Python horovod.tensorflow.allreduce() Examples

def update_state(self, sparse_predictions, samples, logit_length=None):
        """ Accumulate errors and counts """
        validated_label = tf.cast(
            tf.sparse.from_dense(samples["output"]), dtype=tf.int64
        )
        labels_counter = tf.cast(tf.shape(validated_label.values)[0], tf.float32)

        num_errs = tf.edit_distance(
            sparse_predictions, validated_label, normalize=False
        )
        num_errs = tf.reduce_sum(num_errs)
        if self.rank_size > 1:
            num_errs = hvd.allreduce(num_errs, average=False)
            labels_counter = hvd.allreduce(labels_counter, average=False)
        self.error_count(num_errs)
        self.total_count(labels_counter)
        return num_errs, labels_counter 

def get_gradients(self, loss, params):
        """
        Compute gradients of all trainable variables.

        See Optimizer.get_gradients() for more info.

        In DistributedOptimizer, get_gradients() is overriden to also
        allreduce the gradients before returning them.
        """
        gradients = super(self.__class__, self).get_gradients(loss, params)
        if hvd.size() > 1:
            averaged_gradients = []
            with tf.name_scope(self._name + "_Allreduce"):
                for grad in gradients:
                    if grad is not None:
                        avg_grad = hvd.allreduce(grad, device_dense=self._device_dense,
                                                 device_sparse=self._device_sparse)
                        averaged_gradients.append(avg_grad)
                    else:
                        averaged_gradients.append(None)
                return averaged_gradients
        else:
            return gradients 

def DistributedOptimizer(optimizer, name=None, device_dense='', device_sparse=''):
    """
    An optimizer that wraps another keras.optimizers.Optimizer, using an allreduce to
    average gradient values before applying gradients to model weights.

    Args:
        optimizer: Optimizer to use for computing gradients and applying updates.
        name: Optional name prefix for the operations created when applying
              gradients. Defaults to "Distributed" followed by the provided
              optimizer type.
        device_dense: Device to be used for dense tensors. Uses GPU by default
                      if Horovod was build with HOROVOD_GPU_ALLREDUCE.
        device_sparse: Device to be used for sparse tensors. Uses GPU by default
                       if Horovod was build with HOROVOD_GPU_ALLGATHER.
    """
    # We dynamically create a new class that inherits from the optimizer that was passed in.
    # The goal is to override get_gradients() method with an allreduce implementation.
    # This class will have the same name as the optimizer it's wrapping, so that the saved
    # model could be easily restored without Horovod.
    cls = type(optimizer.__class__.__name__, (optimizer.__class__,),
               dict(_DistributedOptimizer.__dict__))
    return cls(name, device_dense, device_sparse, **optimizer.get_config()) 

def _setup_graph(self):
        num_gpu = cfg.TRAIN.NUM_GPUS
        if cfg.TRAINER == 'replicated':
            # TF bug in version 1.11, 1.12: https://github.com/tensorflow/tensorflow/issues/22750
            buggy_tf = get_tf_version_tuple() in [(1, 11), (1, 12)]

            # Use two predictor threads per GPU to get better throughput
            self.num_predictor = num_gpu if buggy_tf else num_gpu * 2
            self.predictors = [self._build_predictor(k % num_gpu) for k in range(self.num_predictor)]
            self.dataflows = [get_eval_dataflow(self._eval_dataset,
                                                shard=k, num_shards=self.num_predictor)
                              for k in range(self.num_predictor)]
        else:
            # Only eval on the first machine,
            # Because evaluation assumes that all horovod workers share the filesystem.
            # Alternatively, can eval on all ranks and use allgather, but allgather sometimes hangs
            self._horovod_run_eval = hvd.rank() == hvd.local_rank()
            if self._horovod_run_eval:
                self.predictor = self._build_predictor(0)
                self.dataflow = get_eval_dataflow(self._eval_dataset,
                                                  shard=hvd.local_rank(), num_shards=hvd.local_size())

            self.barrier = hvd.allreduce(tf.random_normal(shape=[1])) 

def test_horovod_allreduce_cpu_gpu_error(self):
        """Test that the allreduce raises an error if different ranks try to
        perform reduction on CPU and GPU."""
        # Only do this test if there are GPUs available.
        if not tf.test.is_gpu_available(cuda_only=True):
            return

        hvd.init()
        local_rank = hvd.local_rank()
        size = hvd.size()

        # This test does not apply if there is only one worker.
        if size == 1:
            return

        device = "/gpu:%d" % local_rank if local_rank % 2 == 0 else "/cpu:0"
        with self.test_session(config=self.config) as session:
            with tf.device(device):
                # Same rank, different dimension
                dims = [17] * 3
                tensor = tf.ones(dims, dtype=tf.int32)
                with self.assertRaises(tf.errors.FailedPreconditionError):
                    session.run(hvd.allreduce(tensor)) 

def test_horovod_allreduce_type_error(self):
        """Test that the allreduce raises an error if different ranks try to
        send tensors of different type."""
        hvd.init()
        rank = hvd.rank()
        size = hvd.size()

        # This test does not apply if there is only one worker.
        if size == 1:
            return

        with self.test_session(config=self.config) as session:
            # Same rank, different dimension
            dims = [17] * 3
            tensor = tf.ones(dims,
                             dtype=tf.int32 if rank % 2 == 0 else tf.float32)
            with self.assertRaises(tf.errors.FailedPreconditionError):
                session.run(hvd.allreduce(tensor)) 

def test_horovod_allreduce_gpu(self):
        """Test that the allreduce works on GPUs.

        This test will crash badly if used with an MPI implementation that does
        not support GPU memory transfers directly, as it will call MPI_Send on
        a GPU data pointer."""
        # Only do this test if there are GPUs available.
        if not tf.test.is_gpu_available(cuda_only=True):
            return

        hvd.init()
        local_rank = hvd.local_rank()
        size = hvd.size()

        with self.test_session(config=self.config) as session:
            dtypes = [tf.int32, tf.int64, tf.float16, tf.float32, tf.float64]
            dims = [1, 2, 3]
            for dtype, dim in itertools.product(dtypes, dims):
                with tf.device("/gpu:%d" % local_rank):
                    tf.set_random_seed(1234)
                    tensor = tf.random_uniform(
                        [17] * dim, -100, 100, dtype=dtype)
                    summed = hvd.allreduce(tensor, average=False)
                multiplied = tensor * size
                max_difference = tf.reduce_max(tf.abs(summed - multiplied))

                # Threshold for floating point equality depends on number of
                # ranks, since we're comparing against precise multiplication.
                if size <= 3 or dtype in [tf.int32, tf.int64]:
                    threshold = 0
                elif size < 10:
                    threshold = 1e-4
                elif size < 15:
                    threshold = 5e-4
                else:
                    return

                diff = session.run(max_difference)
                self.assertTrue(diff <= threshold,
                                "hvd.allreduce on GPU produces incorrect results") 

def reduce_gradients(grads_and_vars, on_horovod, model=None):
  if on_horovod:
    from horovod.tensorflow import allreduce, size

    if size() > 1:
      averaged_grads_and_vars = []
      with tf.name_scope("all_reduce"):
        for grad, var in grads_and_vars:
          if grad is not None:
            if isinstance(grad, tf.IndexedSlices):
              if model._decoder.params.get('shared_embed', False):
                from tensorflow.python.training.optimizer import _deduplicate_indexed_slices
                summed_values, unique_indices = _deduplicate_indexed_slices(
                    values=grad.values, indices=grad.indices)
                gradient_no_duplicate_indices = tf.IndexedSlices(
                    indices=unique_indices,
                    values=summed_values,
                    dense_shape=grad.dense_shape)
                grad = tf.convert_to_tensor(gradient_no_duplicate_indices)
            avg_grad = allreduce(grad)
            averaged_grads_and_vars.append((avg_grad, var))
          else:
            averaged_grads_and_vars.append((None, var))
      return averaged_grads_and_vars
    else:
      return grads_and_vars
  else:
    raise NotImplementedError("Reduce in tower-mode is not implemented.") 

def _add_aggregation_ops(gradients_info, op_to_control_consumer_ops, config):
    grad_tensor = gradients_info._grad
    if isinstance(grad_tensor, tf.Tensor):
        grad = grad_tensor
        grad_consumers = [c for c in grad.consumers()]
        agg_grad = hvd.allreduce(grad,
                                 average=True)
        update_consumers(grad_consumers, grad, agg_grad)
        update_control_consumers(op_to_control_consumer_ops[grad.op],
                                 grad.op, agg_grad.op)
    else:
        grad = grad_tensor.values
        indices = grad_tensor.indices
        dense_shape = grad_tensor.dense_shape
        grad_consumers = [c for c in grad.consumers()]
        indices_consumers = [c for c in indices.consumers()]
        agg_grad = \
            hvd.allreduce(tf.IndexedSlices(grad, indices, dense_shape),
                          average=config.average_sparse)
        update_consumers(grad_consumers, grad, agg_grad.values)
        update_consumers(indices_consumers, indices, agg_grad.indices)
        update_control_consumers(op_to_control_consumer_ops[grad.op],
                                 grad.op, agg_grad.values.op)
        update_control_consumers(
            op_to_control_consumer_ops[indices.op], indices.op,
            agg_grad.indices.op)
    gradients_info._grad = agg_grad 

def test_horovod_allreduce_grad(self):
        """Test the correctness of the allreduce gradient."""
        hvd.init()
        size = hvd.size()

        with self.test_session(config=self.config) as session:
            # As of TensorFlow v1.9, gradients are not supported on
            # integer tensors
            dtypes = [tf.float32, tf.float64]
            dims = [1, 2, 3]
            for dtype, dim in itertools.product(dtypes, dims):
                with tf.device("/cpu:0"):
                    tf.set_random_seed(1234)
                    tensor = tf.random_uniform(
                        [5] * dim, -100, 100, dtype=dtype)
                    summed = hvd.allreduce(tensor, average=False)

                grad_ys = tf.ones([5] * dim)
                grad = tf.gradients(summed, tensor, grad_ys)[0]
                grad_out = session.run(grad)

                expected = np.ones([5] * dim) * size
                err = np.linalg.norm(expected - grad_out)
                self.assertLess(err, 0.00000001,
                                "gradient %s differs from expected %s, "
                                "error: %s" % (grad_out, expected, str(err))) 

def _setup_graph(self):
        self._placeholder = tf.placeholder(tf.float32, shape=[2], name='to_be_reduced')
        self._reduced = hvd.allreduce(self._placeholder, average=False) 

def test_horovod_allreduce_cpu_fused(self):
        """Test on CPU that the allreduce correctly sums 1D, 2D, 3D tensors
        with Tensor Fusion."""
        hvd.init()
        size = hvd.size()
        with self.test_session(config=self.config) as session:
            dtypes = [tf.int32, tf.int64, tf.float32, tf.float64]
            dims = [1, 2, 3]
            tests = []
            for dtype, dim in itertools.product(dtypes, dims):
                with tf.device("/cpu:0"):
                    tf.set_random_seed(1234)
                    tensor = tf.random_uniform(
                        [17] * dim, -100, 100, dtype=dtype)
                    summed = hvd.allreduce(tensor, average=False)
                multiplied = tensor * size
                max_difference = tf.reduce_max(tf.abs(summed - multiplied))

                # Threshold for floating point equality depends on number of
                # ranks, since we're comparing against precise multiplication.
                if size <= 3 or dtype in [tf.int32, tf.int64]:
                    threshold = 0
                elif size < 10:
                    threshold = 1e-4
                elif size < 15:
                    threshold = 5e-4
                else:
                    break

                test = max_difference <= threshold
                tests.append(test)
            self.assertTrue(session.run(tf.reduce_all(tests)),
                            "hvd.allreduce produces incorrect results") 

def test_horovod_allreduce_average(self):
        """Test on CPU that the allreduce correctly sums 1D, 2D, 3D tensors."""
        hvd.init()
        size = hvd.size()
        with self.test_session() as session:
            dtypes = [tf.int32, tf.int64, tf.float32, tf.float64]
            dims = [1, 2, 3]
            for dtype, dim in itertools.product(dtypes, dims):
                with tf.device("/cpu:0"):
                    tf.set_random_seed(1234)
                    tensor = tf.random_uniform(
                        [17] * dim, -100, 100, dtype=dtype)
                    summed = hvd.allreduce(tensor, average=True)
                max_difference = tf.reduce_max(tf.abs(summed - tensor))

                # Threshold for floating point equality depends on number of
                # ranks, since we're comparing against precise multiplication.
                if dtype in [tf.int32, tf.int64]:
                    threshold = hvd.size()
                elif size <= 3:
                    threshold = 0
                elif size < 10:
                    threshold = 1e-4
                elif size < 15:
                    threshold = 5e-4
                else:
                    break

                diff = session.run(max_difference)
                self.assertTrue(diff <= threshold,
                                "hvd.allreduce produces incorrect results") 

def _make_variable(self, metric, value):
        with tf.name_scope('MetricAverageCallback'):
            var = tf.Variable(value, name=metric)
            K.get_session().run(var.initializer)
            allreduce_op = hvd.allreduce(var, device_dense=self.device)
            return var, allreduce_op 

def __init__(self, device=''):
        """
        Construct a new MetricAverageCallback that will average metrics
        across all processes at the end of the epoch.

        Args:
            device: Device to be used for allreduce. Uses GPU by default
                    if Horovod was build with HOROVOD_GPU_ALLREDUCE.
        """
        super(MetricAverageCallback, self).__init__()
        self.variables = {}
        self.allreduce_ops = {}
        self.device = device 

def allreduce(value, name=None, average=True):
    """
    Perform an allreduce on a tensor-compatible value.

    Arguments:
        value: A tensor-compatible value to reduce.
               The shape of the input must be identical across all ranks.
        name: Optional name for the constants created by this operation.
        average: If True, computes the average over all ranks.
                 Otherwise, computes the sum over all ranks.
    """
    allreduce_op = hvd.allreduce(tf.constant(value, name=name), average=average)
    return K.get_session().run(allreduce_op) 

def test_horovod_allreduce_gpu_fused(self):
        """Test that the allreduce works on GPUs with Tensor Fusion.

        This test will crash badly if used with an MPI implementation that does
        not support GPU memory transfers directly, as it will call MPI_Send on
        a GPU data pointer."""
        # Only do this test if there are GPUs available.
        if not tf.test.is_gpu_available(cuda_only=True):
            return

        hvd.init()
        local_rank = hvd.local_rank()
        size = hvd.size()

        with self.test_session(config=self.config) as session:
            dtypes = [tf.int32, tf.int64, tf.float16, tf.float32, tf.float64]
            dims = [1, 2, 3]
            tests = []
            for dtype, dim in itertools.product(dtypes, dims):
                with tf.device("/gpu:%d" % local_rank):
                    tf.set_random_seed(1234)
                    tensor = tf.random_uniform(
                        [17] * dim, -100, 100, dtype=dtype)
                    summed = hvd.allreduce(tensor, average=False)
                multiplied = tensor * size
                max_difference = tf.reduce_max(tf.abs(summed - multiplied))

                # Threshold for floating point equality depends on number of
                # ranks, since we're comparing against precise multiplication.
                if size <= 3 or dtype in [tf.int32, tf.int64]:
                    threshold = 0
                elif size < 10:
                    threshold = 1e-4
                elif size < 15:
                    threshold = 5e-4
                else:
                    return

                test = max_difference <= threshold
                tests.append(test)
            self.assertTrue(session.run(tf.reduce_all(tests)),
                            "hvd.allreduce produces incorrect results") 

def test_horovod_allreduce_multi_gpu(self):
        """Test that the allreduce works on multiple GPUs.

        This test will crash badly if used with an MPI implementation that does
        not support GPU memory transfers directly, as it will call MPI_Send on
        a GPU data pointer."""
        # Only do this test if there are GPUs available.
        if not tf.test.is_gpu_available(cuda_only=True):
            return

        hvd.init()
        local_rank = hvd.local_rank()
        size = hvd.size()

        iter = 0
        gpu_ids = [local_rank * 2, local_rank * 2 + 1]
        with self.test_session(config=self.config) as session:
            dtypes = [tf.int32, tf.int64, tf.float16, tf.float32, tf.float64]
            dims = [1, 2, 3]
            for dtype, dim in itertools.product(dtypes, dims):
                iter += 1
                with tf.device("/gpu:%d" % gpu_ids[(iter + local_rank) % 2]):
                    tf.set_random_seed(1234)
                    tensor = tf.random_uniform(
                        [17] * dim, -100, 100, dtype=dtype)
                    summed = hvd.allreduce(tensor, average=False)
                multiplied = tensor * size
                max_difference = tf.reduce_max(tf.abs(summed - multiplied))

                # Threshold for floating point equality depends on number of
                # ranks, since we're comparing against precise multiplication.
                if size <= 3 or dtype in [tf.int32, tf.int64]:
                    threshold = 0
                elif size < 10:
                    threshold = 1e-4
                elif size < 15:
                    threshold = 5e-4
                else:
                    return

                diff = session.run(max_difference)
                self.assertTrue(diff <= threshold,
                                "hvd.allreduce on GPU produces incorrect results") 

def get_sync_bn_mean_var(inputs, red_axis, sync_statistics):
    ctx = get_current_tower_context()
    batch_mean = tf.reduce_mean(inputs, axis=red_axis)
    batch_mean_square = tf.reduce_mean(tf.square(inputs), axis=red_axis)

    TF_version = get_tf_version_tuple()

    if sync_statistics == 'nccl':
        num_dev = ctx.total
        if num_dev == 1:
            logger.warn("BatchNorm(sync_statistics='nccl') is used with only one tower!")
        else:
            assert TF_version >= (1, 10), \
                "Cross-GPU BatchNorm is only supported in TF>=1.10 ." \
                "Upgrade TF or apply this patch manually: https://github.com/tensorflow/tensorflow/pull/20360"

            if TF_version <= (1, 12):
                try:
                    from tensorflow.contrib.nccl.python.ops.nccl_ops import _validate_and_load_nccl_so  # deprecated
                except Exception:
                    pass
                else:
                    _validate_and_load_nccl_so()
                from tensorflow.contrib.nccl.ops import gen_nccl_ops  # deprecated
            else:
                from tensorflow.python.ops import gen_nccl_ops
            shared_name = re.sub('tower[0-9]+/', '', tf.get_variable_scope().name)
            batch_mean = gen_nccl_ops.nccl_all_reduce(
                input=batch_mean,
                reduction='sum',
                num_devices=num_dev,
                shared_name=shared_name + '_NCCL_mean') * (1.0 / num_dev)
            batch_mean_square = gen_nccl_ops.nccl_all_reduce(
                input=batch_mean_square,
                reduction='sum',
                num_devices=num_dev,
                shared_name=shared_name + '_NCCL_mean_square') * (1.0 / num_dev)
    elif sync_statistics == 'horovod':
        # Require https://github.com/uber/horovod/pull/331
        import horovod.tensorflow as hvd
        if hvd.size() == 1:
            logger.warn("BatchNorm(sync_statistics='horovod') is used with only one process!")
        else:
            import horovod
            hvd_version = tuple(map(int, horovod.__version__.split('.')[:3]))
            assert hvd_version >= (0, 13, 6), "sync_statistics=horovod needs horovod>=0.13.6 !"

            batch_mean = hvd.allreduce(batch_mean, average=True)
            batch_mean_square = hvd.allreduce(batch_mean_square, average=True)
    batch_var = batch_mean_square - tf.square(batch_mean)
    return batch_mean, batch_var