/core/grappler/optimizers/data/auto_shard.cc:695] AUTO sharding policy will apply DATA sharding policy as it failed to apply FILE sharding policy because of the following reason: Found an unshardable source dataset: name: "TensorDataset/_2"
op: "TensorDataset"
input: "Placeholder/_0"
input: "Placeholder/_1"
attr {
  key: "Toutput_types"
  value {
    list {
      type: DT_FLOAT
      type: DT_FLOAT
    }
  }
}
attr {
  key: "output_shapes"
  value {
    list {
      shape {
        dim {
          size: 1
        }
      }
      shape {
        dim {
          size: 1
        }
      }
    }
  }
}

Then, define one step of the training. Use tf.GradientTape to compute gradients and optimizer to apply those gradients to update your model's variables. To distribute this training step, put it in a function train_step and pass it to tf.distrbute.Strategy.run along with the dataset inputs you got from the dist_dataset created before:

loss_object = tf.keras.losses.BinaryCrossentropy(
  from_logits=True,
  reduction=tf.keras.losses.Reduction.NONE)

def compute_loss(labels, predictions):
  per_example_loss = loss_object(labels, predictions)
  return tf.nn.compute_average_loss(per_example_loss, global_batch_size=global_batch_size)

def train_step(inputs):
  features, labels = inputs

  with tf.GradientTape() as tape:
    predictions = model(features, training=True)
    loss = compute_loss(labels, predictions)

  gradients = tape.gradient(loss, model.trainable_variables)
  optimizer.apply_gradients(zip(gradients, model.trainable_variables))
  return loss

@tf.function
def distributed_train_step(dist_inputs):
  per_replica_losses = mirrored_strategy.run(train_step, args=(dist_inputs,))
  return mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, per_replica_losses,
                         axis=None)

A few other things to note in the code above:

  1. You used tf.nn.compute_average_loss to compute the loss. tf.nn.compute_average_loss sums the per example loss and divide the sum by the global_batch_size. This is important because later after the gradients are calculated on each replica, they are aggregated across the replicas by summing them.
  2. You also used the tf.distribute.Strategy.reduce API to aggregate the results returned by tf.distribute.Strategy.run. tf.distribute.Strategy.run returns results from each local replica in the strategy, and there are multiple ways to consume this result. You can reduce them to get an aggregated value. You can also do tf.distribute.Strategy.experimental_local_results to get the list of values contained in the result, one per local replica.
  3. When you call apply_gradients within a distribution strategy scope, its behavior is modified. Specifically, before applying gradients on each parallel instance during synchronous training, it performs a sum-over-all-replicas of the gradients.

Finally, once you have defined the training step, you can iterate over dist_dataset and run the training in a loop:

for dist_inputs in dist_dataset:
  print(distributed_train_step(dist_inputs))
tf.Tensor(1.6688763, shape=(), dtype=float32)
tf.Tensor(1.6557245, shape=(), dtype=float32)
tf.Tensor(1.6426537, shape=(), dtype=float32)
tf.Tensor(1.6296641, shape=(), dtype=float32)
tf.Tensor(1.6167561, shape=(), dtype=float32)
tf.Tensor(1.6039299, shape=(), dtype=float32)
tf.Tensor(1.5911858, shape=(), dtype=float32)
tf.Tensor(1.578524, shape=(), dtype=float32)
tf.Tensor(1.5659449, shape=(), dtype=float32)
tf.Tensor(1.5534486, shape=(), dtype=float32)
tf.Tensor(1.5410352, shape=(), dtype=float32)
tf.Tensor(1.5287051, shape=(), dtype=float32)
tf.Tensor(1.5164585, shape=(), dtype=float32)
tf.Tensor(1.5042956, shape=(), dtype=float32)
tf.Tensor(1.4922162, shape=(), dtype=float32)
tf.Tensor(1.4802208, shape=(), dtype=float32)
tf.Tensor(1.4683094, shape=(), dtype=float32)
tf.Tensor(1.4564822, shape=(), dtype=float32)
tf.Tensor(1.4447392, shape=(), dtype=float32)
tf.Tensor(1.4330806, shape=(), dtype=float32)

In the example above, you iterated over the dist_dataset to provide input to your training. You are also provided with the tf.distribute.Strategy.make_experimental_numpy_dataset to support NumPy inputs. You can use this API to create a dataset before calling tf.distribute.Strategy.experimental_distribute_dataset.

Another way of iterating over your data is to explicitly use iterators. You may want to do this when you want to run for a given number of steps as opposed to iterating over the entire dataset. The above iteration would now be modified to first create an iterator and then explicitly call next on it to get the input data.

iterator = iter(dist_dataset)
for _ in range(10):
  print(distributed_train_step(next(iterator)))
tf.Tensor(1.4215062, shape=(), dtype=float32)
tf.Tensor(1.4100163, shape=(), dtype=float32)
tf.Tensor(1.398611, shape=(), dtype=float32)
tf.Tensor(1.3872901, shape=(), dtype=float32)
tf.Tensor(1.3760538, shape=(), dtype=float32)
tf.Tensor(1.364902, shape=(), dtype=float32)
tf.Tensor(1.3538347, shape=(), dtype=float32)
tf.Tensor(1.342852, shape=(), dtype=float32)
tf.Tensor(1.3319536, shape=(), dtype=float32)
tf.Tensor(1.3211398, shape=(), dtype=float32)

This covers the simplest case of using tf.distribute.Strategy API to distribute custom training loops.

What's supported now?

Training API MirroredStrategy TPUStrategy MultiWorkerMirroredStrategy ParameterServerStrategy CentralStorageStrategy
Custom Training Loop Supported Supported Experimental support Experimental support Experimental support

Examples and tutorials

Here are some examples for using distribution strategy with custom training loops:

  1. Tutorial to train MNIST using MirroredStrategy.
  2. Guide on training MNIST using TPUStrategy.
  3. TensorFlow Model Garden repository containing collections of state-of-the-art models implemented using various strategies.

Other topics

This section covers some topics that are relevant to multiple use cases.

Setting up the TF_CONFIG environment variable

For multi-worker training, as mentioned before, you need to set up the TF_CONFIG environment variable for each binary running in your cluster. The TF_CONFIG environment variable is a JSON string which specifies what tasks constitute a cluster, their addresses and each task's role in the cluster. The tensorflow/ecosystem repo provides a Kubernetes template, which sets up TF_CONFIG for your training tasks.

There are two components of TF_CONFIG: a cluster and a task.

  • A cluster provides information about the training cluster, which is a dict consisting of different types of jobs such as worker. In multi-worker training, there is usually one worker that takes on a little more responsibility like saving checkpoint and writing summary file for TensorBoard in addition to what a regular worker does. Such worker is referred to as the "chief" worker, and it is customary that the worker with index 0 is appointed as the chief worker (in fact this is how tf.distribute.Strategy is implemented).
  • A task on the other hand provides information of the current task. The first component cluster is the same for all workers, and the second component task is different on each worker and specifies the type and index of that worker.

One example of TF_CONFIG is:

os.environ["TF_CONFIG"] = json.dumps({
    "cluster": {
        "worker": ["host1:port", "host2:port", "host3:port"],
        "ps": ["host4:port", "host5:port"]
    },
   "task": {"type": "worker", "index": 1}
})

This TF_CONFIG specifies that there are three workers and two "ps" tasks in the "cluster" along with their hosts and ports. The "task" part specifies that the role of the current task in the "cluster"—worker 1 (the second worker). Valid roles in a cluster are "chief", "worker", "ps", and "evaluator". There should be no "ps" job except when using tf.distribute.experimental.ParameterServerStrategy.

What's next?

tf.distribute.Strategy is actively under development. Try it out and provide and your feedback using GitHub issues.