Migrate single-worker multiple-GPU training

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

This guide demonstrates how to migrate the single-worker multiple-GPU workflows from TensorFlow 1 to TensorFlow 2.

To perform synchronous training across multiple GPUs on one machine:

Setup

Start with imports and a simple dataset for demonstration purposes:

import tensorflow as tf
import tensorflow.compat.v1 as tf1
features = [[1., 1.5], [2., 2.5], [3., 3.5]]
labels = [[0.3], [0.5], [0.7]]
eval_features = [[4., 4.5], [5., 5.5], [6., 6.5]]
eval_labels = [[0.8], [0.9], [1.]]

TensorFlow 1: Single-worker distributed training with tf.estimator.Estimator

This example demonstrates the TensorFlow 1 canonical workflow of single-worker multiple-GPU training. You need to set the distribution strategy (tf.distribute.MirroredStrategy) through the config parameter of the tf.estimator.Estimator:

def _input_fn():
  return tf1.data.Dataset.from_tensor_slices((features, labels)).batch(1)

def _eval_input_fn():
  return tf1.data.Dataset.from_tensor_slices(
      (eval_features, eval_labels)).batch(1)

def _model_fn(features, labels, mode):
  logits = tf1.layers.Dense(1)(features)
  loss = tf1.losses.mean_squared_error(labels=labels, predictions=logits)
  optimizer = tf1.train.AdagradOptimizer(0.05)
  train_op = optimizer.minimize(loss, global_step=tf1.train.get_global_step())
  return tf1.estimator.EstimatorSpec(mode, loss=loss, train_op=train_op)

strategy = tf1.distribute.MirroredStrategy()
config = tf1.estimator.RunConfig(
    train_distribute=strategy, eval_distribute=strategy)
estimator = tf1.estimator.Estimator(model_fn=_model_fn, config=config)

train_spec = tf1.estimator.TrainSpec(input_fn=_input_fn)
eval_spec = tf1.estimator.EvalSpec(input_fn=_eval_input_fn)
tf1.estimator.train_and_evaluate(estimator, train_spec, eval_spec)
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)
INFO:tensorflow:Initializing RunConfig with distribution strategies.
INFO:tensorflow:Not using Distribute Coordinator.
WARNING:tensorflow:Using temporary folder as model directory: /tmp/tmp5g_f_ufk
INFO:tensorflow:Using config: {'_model_dir': '/tmp/tmp5g_f_ufk', '_tf_random_seed': None, '_save_summary_steps': 100, '_save_checkpoints_steps': None, '_save_checkpoints_secs': 600, '_session_config': allow_soft_placement: true
graph_options {
  rewrite_options {
    meta_optimizer_iterations: ONE
  }
}
, '_keep_checkpoint_max': 5, '_keep_checkpoint_every_n_hours': 10000, '_log_step_count_steps': 100, '_train_distribute': <tensorflow.python.distribute.mirrored_strategy.MirroredStrategyV1 object at 0x7f6853562450>, '_device_fn': None, '_protocol': None, '_eval_distribute': <tensorflow.python.distribute.mirrored_strategy.MirroredStrategyV1 object at 0x7f6853562450>, '_experimental_distribute': None, '_experimental_max_worker_delay_secs': None, '_session_creation_timeout_secs': 7200, '_checkpoint_save_graph_def': True, '_service': None, '_cluster_spec': ClusterSpec({}), '_task_type': 'worker', '_task_id': 0, '_global_id_in_cluster': 0, '_master': '', '_evaluation_master': '', '_is_chief': True, '_num_ps_replicas': 0, '_num_worker_replicas': 1, '_distribute_coordinator_mode': None}
INFO:tensorflow:Not using Distribute Coordinator.
INFO:tensorflow:Running training and evaluation locally (non-distributed).
INFO:tensorflow:Start train and evaluate loop. The evaluate will happen after every checkpoint. Checkpoint frequency is determined based on RunConfig arguments: save_checkpoints_steps None or save_checkpoints_secs 600.
/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/tensorflow/python/data/ops/dataset_ops.py:374: UserWarning: To make it possible to preserve tf.data options across serialization boundaries, their implementation has moved to be part of the TensorFlow graph. As a consequence, the options value is in general no longer known at graph construction time. Invoking this method in graph mode retains the legacy behavior of the original implementation, but note that the returned value might not reflect the actual value of the options.
  warnings.warn("To make it possible to preserve tf.data options across "
INFO:tensorflow:Calling model_fn.
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.7/site-packages/tensorflow/python/training/adagrad.py:77: calling Constant.__init__ (from tensorflow.python.ops.init_ops) with dtype is deprecated and will be removed in a future version.
Instructions for updating:
Call initializer instance with the dtype argument instead of passing it to the constructor
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Create CheckpointSaverHook.
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.7/site-packages/tensorflow_estimator/python/estimator/util.py:95: DistributedIteratorV1.initialize (from tensorflow.python.distribute.input_lib) is deprecated and will be removed in a future version.
Instructions for updating:
Use the iterator's `initializer` property instead.
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Calling checkpoint listeners before saving checkpoint 0...
INFO:tensorflow:Saving checkpoints for 0 into /tmp/tmp5g_f_ufk/model.ckpt.
INFO:tensorflow:Calling checkpoint listeners after saving checkpoint 0...
2021-09-22 20:15:43.228503: W tensorflow/core/grappler/utils/graph_view.cc:836] No registered 'MultiDeviceIteratorFromStringHandle' OpKernel for GPU devices compatible with node { {node MultiDeviceIteratorFromStringHandle} }
    .  Registered:  device='CPU'

2021-09-22 20:15:43.229960: W tensorflow/core/grappler/utils/graph_view.cc:836] No registered 'MultiDeviceIteratorGetNextFromShard' OpKernel for GPU devices compatible with node { {node MultiDeviceIteratorGetNextFromShard} }
    .  Registered:  device='CPU'
INFO:tensorflow:loss = 0.14477473, step = 0
INFO:tensorflow:Calling checkpoint listeners before saving checkpoint 3...
INFO:tensorflow:Saving checkpoints for 3 into /tmp/tmp5g_f_ufk/model.ckpt.
INFO:tensorflow:Calling checkpoint listeners after saving checkpoint 3...
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Reduce to /replica:0/task:0/device:CPU:0 then broadcast to ('/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /replica:0/task:0/device:CPU:0 then broadcast to ('/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /replica:0/task:0/device:CPU:0 then broadcast to ('/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /replica:0/task:0/device:CPU:0 then broadcast to ('/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Starting evaluation at 2021-09-22T20:15:43
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Restoring parameters from /tmp/tmp5g_f_ufk/model.ckpt-3
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Inference Time : 0.17626s
INFO:tensorflow:Finished evaluation at 2021-09-22-20:15:44
INFO:tensorflow:Saving dict for global step 3: global_step = 3, loss = 1.1251448
INFO:tensorflow:Saving 'checkpoint_path' summary for global step 3: /tmp/tmp5g_f_ufk/model.ckpt-3
INFO:tensorflow:Loss for final step: 0.45722964.
2021-09-22 20:15:44.095116: W tensorflow/core/grappler/utils/graph_view.cc:836] No registered 'MultiDeviceIteratorFromStringHandle' OpKernel for GPU devices compatible with node { {node MultiDeviceIteratorFromStringHandle} }
    .  Registered:  device='CPU'

2021-09-22 20:15:44.096454: W tensorflow/core/grappler/utils/graph_view.cc:836] No registered 'MultiDeviceIteratorGetNextFromShard' OpKernel for GPU devices compatible with node { {node MultiDeviceIteratorGetNextFromShard} }
    .  Registered:  device='CPU'
({'loss': 1.1251448, 'global_step': 3}, [])

TensorFlow 2: Single-worker training with Keras

When migrating to TensorFlow 2, you can use the Keras APIs with tf.distribute.MirroredStrategy.

If you use the tf.keras APIs for model building and Keras Model.fit for training, the main difference is instantiating the Keras model, an optimizer, and metrics in the context of Strategy.scope, instead of defining a config for tf.estimator.Estimator.

If you need to use a custom training loop, check out the Using tf.distribute.Strategy with custom training loops guide.

dataset = tf.data.Dataset.from_tensor_slices((features, labels)).batch(1)
eval_dataset = tf.data.Dataset.from_tensor_slices(
      (eval_features, eval_labels)).batch(1)
strategy = tf.distribute.MirroredStrategy()
with strategy.scope():
  model = tf.keras.models.Sequential([tf.keras.layers.Dense(1)])
  optimizer = tf.keras.optimizers.Adagrad(learning_rate=0.05)

model.compile(optimizer=optimizer, loss='mse')
model.fit(dataset)
model.evaluate(eval_dataset, return_dict=True)
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',).
2021-09-22 20:15:44.265351: 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: "TensorSliceDataset/_2"
op: "TensorSliceDataset"
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: 2
        }
      }
      shape {
        dim {
          size: 1
        }
      }
    }
  }
}
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',).
3/3 [==============================] - 2s 3ms/step - loss: 0.2363
2021-09-22 20:15:46.836745: 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: "TensorSliceDataset/_2"
op: "TensorSliceDataset"
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: 2
        }
      }
      shape {
        dim {
          size: 1
        }
      }
    }
  }
}
3/3 [==============================] - 1s 3ms/step - loss: 0.0079
{'loss': 0.007883546873927116}

Next steps

To learn more about distributed training with tf.distribute.MirroredStrategy in TensorFlow 2, check out the following documentation: