def tf_shapes(self):
"""
:return: a dictionary of sampler output tensor shapes
"""
output_shapes = nest.map_structure_up_to(
self.tf_dtypes, tf.TensorShape, self.shapes)
return output_shapes
def testMapStructureUpTo(self):
ab_tuple = collections.namedtuple("ab_tuple", "a, b")
op_tuple = collections.namedtuple("op_tuple", "add, mul")
inp_val = ab_tuple(a=2, b=3)
inp_ops = ab_tuple(a=op_tuple(add=1, mul=2), b=op_tuple(add=2, mul=3))
out = nest.map_structure_up_to(
inp_val, lambda val, ops: (val + ops.add) * ops.mul, inp_val, inp_ops)
self.assertEqual(out.a, 6)
self.assertEqual(out.b, 15)
data_list = ((2, 4, 6, 8), ((1, 3, 5, 7, 9), (3, 5, 7)))
name_list = ("evens", ("odds", "primes"))
out = nest.map_structure_up_to(
name_list, lambda name, sec: "first_{}_{}".format(len(sec), name),
name_list, data_list)
self.assertEqual(out, ("first_4_evens", ("first_5_odds", "first_3_primes")))
def from_string_handle(string_handle, output_types, output_shapes=None):
"""Creates a new, uninitialized `Iterator` based on the given handle.
This method allows you to define a "feedable" iterator where you can choose
between concrete iterators by feeding a value in a @{tf.Session.run} call.
In that case, `string_handle` would a @{tf.placeholder}, and you would feed
it with the value of @{tf.data.Iterator.string_handle} in each step.
For example, if you had two iterators that marked the current position in
a training dataset and a test dataset, you could choose which to use in
each step as follows:
```python
train_iterator = tf.data.Dataset(...).make_one_shot_iterator()
train_iterator_handle = sess.run(train_iterator.string_handle())
test_iterator = tf.data.Dataset(...).make_one_shot_iterator()
test_iterator_handle = sess.run(test_iterator.string_handle())
handle = tf.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(
handle, train_iterator.output_types)
next_element = iterator.get_next()
loss = f(next_element)
train_loss = sess.run(loss, feed_dict={handle: train_iterator_handle})
test_loss = sess.run(loss, feed_dict={handle: test_iterator_handle})
```
Args:
string_handle: A scalar `tf.Tensor` of type `tf.string` that evaluates
to a handle produced by the `Iterator.string_handle()` method.
output_types: A nested structure of `tf.DType` (or `tf.data.SparseType`)
objects corresponding to each `tf.Tensor` (or `tf.SparseTensor`)
component of an element of this dataset.
output_shapes: (Optional.) A nested structure of `tf.TensorShape` objects
corresponding to each component of an element of this dataset. If
omitted, each component will have an unconstrainted shape.
Returns:
An `Iterator`.
"""
output_types = nest.map_structure(dtypes.as_dtype, output_types)
if output_shapes is None:
output_shapes = nest.map_structure(
lambda _: tensor_shape.TensorShape(None), output_types)
else:
output_shapes = nest.map_structure_up_to(
output_types, tensor_shape.as_shape, output_shapes)
nest.assert_same_structure(output_types, output_shapes)
string_handle = ops.convert_to_tensor(string_handle, dtype=dtypes.string)
iterator_resource = gen_dataset_ops.iterator_from_string_handle(
string_handle,
output_types=nest.flatten(sparse.unwrap_sparse_types(output_types)),
output_shapes=nest.flatten(output_shapes))
return Iterator(iterator_resource, None, output_types, output_shapes)
def set_spatial_shape(self, spatial_window, source_names=None):
"""
Set all spatial window of the window.
spatial_window should be a dictionary of window sizes tuples
or single window size tuple. In the latter case the size
will be used by all output image windows.
:param spatial_window: tuple of integers specifying new shape
:param source_names: list/dictionary of input source names
:return:
"""
win_sizes = copy.deepcopy(spatial_window)
if isinstance(spatial_window, dict):
for name in list(spatial_window):
window_size = spatial_window[name]
if isinstance(window_size,
(ParserNamespace, argparse.Namespace)):
window_size = vars(window_size)
if not isinstance(window_size, dict):
win_sizes[name] = tuple(window_size)
elif 'spatial_window_size' in window_size:
win_sizes[name] = tuple(
window_size['spatial_window_size'])
else:
raise ValueError(
'window_sizes should be a nested dictionary')
elif isinstance(spatial_window, (list, tuple)):
# list or tuple of single window sizes
win_sizes = {name: spatial_window for name in list(self._dtypes)}
# complete window shapes based on user input and input_image sizes
if source_names:
spatial_shapes = _read_window_sizes(source_names, win_sizes)
else:
try:
spatial_shapes = {}
for name in list(self._dtypes):
spatial_shapes[name] = \
tuple(int(win_size) for win_size in win_sizes[name])
except ValueError:
tf.logging.fatal("spatial window should be an array of int")
raise
spatial_shapes = nest.map_structure_up_to(
self._dtypes, tuple, spatial_shapes)
self._shapes = {
name: _complete_partial_window_sizes(spatial_shapes[name],
self._shapes[name])
for name in list(self._shapes)}
# update based on the latest spatial shapes
self.has_dynamic_shapes = self._check_dynamic_shapes()
if self._placeholders_dict is not None:
self._update_placeholders_dict(n_samples=self.n_samples)
def __init__(self, dataset, output_types, output_shapes=None):
"""Creates a new dataset with the given output types and shapes.
The given `dataset` must have a structure that is convertible:
* `dataset.output_types` must be the same as `output_types` module nesting.
* Each shape in `dataset.output_shapes` must be compatible with each shape
in `output_shapes` (if given).
Note: This helper permits "unsafe casts" for shapes, equivalent to using
`tf.Tensor.set_shape()` where domain-specific knowledge is available.
Args:
dataset: A `Dataset` object.
output_types: A nested structure of `tf.DType` objects.
output_shapes: (Optional.) A nested structure of `tf.TensorShape` objects.
If omitted, the shapes will be inherited from `dataset`.
Raises:
ValueError: If either `output_types` or `output_shapes` is not compatible
with the structure of `dataset`.
"""
super(_RestructuredDataset, self).__init__()
self._dataset = dataset
# Validate that the types are compatible.
output_types = nest.map_structure(dtypes.as_dtype, output_types)
flat_original_types = nest.flatten(dataset.output_types)
flat_new_types = nest.flatten(output_types)
if flat_original_types != flat_new_types:
raise ValueError(
"Dataset with output types %r cannot be restructured to have output "
"types %r" % (dataset.output_types, output_types))
self._output_types = output_types
if output_shapes is None:
# Inherit shapes from the original `dataset`.
self._output_shapes = nest.pack_sequence_as(output_types,
nest.flatten(
dataset.output_shapes))
else:
# Validate that the shapes are compatible.
nest.assert_same_structure(output_types, output_shapes)
flat_original_shapes = nest.flatten(dataset.output_shapes)
flat_new_shapes = nest.flatten_up_to(output_types, output_shapes)
for original_shape, new_shape in zip(flat_original_shapes,
flat_new_shapes):
if not original_shape.is_compatible_with(new_shape):
raise ValueError(
"Dataset with output shapes %r cannot be restructured to have "
"incompatible output shapes %r" % (dataset.output_shapes,
output_shapes))
self._output_shapes = nest.map_structure_up_to(
output_types, tensor_shape.as_shape, output_shapes)
def __init__(self,
pipeline,
output_dtypes=None,
output_shapes=None,
*,
batch_size=1,
num_threads=4,
device_id=0,
exec_separated=False,
prefetch_queue_depth=2,
cpu_prefetch_queue_depth=2,
gpu_prefetch_queue_depth=2,
dtypes=None,
shapes=None):
output_shapes = self._handle_deprecation(output_shapes, shapes,
"shapes")
output_dtypes = self._handle_deprecation(output_dtypes, dtypes,
"dtypes")
if not self._check_output_dtypes(output_dtypes):
raise TypeError(("`output_dtypes` should be provided as single tf.DType value " +
"or a tuple of tf.DType values. Got value `{}` of type `{}`.") \
.format(output_dtypes, type(output_dtypes)))
if output_shapes is None:
output_shapes = nest.map_structure(
lambda _: tensor_shape.TensorShape(None), output_dtypes)
else:
output_shapes = nest.map_structure_up_to(
output_dtypes, tensor_shape.as_shape, output_shapes)
if not isinstance(output_dtypes, tuple):
output_dtypes = (output_dtypes, )
output_shapes = (output_shapes, )
output_classes = nest.map_structure(lambda _: ops.Tensor,
output_dtypes)
self._pipeline = serialize_pipeline(pipeline)
self._batch_size = batch_size
self._num_threads = num_threads
self._device_id = device_id
self._exec_separated = exec_separated
self._prefetch_queue_depth = prefetch_queue_depth
self._cpu_prefetch_queue_depth = cpu_prefetch_queue_depth
self._gpu_prefetch_queue_depth = gpu_prefetch_queue_depth
self._output_shapes = output_shapes
self._output_dtypes = output_dtypes
self._structure = structure.convert_legacy_structure(
self._output_dtypes, self._output_shapes, output_classes)
super(_DALIDatasetV2, self).__init__(self._as_variant_tensor())
def __init__(self, dataset, output_types, output_shapes=None):
"""Creates a new dataset with the given output types and shapes.
The given `dataset` must have a structure that is convertible:
* `dataset.output_types` must be the same as `output_types` module nesting.
* Each shape in `dataset.output_shapes` must be compatible with each shape
in `output_shapes` (if given).
Note: This helper permits "unsafe casts" for shapes, equivalent to using
`tf.Tensor.set_shape()` where domain-specific knowledge is available.
Args:
dataset: A `Dataset` object.
output_types: A nested structure of `tf.DType` objects.
output_shapes: (Optional.) A nested structure of `tf.TensorShape` objects.
If omitted, the shapes will be inherited from `dataset`.
Raises:
ValueError: If either `output_types` or `output_shapes` is not compatible
with the structure of `dataset`.
"""
super(_RestructuredDataset, self).__init__()
self._dataset = dataset
# Validate that the types are compatible.
output_types = nest.map_structure(dtypes.as_dtype, output_types)
flat_original_types = nest.flatten(dataset.output_types)
flat_new_types = nest.flatten(output_types)
if flat_original_types != flat_new_types:
raise ValueError(
"Dataset with output types %r cannot be restructured to have output "
"types %r" % (dataset.output_types, output_types))
self._output_types = output_types
if output_shapes is None:
# Inherit shapes from the original `dataset`.
self._output_shapes = nest.pack_sequence_as(
output_types, nest.flatten(dataset.output_shapes))
else:
# Validate that the shapes are compatible.
nest.assert_same_structure(output_types, output_shapes)
flat_original_shapes = nest.flatten(dataset.output_shapes)
flat_new_shapes = nest.flatten_up_to(output_types, output_shapes)
for original_shape, new_shape in zip(flat_original_shapes,
flat_new_shapes):
if not original_shape.is_compatible_with(new_shape):
raise ValueError(
"Dataset with output shapes %r cannot be restructured to have "
"incompatible output shapes %r" %
(dataset.output_shapes, output_shapes))
self._output_shapes = nest.map_structure_up_to(
output_types, tensor_shape.as_shape, output_shapes)
def __init__(self, input_dataset, batch_size, padded_shapes, padding_values):
"""Initialize `PrependFromQueueAndPaddedBatchDataset`."""
super(_PrependFromQueueAndPaddedBatchDataset, self).__init__()
if sparse.any_sparse(input_dataset.output_classes):
raise TypeError(
"Batching of padded sparse tensors is not currently supported")
self._input_dataset = input_dataset
self._batch_size = ops.convert_to_tensor(
batch_size, dtype=dtypes.int64, name="batch_size")
# pylint: disable=protected-access
if padded_shapes is None:
self._padded_shapes = nest.map_structure(
dataset_ops._partial_shape_to_tensor, input_dataset.output_shapes)
else:
self._padded_shapes = nest.map_structure_up_to(
input_dataset.output_shapes, dataset_ops._partial_shape_to_tensor,
padded_shapes)
padding_values = (
padding_values if padding_values is not None else
dataset_ops._default_padding(input_dataset))
self._padding_values = nest.map_structure_up_to(
input_dataset.output_shapes, dataset_ops._padding_value_to_tensor,
padding_values, input_dataset.output_types)
def __init__(self, variant_tensor, output_shapes, output_types,
output_classes):
# TODO(b/110122868): Consolidate the structure validation logic with the
# similar logic in `Iterator.from_structure()` and
# `Dataset.from_generator()`.
output_types = nest.map_structure(dtypes.as_dtype, output_types)
output_shapes = nest.map_structure_up_to(
output_types, tensor_shape.as_shape, output_shapes)
nest.assert_same_structure(output_types, output_shapes)
nest.assert_same_structure(output_types, output_classes)
self._variant_tensor = variant_tensor
self._output_shapes = output_shapes
self._output_types = output_types
self._output_classes = output_classes
def __init__(self, variant_tensor, output_shapes, output_types,
output_classes):
# TODO(b/110122868): Consolidate the structure validation logic with the
# similar logic in `Iterator.from_structure()` and
# `Dataset.from_generator()`.
output_types = nest.map_structure(dtypes.as_dtype, output_types)
output_shapes = nest.map_structure_up_to(
output_types, tensor_shape.as_shape, output_shapes)
nest.assert_same_structure(output_types, output_shapes)
nest.assert_same_structure(output_types, output_classes)
self._variant_tensor = variant_tensor
self._output_shapes = output_shapes
self._output_types = output_types
self._output_classes = output_classes
def __init__(self, input_dataset, batch_size, padded_shapes, padding_values):
"""Initialize `PrependFromQueueAndPaddedBatchDataset`."""
super(_PrependFromQueueAndPaddedBatchDataset, self).__init__()
if sparse.any_sparse(input_dataset.output_classes):
raise TypeError(
"Batching of padded sparse tensors is not currently supported")
self._input_dataset = input_dataset
self._batch_size = ops.convert_to_tensor(
batch_size, dtype=dtypes.int64, name="batch_size")
# pylint: disable=protected-access
if padded_shapes is None:
self._padded_shapes = nest.map_structure(
dataset_ops._partial_shape_to_tensor, input_dataset.output_shapes)
else:
self._padded_shapes = nest.map_structure_up_to(
input_dataset.output_shapes, dataset_ops._partial_shape_to_tensor,
padded_shapes)
padding_values = (
padding_values if padding_values is not None else
dataset_ops._default_padding(input_dataset))
self._padding_values = nest.map_structure_up_to(
input_dataset.output_shapes, dataset_ops._padding_value_to_tensor,
padding_values, input_dataset.output_types)
def from_data_reader_properties(cls,
source_names,
image_shapes,
image_dtypes,
window_sizes=None,
allow_dynamic=False):
"""
Create a window instance with input data properties
each property is grouped into dict, with pairs of
image_name: data_value. Some input images is a
concatenated data array from multiple data sources.
example of input::
source_names={
'image': (u'modality1', u'modality2'),
'label': (u'modality3',)},
image_shapes={
'image': (192, 160, 192, 1, 2),
'label': (192, 160, 192, 1, 1)},
image_dtypes={
'image': tf.float32,
'label': tf.float32},
window_sizes={
'image': (10, 10, 2),
'label': (10, 10, 2)}
the ``window_sizes`` can also be::
window_sizes={
'modality1': (10, 10, 2),
'modality3': (10, 10, 2)}
or using a nested dictionary with 'spatial_window_size' (deprecating)::
window_sizes={
'modality1': {'spatial_window_size': (10, 10, 2)},
'modality2': {'spatial_window_size': (10, 10, 2)},
'modality3': {'spatial_window_size': (5, 5, 1)}}
see ``niftynet.io.ImageReader`` for more details.
:param source_names: input image names
:param image_shapes: tuple of image window shapes
:param image_dtypes: tuple of image window data types
:param window_sizes: window sizes for the image image
:param allow_dynamic: if True, window_sizes negative or 0 indicates
dynamic window sizes; . Otherwise the dynamic sizes will be fixed
as the image shapes; this assumes the same image size across the
dataset.
:return: an ImageWindow instance
"""
try:
image_shapes = nest.map_structure_up_to(
image_dtypes, tuple, image_shapes)
except KeyError:
tf.logging.fatal('window_sizes wrong format %s', window_sizes)
raise
# create ImageWindow instance
window_instance = cls(shapes=image_shapes, dtypes=image_dtypes)
if not window_sizes:
# image window sizes not specified, defaulting to image sizes.
return window_instance
window_instance.set_spatial_shape(window_sizes, source_names)
if not allow_dynamic:
full_shape = window_instance.match_image_shapes(image_shapes)
window_instance.set_spatial_shape(full_shape)
return window_instance
def _from_components(self, value):
return nest.map_structure_up_to(self._nested_structure,
lambda t, v: t._from_components(v),
self._nested_structure, value)
def from_generator(generator, output_types, output_shapes=None):
"""Creates a `Dataset` whose elements are generated by `generator`.
The `generator` argument must be a callable object that returns
an object that support the `iter()` protocol (e.g. a generator function).
The elements generated by `generator` must be compatible with the given
`output_types` and (optional) `output_shapes` arguments.
For example:
```python
import itertools
def gen():
for i in itertools.count(1):
yield (i, [1] * i)
ds = Dataset.from_generator(
gen, (tf.int64, tf.int64), (tf.TensorShape([]), tf.TensorShape([None])))
value = ds.make_one_shot_iterator().get_next()
sess.run(value) # (1, array([1]))
sess.run(value) # (2, array([1, 1]))
```
Args:
generator: A callable object that takes no arguments and returns an
object that supports the `iter()` protocol.
output_types: A nested structure of `tf.DType` objects corresponding to
each component of an element yielded by `generator`.
output_shapes: (Optional.) A nested structure of `tf.TensorShape`
objects corresponding to each component of an element yielded by
`generator`.
Returns:
A `Dataset`.
"""
if not callable(generator):
raise TypeError("`generator` must be callable.")
if output_shapes is None:
output_shapes = nest.map_structure(
lambda _: tensor_shape.TensorShape(None), output_types)
else:
output_shapes = nest.map_structure_up_to(
output_types, tensor_shape.as_shape, output_shapes)
flattened_types = nest.flatten(output_types)
flattened_shapes = nest.flatten(output_shapes)
generator_state = dataset_ops.Dataset._GeneratorState(generator)
def get_iterator_id_map_fn(unused_dummy):
"""Creates a unique `iterator_id` for each pass over the dataset.
The "iterator_id" disambiguates between multiple concurrently
existing iterators.
Args:
unused_dummy: Ignored value.
Returns:
A `tf.int64` tensor whose value uniquely identifies an iterator in
`generator_state`.
"""
return script_ops.py_func(
generator_state.get_next_id, [], dtypes.int64, stateful=True)
def generator_map_fn(iterator_id_t):
"""Generates the next element from iterator with ID `iterator_id_t`.
We map this function across an infinite repetition of the
`iterator_id_t`, and raise `StopIteration` to terminate the iteration.
Args:
iterator_id_t: A `tf.int64` tensor whose value uniquely identifies
the iterator in `generator_state` from which to generate an element.
Returns:
A nested structure of tensors representing an element from the iterator.
"""
def generator_py_func(iterator_id):
"""A `py_func` that will be called to invoke the iterator."""
try:
values = next(generator_state.get_iterator(iterator_id))
except StopIteration:
generator_state.iterator_completed(iterator_id)
raise StopIteration("Iteration finished.")
# Use the same _convert function from the py_func() implementation to
# convert the returned values to arrays early, so that we can inspect
# their values.
# pylint: disable=protected-access
ret_arrays = [
script_ops.FuncRegistry._convert(ret, dtype=dtype.as_numpy_dtype)
for ret, dtype in zip(nest.flatten_up_to(output_types, values),
flattened_types)
]
# pylint: enable=protected-access
# Additional type and shape checking to ensure that the components
# of the generated element match the `output_types` and `output_shapes`
# arguments.
for (ret_array, expected_dtype, expected_shape) in zip(
ret_arrays, flattened_types, flattened_shapes):
if ret_array.dtype != expected_dtype.as_numpy_dtype:
raise TypeError(
"`generator` yielded an element of type %s where an element "
"of type %s was expected." % (ret_array.dtype,
expected_dtype.as_numpy_dtype))
if not expected_shape.is_compatible_with(ret_array.shape):
raise ValueError(
"`generator` yielded an element of shape %s where an element "
"of shape %s was expected." % (ret_array.shape, expected_shape))
return ret_arrays
flat_values = script_ops.py_func(
generator_py_func, [iterator_id_t], flattened_types, stateful=True)
# The `py_func()` op drops the inferred shapes, so we add them back in
# here.
if output_shapes is not None:
for ret_t, shape in zip(flat_values, flattened_shapes):
ret_t.set_shape(shape)
return nest.pack_sequence_as(output_types, flat_values)
# This function associates each traversal of `generator` with a unique
# iterator ID.
def flat_map_fn(iterator_id_t):
# First, generate an infinite dataset containing the iterator ID repeated
# forever.
repeated_id = Dataset.from_tensors(iterator_id_t).repeat(None)
# The `generator_map_fn` gets the next element from the iterator with the
# relevant ID, and raises StopIteration when that iterator contains no
# more elements.
return repeated_id.map(generator_map_fn)
# A single-element dataset that, each time it is evaluated, contains a
# freshly-generated and unique (for the returned dataset) int64
# ID that will be used to identify the appropriate Python state, which
# is encapsulated in `generator_state`, and captured in
# `get_iterator_id_map_fn`.
dummy = 0
id_dataset = Dataset.from_tensors(dummy).map(get_iterator_id_map_fn)
# A dataset that contains all of the elements generated by a
# single iterator created from `generator`, identified by the
# iterator ID contained in `id_dataset`. Lifting the iteration
# into a flat_map here enables multiple repetitions and/or nested
# versions of the returned dataset to be created, because it forces
# the generation of a new ID for each version.
return id_dataset.flat_map(flat_map_fn)
def from_structure(output_types,
output_shapes=None,
shared_name=None,
output_classes=None):
"""Creates a new, uninitialized `Iterator` with the given structure.
This iterator-constructing method can be used to create an iterator that
is reusable with many different datasets.
The returned iterator is not bound to a particular dataset, and it has
no `initializer`. To initialize the iterator, run the operation returned by
`Iterator.make_initializer(dataset)`.
The following is an example
```python
iterator = Iterator.from_structure(tf.int64, tf.TensorShape([]))
dataset_range = Dataset.range(10)
range_initializer = iterator.make_initializer(dataset_range)
dataset_evens = dataset_range.filter(lambda x: x % 2 == 0)
evens_initializer = iterator.make_initializer(dataset_evens)
# Define a model based on the iterator; in this example, the model_fn
# is expected to take scalar tf.int64 Tensors as input (see
# the definition of 'iterator' above).
prediction, loss = model_fn(iterator.get_next())
# Train for `num_epochs`, where for each epoch, we first iterate over
# dataset_range, and then iterate over dataset_evens.
for _ in range(num_epochs):
# Initialize the iterator to `dataset_range`
sess.run(range_initializer)
while True:
try:
pred, loss_val = sess.run([prediction, loss])
except tf.errors.OutOfRangeError:
break
# Initialize the iterator to `dataset_evens`
sess.run(evens_initializer)
while True:
try:
pred, loss_val = sess.run([prediction, loss])
except tf.errors.OutOfRangeError:
break
```
Args:
output_types: A nested structure of `tf.DType` objects corresponding to
each component of an element of this dataset.
output_shapes: (Optional.) A nested structure of `tf.TensorShape` objects
corresponding to each component of an element of this dataset. If
omitted, each component will have an unconstrainted shape.
shared_name: (Optional.) If non-empty, this iterator will be shared under
the given name across multiple sessions that share the same devices
(e.g. when using a remote server).
output_classes: (Optional.) A nested structure of Python `type` objects
corresponding to each component of an element of this iterator. If
omitted, each component is assumed to be of type `tf.Tensor`.
Returns:
An `Iterator`.
Raises:
TypeError: If the structures of `output_shapes` and `output_types` are
not the same.
"""
output_types = nest.map_structure(dtypes.as_dtype, output_types)
if output_shapes is None:
output_shapes = nest.map_structure(
lambda _: tensor_shape.TensorShape(None), output_types)
else:
output_shapes = nest.map_structure_up_to(
output_types, tensor_shape.as_shape, output_shapes)
if output_classes is None:
output_classes = nest.map_structure(lambda _: ops.Tensor, output_types)
nest.assert_same_structure(output_types, output_shapes)
if shared_name is None:
shared_name = ""
iterator_resource = gen_dataset_ops.iterator(
container="",
shared_name=shared_name,
output_types=nest.flatten(
sparse.as_dense_types(output_types, output_classes)),
output_shapes=nest.flatten(
sparse.as_dense_shapes(output_shapes, output_classes)))
return Iterator(iterator_resource, None, output_types, output_shapes,
output_classes)
def from_generator(generator, output_types, output_shapes=None):
"""Creates a `Dataset` whose elements are generated by `generator`.
The `generator` argument must be a callable object that returns
an object that support the `iter()` protocol (e.g. a generator function).
The elements generated by `generator` must be compatible with the given
`output_types` and (optional) `output_shapes` arguments.
For example:
```python
import itertools
def gen():
for i in itertools.count(1):
yield (i, [1] * i)
ds = Dataset.from_generator(
gen, (tf.int64, tf.int64), (tf.TensorShape([]), tf.TensorShape([None])))
value = ds.make_one_shot_iterator().get_next()
sess.run(value) # (1, array([1]))
sess.run(value) # (2, array([1, 1]))
```
Args:
generator: A callable object that takes no arguments and returns an
object that supports the `iter()` protocol.
output_types: A nested structure of `tf.DType` objects corresponding to
each component of an element yielded by `generator`.
output_shapes: (Optional.) A nested structure of `tf.TensorShape`
objects corresponding to each component of an element yielded by
`generator`.
Returns:
A `Dataset`.
"""
if not callable(generator):
raise TypeError("`generator` must be callable.")
if output_shapes is None:
output_shapes = nest.map_structure(
lambda _: tensor_shape.TensorShape(None), output_types)
else:
output_shapes = nest.map_structure_up_to(output_types,
tensor_shape.as_shape,
output_shapes)
flattened_types = nest.flatten(output_types)
flattened_shapes = nest.flatten(output_shapes)
generator_state = dataset_ops.Dataset._GeneratorState(generator)
def get_iterator_id_map_fn(unused_dummy):
"""Creates a unique `iterator_id` for each pass over the dataset.
The "iterator_id" disambiguates between multiple concurrently
existing iterators.
Args:
unused_dummy: Ignored value.
Returns:
A `tf.int64` tensor whose value uniquely identifies an iterator in
`generator_state`.
"""
return script_ops.py_func(generator_state.get_next_id, [],
dtypes.int64,
stateful=True)
def generator_map_fn(iterator_id_t):
"""Generates the next element from iterator with ID `iterator_id_t`.
We map this function across an infinite repetition of the
`iterator_id_t`, and raise `StopIteration` to terminate the iteration.
Args:
iterator_id_t: A `tf.int64` tensor whose value uniquely identifies
the iterator in `generator_state` from which to generate an element.
Returns:
A nested structure of tensors representing an element from the iterator.
"""
def generator_py_func(iterator_id):
"""A `py_func` that will be called to invoke the iterator."""
try:
values = next(generator_state.get_iterator(iterator_id))
except StopIteration:
generator_state.iterator_completed(iterator_id)
raise StopIteration("Iteration finished.")
# Use the same _convert function from the py_func() implementation to
# convert the returned values to arrays early, so that we can inspect
# their values.
# pylint: disable=protected-access
ret_arrays = [
script_ops.FuncRegistry._convert(
ret, dtype=dtype.as_numpy_dtype) for ret, dtype in zip(
nest.flatten_up_to(output_types, values),
flattened_types)
]
# pylint: enable=protected-access
# Additional type and shape checking to ensure that the components
# of the generated element match the `output_types` and `output_shapes`
# arguments.
for (ret_array, expected_dtype,
expected_shape) in zip(ret_arrays, flattened_types,
flattened_shapes):
if ret_array.dtype != expected_dtype.as_numpy_dtype:
raise TypeError(
"`generator` yielded an element of type %s where an element "
"of type %s was expected." %
(ret_array.dtype, expected_dtype.as_numpy_dtype))
if not expected_shape.is_compatible_with(ret_array.shape):
raise ValueError(
"`generator` yielded an element of shape %s where an element "
"of shape %s was expected." %
(ret_array.shape, expected_shape))
return ret_arrays
flat_values = script_ops.py_func(generator_py_func,
[iterator_id_t],
flattened_types,
stateful=True)
# The `py_func()` op drops the inferred shapes, so we add them back in
# here.
if output_shapes is not None:
for ret_t, shape in zip(flat_values, flattened_shapes):
ret_t.set_shape(shape)
return nest.pack_sequence_as(output_types, flat_values)
# This function associates each traversal of `generator` with a unique
# iterator ID.
def flat_map_fn(iterator_id_t):
# First, generate an infinite dataset containing the iterator ID repeated
# forever.
repeated_id = Dataset.from_tensors(iterator_id_t).repeat(None)
# The `generator_map_fn` gets the next element from the iterator with the
# relevant ID, and raises StopIteration when that iterator contains no
# more elements.
return repeated_id.map(generator_map_fn)
# A single-element dataset that, each time it is evaluated, contains a
# freshly-generated and unique (for the returned dataset) int64
# ID that will be used to identify the appropriate Python state, which
# is encapsulated in `generator_state`, and captured in
# `get_iterator_id_map_fn`.
dummy = 0
id_dataset = Dataset.from_tensors(dummy).map(get_iterator_id_map_fn)
# A dataset that contains all of the elements generated by a
# single iterator created from `generator`, identified by the
# iterator ID contained in `id_dataset`. Lifting the iteration
# into a flat_map here enables multiple repetitions and/or nested
# versions of the returned dataset to be created, because it forces
# the generation of a new ID for each version.
return id_dataset.flat_map(flat_map_fn)
def __init__(self,
pipeline,
output_dtypes=None,
output_shapes=None,
fail_on_device_mismatch=True,
*,
input_datasets=None,
batch_size=1,
num_threads=4,
device_id=0,
exec_separated=False,
prefetch_queue_depth=2,
cpu_prefetch_queue_depth=2,
gpu_prefetch_queue_depth=2,
dtypes=None,
shapes=None):
output_shapes = self._handle_deprecation(output_shapes, shapes,
"shapes")
output_dtypes = self._handle_deprecation(output_dtypes, dtypes,
"dtypes")
if not self._check_dtypes(output_dtypes, tf.DType):
raise TypeError(
"`output_dtypes` should be provided as single tf.DType value "
f"or a tuple of tf.DType values. Got value `{output_dtypes}` "
f"of the type `{type(output_dtypes)}`.")
if output_shapes is None:
output_shapes = nest.map_structure(
lambda _: tensor_shape.TensorShape(None), output_dtypes)
else:
output_shapes = nest.map_structure_up_to(
output_dtypes, tensor_shape.as_shape, output_shapes)
if not isinstance(output_dtypes, tuple):
output_dtypes = (output_dtypes, )
output_shapes = (output_shapes, )
output_classes = nest.map_structure(lambda _: ops.Tensor,
output_dtypes)
self._pipeline_instance = pipeline # keep the live Pipeline object
self._pipeline_serialized = serialize_pipeline(pipeline)
self._batch_size = batch_size
self._num_threads = num_threads
if device_id is None:
device_id = types.CPU_ONLY_DEVICE_ID
self._device_id = device_id
self._exec_separated = exec_separated
self._prefetch_queue_depth = prefetch_queue_depth
self._cpu_prefetch_queue_depth = cpu_prefetch_queue_depth
self._gpu_prefetch_queue_depth = gpu_prefetch_queue_depth
self._output_shapes = output_shapes
self._output_dtypes = output_dtypes
self._fail_on_device_mismatch = fail_on_device_mismatch
self._setup_inputs(input_datasets)
self._structure = structure.convert_legacy_structure(
self._output_dtypes, self._output_shapes, output_classes)
super(_DALIDatasetV2, self).__init__(self._as_variant_tensor())
def from_structure(output_types,
output_shapes=None,
shared_name=None,
output_classes=None):
"""Creates a new, uninitialized `Iterator` with the given structure.
This iterator-constructing method can be used to create an iterator that
is reusable with many different datasets.
The returned iterator is not bound to a particular dataset, and it has
no `initializer`. To initialize the iterator, run the operation returned by
`Iterator.make_initializer(dataset)`.
The following is an example
```python
iterator = Iterator.from_structure(tf.int64, tf.TensorShape([]))
dataset_range = Dataset.range(10)
range_initializer = iterator.make_initializer(dataset_range)
dataset_evens = dataset_range.filter(lambda x: x % 2 == 0)
evens_initializer = iterator.make_initializer(dataset_evens)
# Define a model based on the iterator; in this example, the model_fn
# is expected to take scalar tf.int64 Tensors as input (see
# the definition of 'iterator' above).
prediction, loss = model_fn(iterator.get_next())
# Train for `num_epochs`, where for each epoch, we first iterate over
# dataset_range, and then iterate over dataset_evens.
for _ in range(num_epochs):
# Initialize the iterator to `dataset_range`
sess.run(range_initializer)
while True:
try:
pred, loss_val = sess.run([prediction, loss])
except tf.errors.OutOfRangeError:
break
# Initialize the iterator to `dataset_evens`
sess.run(evens_initializer)
while True:
try:
pred, loss_val = sess.run([prediction, loss])
except tf.errors.OutOfRangeError:
break
```
Args:
output_types: A nested structure of `tf.DType` objects corresponding to
each component of an element of this dataset.
output_shapes: (Optional.) A nested structure of `tf.TensorShape` objects
corresponding to each component of an element of this dataset. If
omitted, each component will have an unconstrainted shape.
shared_name: (Optional.) If non-empty, this iterator will be shared under
the given name across multiple sessions that share the same devices
(e.g. when using a remote server).
output_classes: (Optional.) A nested structure of Python `type` objects
corresponding to each component of an element of this iterator. If
omitted, each component is assumed to be of type `tf.Tensor`.
Returns:
An `Iterator`.
Raises:
TypeError: If the structures of `output_shapes` and `output_types` are
not the same.
"""
output_types = nest.map_structure(dtypes.as_dtype, output_types)
if output_shapes is None:
output_shapes = nest.map_structure(
lambda _: tensor_shape.TensorShape(None), output_types)
else:
output_shapes = nest.map_structure_up_to(output_types,
tensor_shape.as_shape,
output_shapes)
if output_classes is None:
output_classes = nest.map_structure(lambda _: ops.Tensor,
output_types)
nest.assert_same_structure(output_types, output_shapes)
output_structure = structure.convert_legacy_structure(
output_types, output_shapes, output_classes)
if shared_name is None:
shared_name = ""
if _device_stack_is_empty():
with ops.device("/cpu:0"):
iterator_resource = gen_dataset_ops.iterator_v2(
container="",
shared_name=shared_name,
output_types=structure.get_flat_tensor_types(
output_structure),
output_shapes=structure.get_flat_tensor_shapes(
output_structure))
else:
iterator_resource = gen_dataset_ops.iterator_v2(
container="",
shared_name=shared_name,
output_types=structure.get_flat_tensor_types(output_structure),
output_shapes=structure.get_flat_tensor_shapes(
output_structure))
return Iterator(iterator_resource, None, output_types, output_shapes,
output_classes)
def from_string_handle(string_handle,
output_types,
output_shapes=None,
output_classes=None):
"""Creates a new, uninitialized `Iterator` based on the given handle.
This method allows you to define a "feedable" iterator where you can choose
between concrete iterators by feeding a value in a `tf.Session.run` call.
In that case, `string_handle` would be a `tf.compat.v1.placeholder`, and you
would
feed it with the value of `tf.data.Iterator.string_handle` in each step.
For example, if you had two iterators that marked the current position in
a training dataset and a test dataset, you could choose which to use in
each step as follows:
```python
train_iterator = tf.data.Dataset(...).make_one_shot_iterator()
train_iterator_handle = sess.run(train_iterator.string_handle())
test_iterator = tf.data.Dataset(...).make_one_shot_iterator()
test_iterator_handle = sess.run(test_iterator.string_handle())
handle = tf.compat.v1.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(
handle, train_iterator.output_types)
next_element = iterator.get_next()
loss = f(next_element)
train_loss = sess.run(loss, feed_dict={handle: train_iterator_handle})
test_loss = sess.run(loss, feed_dict={handle: test_iterator_handle})
```
Args:
string_handle: A scalar `tf.Tensor` of type `tf.string` that evaluates to
a handle produced by the `Iterator.string_handle()` method.
output_types: A nested structure of `tf.DType` objects corresponding to
each component of an element of this dataset.
output_shapes: (Optional.) A nested structure of `tf.TensorShape` objects
corresponding to each component of an element of this dataset. If
omitted, each component will have an unconstrainted shape.
output_classes: (Optional.) A nested structure of Python `type` objects
corresponding to each component of an element of this iterator. If
omitted, each component is assumed to be of type `tf.Tensor`.
Returns:
An `Iterator`.
"""
output_types = nest.map_structure(dtypes.as_dtype, output_types)
if output_shapes is None:
output_shapes = nest.map_structure(
lambda _: tensor_shape.TensorShape(None), output_types)
else:
output_shapes = nest.map_structure_up_to(output_types,
tensor_shape.as_shape,
output_shapes)
if output_classes is None:
output_classes = nest.map_structure(lambda _: ops.Tensor,
output_types)
nest.assert_same_structure(output_types, output_shapes)
output_structure = structure.convert_legacy_structure(
output_types, output_shapes, output_classes)
string_handle = ops.convert_to_tensor(string_handle,
dtype=dtypes.string)
if _device_stack_is_empty():
with ops.device("/cpu:0"):
iterator_resource = gen_dataset_ops.iterator_from_string_handle_v2(
string_handle,
output_types=structure.get_flat_tensor_types(
output_structure),
output_shapes=structure.get_flat_tensor_shapes(
output_structure))
else:
iterator_resource = gen_dataset_ops.iterator_from_string_handle_v2(
string_handle,
output_types=structure.get_flat_tensor_types(output_structure),
output_shapes=structure.get_flat_tensor_shapes(
output_structure))
return Iterator(iterator_resource, None, output_types, output_shapes,
output_classes)
def from_string_handle(string_handle,
output_types,
output_shapes=None,
output_classes=None):
"""Creates a new, uninitialized `Iterator` based on the given handle.
This method allows you to define a "feedable" iterator where you can choose
between concrete iterators by feeding a value in a `tf.Session.run` call.
In that case, `string_handle` would be a `tf.placeholder`, and you would
feed it with the value of `tf.data.Iterator.string_handle` in each step.
For example, if you had two iterators that marked the current position in
a training dataset and a test dataset, you could choose which to use in
each step as follows:
```python
train_iterator = tf.data.Dataset(...).make_one_shot_iterator()
train_iterator_handle = sess.run(train_iterator.string_handle())
test_iterator = tf.data.Dataset(...).make_one_shot_iterator()
test_iterator_handle = sess.run(test_iterator.string_handle())
handle = tf.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(
handle, train_iterator.output_types)
next_element = iterator.get_next()
loss = f(next_element)
train_loss = sess.run(loss, feed_dict={handle: train_iterator_handle})
test_loss = sess.run(loss, feed_dict={handle: test_iterator_handle})
```
Args:
string_handle: A scalar `tf.Tensor` of type `tf.string` that evaluates
to a handle produced by the `Iterator.string_handle()` method.
output_types: A nested structure of `tf.DType` objects corresponding to
each component of an element of this dataset.
output_shapes: (Optional.) A nested structure of `tf.TensorShape` objects
corresponding to each component of an element of this dataset. If
omitted, each component will have an unconstrainted shape.
output_classes: (Optional.) A nested structure of Python `type` objects
corresponding to each component of an element of this iterator. If
omitted, each component is assumed to be of type `tf.Tensor`.
Returns:
An `Iterator`.
"""
output_types = nest.map_structure(dtypes.as_dtype, output_types)
if output_shapes is None:
output_shapes = nest.map_structure(
lambda _: tensor_shape.TensorShape(None), output_types)
else:
output_shapes = nest.map_structure_up_to(
output_types, tensor_shape.as_shape, output_shapes)
if output_classes is None:
output_classes = nest.map_structure(lambda _: ops.Tensor, output_types)
nest.assert_same_structure(output_types, output_shapes)
output_structure = structure_lib.convert_legacy_structure(
output_types, output_shapes, output_classes)
string_handle = ops.convert_to_tensor(string_handle, dtype=dtypes.string)
# pylint: disable=protected-access
if compat.forward_compatible(2018, 8, 3):
if _device_stack_is_empty():
with ops.device("/cpu:0"):
iterator_resource = gen_dataset_ops.iterator_from_string_handle_v2(
string_handle,
output_types=output_structure._flat_types,
output_shapes=output_structure._flat_shapes)
else:
iterator_resource = gen_dataset_ops.iterator_from_string_handle_v2(
string_handle,
output_types=output_structure._flat_types,
output_shapes=output_structure._flat_shapes)
else:
iterator_resource = gen_dataset_ops.iterator_from_string_handle(
string_handle,
output_types=output_structure._flat_types,
output_shapes=output_structure._flat_shapes)
# pylint: enable=protected-access
return Iterator(iterator_resource, None, output_types, output_shapes,
output_classes)
def __init__(self,
dataset,
output_types,
output_shapes=None,
output_classes=None,
allow_unsafe_cast=False):
"""Creates a new dataset with the given output types and shapes.
The given `dataset` must have a structure that is convertible:
* `dataset.output_types` must be the same as `output_types` module nesting.
* Each shape in `dataset.output_shapes` must be compatible with each shape
in `output_shapes` (if given).
Note: This helper permits "unsafe casts" for shapes, equivalent to using
`tf.Tensor.set_shape()` where domain-specific knowledge is available.
Args:
dataset: A `Dataset` object.
output_types: A nested structure of `tf.DType` objects.
output_shapes: (Optional.) A nested structure of `tf.TensorShape` objects.
If omitted, the shapes will be inherited from `dataset`.
output_classes: (Optional.) A nested structure of class types. If omitted,
the class types will be inherited from `dataset`.
allow_unsafe_cast: (Optional.) If `True`, the caller may switch the
reported output types and shapes of the restructured dataset, e.g. to
switch a sparse tensor represented as `tf.variant` to its user-visible
type and shape.
Raises:
ValueError: If either `output_types` or `output_shapes` is not compatible
with the structure of `dataset`.
"""
self._input_dataset = dataset
input_types = dataset_ops.get_legacy_output_types(dataset)
if not allow_unsafe_cast:
# Validate that the types are compatible.
output_types = nest.map_structure(dtypes.as_dtype, output_types)
flat_original_types = nest.flatten(input_types)
flat_new_types = nest.flatten(output_types)
if flat_original_types != flat_new_types:
raise ValueError(
"Dataset with output types %r cannot be restructured to have "
"output types %r" %
(dataset_ops.get_legacy_output_types(dataset),
output_types))
input_shapes = dataset_ops.get_legacy_output_shapes(dataset)
if output_shapes is None:
# Inherit shapes from the original `dataset`.
output_shapes = nest.pack_sequence_as(output_types,
nest.flatten(input_shapes))
else:
if not allow_unsafe_cast:
# Validate that the shapes are compatible.
nest.assert_same_structure(output_types, output_shapes)
flat_original_shapes = nest.flatten(input_shapes)
flat_new_shapes = nest.flatten_up_to(output_types,
output_shapes)
for original_shape, new_shape in zip(flat_original_shapes,
flat_new_shapes):
if not original_shape.is_compatible_with(new_shape):
raise ValueError(
"Dataset with output shapes %r cannot be restructured to have "
"incompatible output shapes %r" %
(input_shapes, output_shapes))
output_shapes = nest.map_structure_up_to(output_types,
tensor_shape.as_shape,
output_shapes)
input_classes = dataset_ops.get_legacy_output_classes(dataset)
if output_classes is None:
# Inherit class types from the original `dataset`.
output_classes = nest.pack_sequence_as(output_types,
nest.flatten(input_classes))
self._structure = structure.convert_legacy_structure(
output_types, output_shapes, output_classes)
variant_tensor = self._input_dataset._variant_tensor # pylint: disable=protected-access
super(_RestructuredDataset, self).__init__(dataset, variant_tensor)
def __init__(self,
dataset,
output_types,
output_shapes=None,
output_classes=None,
allow_unsafe_cast=False):
"""Creates a new dataset with the given output types and shapes.
The given `dataset` must have a structure that is convertible:
* `dataset.output_types` must be the same as `output_types` module nesting.
* Each shape in `dataset.output_shapes` must be compatible with each shape
in `output_shapes` (if given).
Note: This helper permits "unsafe casts" for shapes, equivalent to using
`tf.Tensor.set_shape()` where domain-specific knowledge is available.
Args:
dataset: A `Dataset` object.
output_types: A nested structure of `tf.DType` objects.
output_shapes: (Optional.) A nested structure of `tf.TensorShape` objects.
If omitted, the shapes will be inherited from `dataset`.
output_classes: (Optional.) A nested structure of class types.
If omitted, the class types will be inherited from `dataset`.
allow_unsafe_cast: (Optional.) If `True`, the caller may switch the
reported output types and shapes of the restructured dataset, e.g. to
switch a sparse tensor represented as `tf.variant` to its user-visible
type and shape.
Raises:
ValueError: If either `output_types` or `output_shapes` is not compatible
with the structure of `dataset`.
"""
self._input_dataset = dataset
input_types = dataset_ops.get_legacy_output_types(dataset)
if not allow_unsafe_cast:
# Validate that the types are compatible.
output_types = nest.map_structure(dtypes.as_dtype, output_types)
flat_original_types = nest.flatten(input_types)
flat_new_types = nest.flatten(output_types)
if flat_original_types != flat_new_types:
raise ValueError(
"Dataset with output types %r cannot be restructured to have "
"output types %r" %
(dataset_ops.get_legacy_output_types(dataset), output_types))
input_shapes = dataset_ops.get_legacy_output_shapes(dataset)
if output_shapes is None:
# Inherit shapes from the original `dataset`.
output_shapes = nest.pack_sequence_as(
output_types, nest.flatten(input_shapes))
else:
if not allow_unsafe_cast:
# Validate that the shapes are compatible.
nest.assert_same_structure(output_types, output_shapes)
flat_original_shapes = nest.flatten(input_shapes)
flat_new_shapes = nest.flatten_up_to(output_types, output_shapes)
for original_shape, new_shape in zip(flat_original_shapes,
flat_new_shapes):
if not original_shape.is_compatible_with(new_shape):
raise ValueError(
"Dataset with output shapes %r cannot be restructured to have "
"incompatible output shapes %r" % (input_shapes,
output_shapes))
output_shapes = nest.map_structure_up_to(
output_types, tensor_shape.as_shape, output_shapes)
input_classes = dataset_ops.get_legacy_output_classes(dataset)
if output_classes is None:
# Inherit class types from the original `dataset`.
output_classes = nest.pack_sequence_as(
output_types, nest.flatten(input_classes))
self._structure = structure.convert_legacy_structure(
output_types, output_shapes, output_classes)
variant_tensor = self._input_dataset._variant_tensor # pylint: disable=protected-access
super(_RestructuredDataset, self).__init__(dataset, variant_tensor)