How to run Tensorflow Estimator on multiple GPUs with data parallelism

The standard example is: https://github.com/tensorflow/tensorflow/blob/r1.4/tensorflow/contrib/learn/python/learn/estimators/estimator.py

One way to run it data-parallel would be to loop over available GPU devices, and send chunks of your batch to copied versions of your model (all done within your model_fn), then merge the results.

You can use scope and device for that:

 with tf.variable_scope(tf.get_variable_scope()):
  for i in xrange(FLAGS.num_gpus):
    with tf.device('/gpu:%d' % i):
      with tf.name_scope('%s_%d' % (cifar10.TOWER_NAME, i)) as scope:

Full example there: https://github.com/tensorflow/models/blob/master/tutorials/image/cifar10/cifar10_multi_gpu_train.py

I think this is all you need.

Link: https://www.youtube.com/watch?v=bRMGoPqsn20

More Details: https://www.tensorflow.org/api_docs/python/tf/distribute/Strategy

Explained: https://medium.com/tensorflow/multi-gpu-training-with-estimators-tf-keras-and-tf-data-ba584c3134db

NUM_GPUS = 8
dist_strategy = tf.contrib.distribute.MirroredStrategy(num_gpus=NUM_GPUS)
config = tf.estimator.RunConfig(train_distribute=dist_strategy)
estimator = tf.estimator.Estimator(model_fn,model_dir,config=config)

UPDATED

With TF-2.0 and Keras you may use this (https://www.tensorflow.org/tutorials/distribute/keras)

I think tf.contrib.estimator.replicate_model_fn is a cleaner solution. The following is from tf.contrib.estimator.replicate_model_fn documentation,

...
def model_fn(...):  # See `model_fn` in `Estimator`.
  loss = ...
  optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)
  optimizer = tf.contrib.estimator.TowerOptimizer(optimizer)
  if mode == tf.estimator.ModeKeys.TRAIN:
    #  See the section below on `EstimatorSpec.train_op`.
    return EstimatorSpec(mode=mode, loss=loss,
                         train_op=optimizer.minimize(loss))

  #  No change for `ModeKeys.EVAL` or `ModeKeys.PREDICT`.
  return EstimatorSpec(...)
...
classifier = tf.estimator.Estimator(
  model_fn=tf.contrib.estimator.replicate_model_fn(model_fn))

What you need to do is to wrap optimizer with tf.contrib.estimator.TowerOptimize and model_fn() with tf.contrib.estimator.replicate_model_fn(). I followed the description and make an TPU squeezenet model work on a machine with 4 GPUs. My modifications here.