def __init__(self,
input_dataset,
device,
buffer_size):
with ops.device("/device:CPU:0"):
super(_PrefetchToDeviceEagerIterator, self).__init__(input_dataset)
input_iterator_handle = core_gen_dataset_ops.iterator_to_string_handle(
self._resource)
self._device = device
@function.Defun(dtypes.string)
def _prefetch_fn(handle):
"""Prefetches one element from `input_iterator`."""
remote_iterator = iterator_ops.Iterator.from_string_handle(
handle, self.output_types, self.output_shapes, self.output_classes)
ret = remote_iterator.get_next()
return nest.flatten(sparse.serialize_sparse_tensors(ret))
_prefetch_fn.add_to_graph(None)
with ops.device(device):
self._buffering_resource = function_buffering_resource(
f=_prefetch_fn,
output_types=self._flat_output_types,
target_device=gen_dataset_ops.iterator_get_device(self._resource),
string_arg=input_iterator_handle,
buffer_size=buffer_size,
shared_name=iterator_ops._generate_shared_name(
"function_buffer_resource"))
def __init__(self, iterator_resource, initializer, output_types,
output_shapes, output_classes):
"""Creates a new iterator from the given iterator resource.
Note: Most users will not call this initializer directly, and will
instead use `Dataset.make_initializable_iterator()` or
`Dataset.make_one_shot_iterator()`.
Args:
iterator_resource: A `tf.resource` scalar `tf.Tensor` representing the
iterator.
initializer: A `tf.Operation` that should be run to initialize this
iterator.
output_types: A nested structure of `tf.DType` objects corresponding to
each component of an element of this dataset.
output_shapes: A nested structure of `tf.TensorShape` objects
corresponding to each component of an element of this dataset.
output_classes: A nested structure of Python `type` object corresponding
to each
component of an element of this iterator.
"""
self._iterator_resource = iterator_resource
self._initializer = initializer
self._output_classes = output_classes
self._output_types = output_types
self._output_shapes = output_shapes
self._string_handle = gen_dataset_ops.iterator_to_string_handle(
self._iterator_resource)
self._get_next_call_count = 0
def __init__(self,
input_dataset,
device,
buffer_size):
with ops.device("/device:CPU:0"):
super(_PrefetchToDeviceEagerIterator, self).__init__(input_dataset)
input_iterator_handle = gen_dataset_ops.iterator_to_string_handle(
self._resource)
self._device = device
@function.defun(input_signature=[tensor_spec.TensorSpec([], dtypes.string)])
def _prefetch_fn(handle):
"""Prefetches one element from `input_iterator`."""
remote_iterator = iterator_ops.Iterator.from_string_handle(
handle, self.output_types, self.output_shapes, self.output_classes)
ret = remote_iterator.get_next()
return nest.flatten(sparse.serialize_sparse_tensors(ret))
self._prefetch_fn = _prefetch_fn._get_concrete_function_internal() # pylint: disable=protected-access
with ops.device(device):
self._buffering_resource = function_buffering_resource(
f=self._prefetch_fn,
output_types=self._flat_output_types,
target_device=ged_ops.experimental_iterator_get_device(
self._resource),
string_arg=input_iterator_handle,
buffer_size=buffer_size,
shared_name=iterator_ops._generate_shared_name(
"function_buffer_resource"))
def __init__(self, iterator_resource, initializer, output_types,
output_shapes, output_classes):
"""Creates a new iterator from the given iterator resource.
Note: Most users will not call this initializer directly, and will
instead use `Dataset.make_initializable_iterator()` or
`Dataset.make_one_shot_iterator()`.
Args:
iterator_resource: A `tf.resource` scalar `tf.Tensor` representing the
iterator.
initializer: A `tf.Operation` that should be run to initialize this
iterator.
output_types: A nested structure of `tf.DType` objects corresponding to
each component of an element of this iterator.
output_shapes: A nested structure of `tf.TensorShape` objects
corresponding to each component of an element of this iterator.
output_classes: A nested structure of Python `type` objects corresponding
to each component of an element of this iterator.
"""
self._iterator_resource = iterator_resource
self._initializer = initializer
if (output_types is None or output_shapes is None
or output_classes is None):
raise ValueError("If `structure` is not specified, all of "
"`output_types`, `output_shapes`, and `output_classes`"
" must be specified.")
self._structure = structure_lib.convert_legacy_structure(
output_types, output_shapes, output_classes)
self._string_handle = gen_dataset_ops.iterator_to_string_handle(
self._iterator_resource)
self._get_next_call_count = 0
ops.add_to_collection(GLOBAL_ITERATORS, self._iterator_resource)
def __init__(self, dataset):
"""Creates a new iterator over the given dataset.
For example:
```python
dataset = tf.data.Dataset.range(4)
for x in Iterator(dataset):
print(x)
```
Tensors produced will be placed on the device on which this iterator object
was created.
Args:
dataset: A `tf.data.Dataset` object.
Raises:
TypeError: If `dataset` is an unsupported type.
RuntimeError: When invoked without eager execution enabled.
"""
if isinstance(dataset, prefetching_ops._PrefetchToDeviceDataset): # pylint: disable=protected-access
raise TypeError(
"`tf.contrib.data.prefetch_to_device()` is not compatible with "
"`tf.contrib.eager.Iterator`. Use `for ... in dataset:` to iterate "
"over the dataset instead.")
super(Iterator, self).__init__(dataset)
if not context.context().device_spec.device_type:
is_remote_device = False
else:
is_remote_device = context.context().device_spec.device_type != "CPU"
self._buffer_resource_handle = None
if is_remote_device:
with ops.device("/device:CPU:0"):
iter_string_handle = gen_dataset_ops.iterator_to_string_handle(
self._resource)
@function.Defun(dtypes.string)
def remote_fn(h):
remote_iterator = iterator_ops.Iterator.from_string_handle(
h, self.output_types, self.output_shapes, self.output_classes)
return remote_iterator.get_next()
remote_fn.add_to_graph(None)
target = constant_op.constant("/device:CPU:0")
with ops.device(self._device):
self._buffer_resource_handle = prefetching_ops.function_buffering_resource( # pylint: disable=line-too-long
string_arg=iter_string_handle,
output_types=self._flat_output_types,
f=remote_fn,
target_device=target,
buffer_size=10,
container="",
shared_name=_generate_shared_name(
"contrib_eager_iterator_function_buffer_resource"))
self._buffer_resource_deleter = resource_variable_ops.EagerResourceDeleter( # pylint: disable=line-too-long
handle=self._buffer_resource_handle,
handle_device=self._device)
def __init__(self, iterator_resource, initializer, output_types,
output_shapes, output_classes):
self._iterator_resource = iterator_resource
self._initializer = initializer
self._output_classes = output_classes
self._output_types = output_types
self._output_shapes = output_shapes
self._string_handle = gen_dataset_ops.iterator_to_string_handle(
self._iterator_resource)
def string_handle(self, name=None):
"""Returns a string-valued `tf.Tensor` that represents this iterator.
Args:
name: (Optional.) A name for the created operation.
Returns:
A scalar `tf.Tensor` of type `tf.string`.
"""
return gen_dataset_ops.iterator_to_string_handle(
self._iterator_resource, name=name)
def _init_func():
"""Creates an iterator for the input dataset.
Returns:
A `string` tensor that encapsulates the iterator created.
"""
ds_variant = gen_dataset_ops.unwrap_dataset_variant(wrap_ds_variant)
resource = gen_dataset_ops.anonymous_iterator(
**dataset_ops.flat_structure(self._input_dataset))
with ops.control_dependencies(
[gen_dataset_ops.make_iterator(ds_variant, resource)]):
return gen_dataset_ops.iterator_to_string_handle(resource)
def _init_func():
"""Creates an iterator for the input dataset.
Returns:
A `string` tensor that encapsulates the iterator created.
"""
# pylint: disable=protected-access
ds_variant = self._input_dataset._as_variant_tensor()
resource = gen_dataset_ops.anonymous_iterator(
output_types=self._flat_output_types,
output_shapes=self._flat_output_shapes)
with ops.control_dependencies(
[gen_dataset_ops.make_iterator(ds_variant, resource)]):
return gen_dataset_ops.iterator_to_string_handle(resource)
def _init_func():
"""Creates an iterator for the input dataset.
Returns:
A `string` tensor that encapsulates the iterator created.
"""
# pylint: disable=protected-access
ds_variant = self._input_dataset._as_variant_tensor()
resource = gen_dataset_ops.anonymous_iterator(
output_types=self._flat_output_types,
output_shapes=self._flat_output_shapes)
with ops.control_dependencies(
[gen_dataset_ops.make_iterator(ds_variant, resource)]):
return gen_dataset_ops.iterator_to_string_handle(resource)
def string_handle(self, name=None):
"""Returns a string-valued `tf.Tensor` that represents this iterator.
Args:
name: (Optional.) A name for the created operation.
Returns:
A scalar `tf.Tensor` of type `tf.string`.
"""
if name is None:
return self._string_handle
else:
return gen_dataset_ops.iterator_to_string_handle(
self._iterator_resource, name=name)
def _init_func():
"""Creates an iterator for the input dataset.
Returns:
A `string` tensor that encapsulates the iterator created.
"""
ds_variant = gen_dataset_ops.unwrap_dataset_variant(
wrap_ds_variant)
resource = gen_dataset_ops.anonymous_iterator(
output_types=self._flat_output_types,
output_shapes=self._flat_output_shapes)
with ops.control_dependencies(
[gen_dataset_ops.make_iterator(ds_variant, resource)]):
return gen_dataset_ops.iterator_to_string_handle(resource)
def __init__(self, iterator_resource, initializer, output_types,
output_shapes, output_classes):
"""Creates a new iterator from the given iterator resource.
Note: Most users will not call this initializer directly, and will
instead use `Dataset.make_initializable_iterator()` or
`Dataset.make_one_shot_iterator()`.
Args:
iterator_resource: A `tf.resource` scalar `tf.Tensor` representing the
iterator.
initializer: A `tf.Operation` that should be run to initialize this
iterator.
output_types: A (nested) structure of `tf.DType` objects corresponding to
each component of an element of this iterator.
output_shapes: A (nested) structure of `tf.TensorShape` objects
corresponding to each component of an element of this iterator.
output_classes: A (nested) structure of Python `type` objects
corresponding to each component of an element of this iterator.
Raises:
TypeError: If `output_types`, `output_shapes`, or `output_classes` is not
specified.
"""
self._iterator_resource = iterator_resource
self._initializer = initializer
if (output_types is None or output_shapes is None
or output_classes is None):
raise ValueError(
"All of `output_types`, `output_shapes`, and `output_classes` "
"must be specified to create an iterator. Got "
f"`output_types` = {output_types!r}, "
f"`output_shapes` = {output_shapes!r}, "
f"`output_classes` = {output_classes!r}.")
self._element_spec = structure.convert_legacy_structure(
output_types, output_shapes, output_classes)
self._flat_tensor_shapes = structure.get_flat_tensor_shapes(
self._element_spec)
self._flat_tensor_types = structure.get_flat_tensor_types(
self._element_spec)
self._string_handle = gen_dataset_ops.iterator_to_string_handle(
self._iterator_resource)
self._get_next_call_count = 0
ops.add_to_collection(GLOBAL_ITERATORS, self._iterator_resource)
def __init__(self, iterator_resource, initializer, output_types,
output_shapes, output_classes):
"""Creates a new iterator from the given iterator resource.
Note: Most users will not call this initializer directly, and will
instead use `Dataset.make_initializable_iterator()` or
`Dataset.make_one_shot_iterator()`.
Args:
iterator_resource: A `tf.resource` scalar `tf.Tensor` representing the
iterator.
initializer: A `tf.Operation` that should be run to initialize this
iterator.
output_types: A nested structure of `tf.DType` objects corresponding to
each component of an element of this iterator.
output_shapes: A nested structure of `tf.TensorShape` objects
corresponding to each component of an element of this iterator.
output_classes: A nested structure of Python `type` objects corresponding
to each component of an element of this iterator.
"""
self._iterator_resource = iterator_resource
self._initializer = initializer
if (output_types is None or output_shapes is None
or output_classes is None):
raise ValueError(
"If `structure` is not specified, all of "
"`output_types`, `output_shapes`, and `output_classes`"
" must be specified.")
# pylint: disable=protected-access
self._structure = structure_lib.Structure._from_legacy_structure(
output_types, output_shapes, output_classes)
# pylint: enable=protected-access
self._string_handle = gen_dataset_ops.iterator_to_string_handle(
self._iterator_resource)
self._get_next_call_count = 0
ops.add_to_collection(GLOBAL_ITERATORS, self._iterator_resource)
def __init__(self, dataset):
"""Creates a new iterator over the given dataset.
For example:
```python
dataset = tf.data.Dataset.range(4)
for x in Iterator(dataset):
print(x)
```
Tensors produced will be placed on the device on which this iterator object
was created.
Args:
dataset: A `tf.data.Dataset` object.
Raises:
RuntimeError: When invoked without eager execution enabled.
"""
if not context.in_eager_mode():
raise RuntimeError(
"{} objects can only be used when eager execution is enabled, use "
"tf.data.Dataset.make_iterator or "
"tf.data.Dataset.make_one_shot_iterator for graph construction"
.format(type(self)))
with ops.device("/device:CPU:0"):
ds_variant = dataset._as_variant_tensor() # pylint: disable=protected-access
self._output_types = dataset.output_types
self._output_shapes = dataset.output_shapes
self._flat_output_types = nest.flatten(dataset.output_types)
self._flat_output_shapes = nest.flatten(dataset.output_shapes)
self._resource = gen_dataset_ops.iterator(
container="",
shared_name=_generate_shared_name("eager_iterator"),
output_types=self._flat_output_types,
output_shapes=self._flat_output_shapes)
gen_dataset_ops.make_iterator(ds_variant, self._resource)
# Delete the resource when this object is deleted
self._resource_deleter = resource_variable_ops.EagerResourceDeleter(
handle=self._resource, handle_device="/device:CPU:0")
self._device = context.context().device_name
self._buffer_resource_handle = None
if not context.context().device_spec.device_type:
is_remote_device = False
else:
is_remote_device = context.context(
).device_spec.device_type != "CPU"
if is_remote_device:
with ops.device("/device:CPU:0"):
iter_string_handle = gen_dataset_ops.iterator_to_string_handle(
self._resource)
@function.Defun(dtypes.string)
def remote_fn(h):
remote_iterator = iterator_ops.Iterator.from_string_handle(
h, self._output_types, self._output_shapes)
return remote_iterator.get_next()
remote_fn.add_to_graph(None)
target = constant_op.constant("/device:CPU:0")
with ops.device(self._device):
self._buffer_resource_handle = prefetching_ops.function_buffering_resource(
string_arg=iter_string_handle,
f=remote_fn,
target_device=target,
buffer_size=10,
thread_pool_size=1,
container="",
shared_name=_generate_shared_name(
"function_buffer_resource"))
self._buffer_resource_deleter = resource_variable_ops.EagerResourceDeleter(
handle=self._buffer_resource_handle,
handle_device=self._device)
def __init__(self, dataset):
"""Creates a new iterator over the given dataset.
For example:
```python
dataset = tf.data.Dataset.range(4)
for x in Iterator(dataset):
print(x)
```
Tensors produced will be placed on the device on which this iterator object
was created.
Args:
dataset: A `tf.data.Dataset` object.
Raises:
RuntimeError: When invoked without eager execution enabled.
"""
if not context.in_eager_mode():
raise RuntimeError(
"{} objects can only be used when eager execution is enabled, use "
"tf.data.Dataset.make_iterator or "
"tf.data.Dataset.make_one_shot_iterator for graph construction".
format(type(self)))
with ops.device("/device:CPU:0"):
ds_variant = dataset._as_variant_tensor() # pylint: disable=protected-access
self._output_types = dataset.output_types
self._output_shapes = dataset.output_shapes
self._flat_output_types = nest.flatten(dataset.output_types)
self._flat_output_shapes = nest.flatten(dataset.output_shapes)
self._resource = gen_dataset_ops.iterator(
container="",
shared_name=_generate_shared_name("eager_iterator"),
output_types=self._flat_output_types,
output_shapes=self._flat_output_shapes)
gen_dataset_ops.make_iterator(ds_variant, self._resource)
# Delete the resource when this object is deleted
self._resource_deleter = resource_variable_ops.EagerResourceDeleter(
handle=self._resource, handle_device="/device:CPU:0")
self._device = context.context().device_name
self._buffer_resource_handle = None
if not context.context().device_spec.device_type:
is_remote_device = False
else:
is_remote_device = context.context().device_spec.device_type != "CPU"
if is_remote_device:
with ops.device("/device:CPU:0"):
iter_string_handle = gen_dataset_ops.iterator_to_string_handle(
self._resource)
@function.Defun(dtypes.string)
def remote_fn(h):
remote_iterator = iterator_ops.Iterator.from_string_handle(
h, self._output_types, self._output_shapes)
return remote_iterator.get_next()
remote_fn.add_to_graph(None)
target = constant_op.constant("/device:CPU:0")
with ops.device(self._device):
self._buffer_resource_handle = prefetching_ops.function_buffering_resource(
string_arg=iter_string_handle,
f=remote_fn,
target_device=target,
buffer_size=10,
thread_pool_size=1,
container="",
shared_name=_generate_shared_name("function_buffer_resource"))
self._buffer_resource_deleter = resource_variable_ops.EagerResourceDeleter(
handle=self._buffer_resource_handle, handle_device=self._device)
