当前位置: 首页>>代码示例>>Python>>正文


Python tensorflow.allreduce方法代码示例

本文整理汇总了Python中horovod.tensorflow.allreduce方法的典型用法代码示例。如果您正苦于以下问题:Python tensorflow.allreduce方法的具体用法?Python tensorflow.allreduce怎么用?Python tensorflow.allreduce使用的例子?那么, 这里精选的方法代码示例或许可以为您提供帮助。您也可以进一步了解该方法所在horovod.tensorflow的用法示例。


在下文中一共展示了tensorflow.allreduce方法的15个代码示例,这些例子默认根据受欢迎程度排序。您可以为喜欢或者感觉有用的代码点赞,您的评价将有助于系统推荐出更棒的Python代码示例。

示例1: update_state

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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 
开发者ID:athena-team,项目名称:athena,代码行数:19,代码来源:metrics.py

示例2: get_gradients

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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 
开发者ID:mlperf,项目名称:training_results_v0.6,代码行数:25,代码来源:__init__.py

示例3: DistributedOptimizer

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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()) 
开发者ID:mlperf,项目名称:training_results_v0.6,代码行数:24,代码来源:__init__.py

示例4: test_horovod_allreduce_type_error

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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)) 
开发者ID:mlperf,项目名称:training_results_v0.6,代码行数:20,代码来源:test_tensorflow.py

示例5: test_horovod_allreduce_cpu_gpu_error

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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)) 
开发者ID:mlperf,项目名称:training_results_v0.6,代码行数:25,代码来源:test_tensorflow.py

示例6: _setup_graph

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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])) 
开发者ID:tensorpack,项目名称:tensorpack,代码行数:25,代码来源:eval.py

示例7: _setup_graph

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
def _setup_graph(self):
        self._placeholder = tf.placeholder(tf.float32, shape=[2], name='to_be_reduced')
        self._reduced = hvd.allreduce(self._placeholder, average=False) 
开发者ID:tensorpack,项目名称:benchmarks,代码行数:5,代码来源:imagenet-resnet-horovod.py

示例8: allreduce

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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) 
开发者ID:mlperf,项目名称:training_results_v0.6,代码行数:15,代码来源:__init__.py

示例9: __init__

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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 
开发者ID:mlperf,项目名称:training_results_v0.6,代码行数:15,代码来源:callbacks.py

示例10: _make_variable

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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 
开发者ID:mlperf,项目名称:training_results_v0.6,代码行数:8,代码来源:callbacks.py

示例11: test_horovod_allreduce_average

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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") 
开发者ID:mlperf,项目名称:training_results_v0.6,代码行数:33,代码来源:test_tensorflow.py

示例12: test_horovod_allreduce_cpu_fused

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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") 
开发者ID:mlperf,项目名称:training_results_v0.6,代码行数:35,代码来源:test_tensorflow.py

示例13: test_horovod_allreduce_gpu

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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") 
开发者ID:mlperf,项目名称:training_results_v0.6,代码行数:42,代码来源:test_tensorflow.py

示例14: test_horovod_allreduce_grad

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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))) 
开发者ID:mlperf,项目名称:training_results_v0.6,代码行数:28,代码来源:test_tensorflow.py

示例15: _add_aggregation_ops

# 需要导入模块: from horovod import tensorflow [as 别名]
# 或者: from horovod.tensorflow import allreduce [as 别名]
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 
开发者ID:snuspl,项目名称:parallax,代码行数:29,代码来源:graph_transform.py


注:本文中的horovod.tensorflow.allreduce方法示例由纯净天空整理自Github/MSDocs等开源代码及文档管理平台,相关代码片段筛选自各路编程大神贡献的开源项目,源码版权归原作者所有,传播和使用请参考对应项目的License;未经允许,请勿转载。