tensorflow::Tensor

#include <tensor.h>

Represents an n-dimensional array of values.

Summary

Constructors and Destructors

Tensor()
Creates a 1-dimensional, 0-element float tensor.
Tensor(DataType type, const TensorShape & shape)
Creates a Tensor of the given type and shape.
Tensor(Allocator *a, DataType type, const TensorShape & shape)
Creates a tensor with the input type and shape, using the allocator a to allocate the underlying buffer.
Tensor(Allocator *a, DataType type, const TensorShape & shape, const AllocationAttributes & allocation_attr)
Creates a tensor with the input type and shape, using the allocator a and the specified "allocation_attr" to allocate the underlying buffer.
Tensor(DataType type)
Creates an empty Tensor of the given data type.
Tensor(float scalar_value)
Tensor(double scalar_value)
Tensor(int32 scalar_value)
Tensor(uint32 scalar_value)
Tensor(uint16 scalar_value)
Tensor(uint8 scalar_value)
Tensor(int16 scalar_value)
Tensor(int8 scalar_value)
Tensor(string scalar_value)
Tensor(complex64 scalar_value)
Tensor(complex128 scalar_value)
Tensor(int64 scalar_value)
Tensor(uint64 scalar_value)
Tensor(bool scalar_value)
Tensor(qint8 scalar_value)
Tensor(quint8 scalar_value)
Tensor(qint16 scalar_value)
Tensor(quint16 scalar_value)
Tensor(qint32 scalar_value)
Tensor(bfloat16 scalar_value)
Tensor(Eigen::half scalar_value)
Tensor(ResourceHandle scalar_value)
Tensor(const char *scalar_value)
Tensor(const Tensor & other)
Copy constructor.
Tensor(Tensor && other)
Move constructor.
~Tensor()

Public functions

AllocatedBytes() const
size_t
AsProtoField(TensorProto *proto) const
void
Fills in proto with *this tensor's content.
AsProtoTensorContent(TensorProto *proto) const
void
BitcastFrom(const Tensor & other, DataType dtype, const TensorShape & shape)
Copy the other tensor into this tensor, reshape it and reinterpret the buffer's datatype.
CopyFrom(const Tensor & other, const TensorShape & shape) TF_MUST_USE_RESULT
bool
Copy the other tensor into this tensor and reshape it.
DebugString(int num_values) const
string
A human-readable summary of the tensor suitable for debugging.
DebugString() const
string
DeviceSafeDebugString() const
string
FillDescription(TensorDescription *description) const
void
Fill in the TensorDescription proto with metadata about the tensor that is useful for monitoring and debugging.
FromProto(const TensorProto & other) TF_MUST_USE_RESULT
bool
Parse other and construct the tensor.
FromProto(Allocator *a, const TensorProto & other) TF_MUST_USE_RESULT
bool
IsAligned() const
bool
Returns true iff this tensor is aligned.
IsInitialized() const
bool
If necessary, has this Tensor been initialized?
IsSameSize(const Tensor & b) const
bool
NumElements() const
int64
Convenience accessor for the tensor shape.
SharesBufferWith(const Tensor & b) const
bool
Slice(int64 dim0_start, int64 dim0_limit) const
Slice this tensor along the 1st dimension.
SubSlice(int64 index) const
Select a subslice from this tensor along the 1st dimension.
SummarizeValue(int64 max_entries, bool print_v2) const
string
Render the first max_entries values in *this into a string.
TotalBytes() const
size_t
Returns the estimated memory usage of this tensor.
UnsafeCopyFromInternal(const Tensor & other, DataType dtype, const TensorShape & shape)
void
Like BitcastFrom, but CHECK fails if any preconditions are not met.
bit_casted_shaped(gtl::ArraySlice< int64 > new_sizes)
TTypes< T, NDIMS >::Tensor
Return the tensor data to an Eigen::Tensor with the new shape specified in new_sizes and cast to a new dtype T.
bit_casted_shaped(gtl::ArraySlice< int64 > new_sizes) const
TTypes< T, NDIMS >::ConstTensor
Return the tensor data to an Eigen::Tensor with the new shape specified in new_sizes and cast to a new dtype T.
bit_casted_tensor()
TTypes< T, NDIMS >::Tensor
Return the tensor data to an Eigen::Tensor with the same size but a bitwise cast to the specified dtype T.
bit_casted_tensor() const
TTypes< T, NDIMS >::ConstTensor
Return the tensor data to an Eigen::Tensor with the same size but a bitwise cast to the specified dtype T.
dim_size(int d) const
int64
Convenience accessor for the tensor shape.
dims() const
int
Convenience accessor for the tensor shape.
dtype() const
DataType
Returns the data type.
flat()
TTypes< T >::Flat
Return the tensor data as an Eigen::Tensor of the data type and a specified shape.
flat() const
TTypes< T >::ConstFlat
flat_inner_dims()
TTypes< T, NDIMS >::Tensor
Returns the data as an Eigen::Tensor with NDIMS dimensions, collapsing all Tensor dimensions but the last NDIMS-1 into the first dimension of the result.
flat_inner_dims() const
TTypes< T, NDIMS >::ConstTensor
flat_inner_outer_dims(int64 begin)
TTypes< T, NDIMS >::Tensor
Returns the data as an Eigen::Tensor with NDIMS dimensions, collapsing the first 'begin' Tensor dimensions into the first dimension of the result and the Tensor dimensions of the last dims() - 'begin' - NDIMS into the last dimension of the result.
flat_inner_outer_dims(int64 begin) const
TTypes< T, NDIMS >::ConstTensor
flat_outer_dims()
TTypes< T, NDIMS >::Tensor
Returns the data as an Eigen::Tensor with NDIMS dimensions, collapsing all Tensor dimensions but the first NDIMS-1 into the last dimension of the result.
flat_outer_dims() const
TTypes< T, NDIMS >::ConstTensor
matrix()
TTypes< T >::Matrix
matrix() const
TTypes< T >::ConstMatrix
operator=(const Tensor & other)
Tensor &
Assign operator. This tensor shares other's underlying storage.
operator=(Tensor && other)
Tensor &
Move operator. See move constructor for details.
reinterpret_last_dimension()
TTypes< T, NDIMS >::Tensor
Return the tensor data to an Eigen::Tensor with the last dimension elements converted into single elements of a larger type.
reinterpret_last_dimension() const
TTypes< T, NDIMS >::ConstTensor
Return the tensor data to an Eigen::Tensor with the last dimension elements converted into single elements of a larger type.
scalar()
TTypes< T >::Scalar
Return the Tensor data as a TensorMap of fixed size 1: TensorMap>.
scalar() const
TTypes< T >::ConstScalar
shape() const
const TensorShape &
Returns the shape of the tensor.
shaped(gtl::ArraySlice< int64 > new_sizes)
TTypes< T, NDIMS >::Tensor
shaped(gtl::ArraySlice< int64 > new_sizes) const
TTypes< T, NDIMS >::ConstTensor
tensor()
TTypes< T, NDIMS >::Tensor
tensor() const
TTypes< T, NDIMS >::ConstTensor
tensor_data() const
StringPiece
Returns a StringPiece mapping the current tensor's buffer.
unaligned_flat()
TTypes< T >::UnalignedFlat
unaligned_flat() const
TTypes< T >::UnalignedConstFlat
unaligned_shaped(gtl::ArraySlice< int64 > new_sizes)
TTypes< T, NDIMS >::UnalignedTensor
unaligned_shaped(gtl::ArraySlice< int64 > new_sizes) const
TTypes< T, NDIMS >::UnalignedConstTensor
vec()
TTypes< T >::Vec
Return the tensor data as an Eigen::Tensor with the type and sizes of this Tensor.
vec() const
TTypes< T >::ConstVec
Const versions of all the methods above.

Public functions

AllocatedBytes

size_t AllocatedBytes() const 

AsProtoField

void AsProtoField(
  TensorProto *proto
) const 

Fills in proto with *this tensor's content.

AsProtoField() fills in the repeated field for proto.dtype(), while AsProtoTensorContent() encodes the content in proto.tensor_content() in a compact form.

AsProtoTensorContent

void AsProtoTensorContent(
  TensorProto *proto
) const 

BitcastFrom

Status BitcastFrom(
  const Tensor & other,
  DataType dtype,
  const TensorShape & shape
)

Copy the other tensor into this tensor, reshape it and reinterpret the buffer's datatype.

If Status::OK() is returned, the two tensors now share the same underlying storage.

This call requires that the other tensor and the given type and shape are "compatible" (i.e. they occupy the same number of bytes).

Specifically:

shape.num_elements() * DataTypeSize(type)

must equal

other.num_elements() * DataTypeSize(other.dtype())

In addition, this function requires:

  • DataTypeSize(other.dtype()) != 0
  • DataTypeSize(type) != 0

If any of the requirements are not met, errors::InvalidArgument is returned.

CopyFrom

bool CopyFrom(
  const Tensor & other,
  const TensorShape & shape
) TF_MUST_USE_RESULT

Copy the other tensor into this tensor and reshape it.

This tensor shares other's underlying storage. Returns true iff other.shape() has the same number of elements of the given shape.

DebugString

string DebugString(
  int num_values
) const 

A human-readable summary of the tensor suitable for debugging.

DebugString

string DebugString() const 

DeviceSafeDebugString

string DeviceSafeDebugString() const 

FillDescription

void FillDescription(
  TensorDescription *description
) const 

Fill in the TensorDescription proto with metadata about the tensor that is useful for monitoring and debugging.

FromProto

bool FromProto(
  const TensorProto & other
) TF_MUST_USE_RESULT

Parse other and construct the tensor.

Returns true iff the parsing succeeds. If the parsing fails, the state of *this is unchanged.

FromProto

bool FromProto(
  Allocator *a,
  const TensorProto & other
) TF_MUST_USE_RESULT

IsAligned

bool IsAligned() const 

Returns true iff this tensor is aligned.

IsInitialized

bool IsInitialized() const 

If necessary, has this Tensor been initialized?

Zero-element Tensors are always considered initialized, even if they have never been assigned to and do not have any memory allocated.

IsSameSize

bool IsSameSize(
  const Tensor & b
) const 

NumElements

int64 NumElements() const 

Convenience accessor for the tensor shape.

SharesBufferWith

bool SharesBufferWith(
  const Tensor & b
) const 

Slice

Tensor Slice(
  int64 dim0_start,
  int64 dim0_limit
) const 

Slice this tensor along the 1st dimension.

I.e., the returned tensor satisfies returned[i, ...] == this[dim0_start + i, ...]. The returned tensor shares the underlying tensor buffer with this tensor.

NOTE: The returned tensor may not satisfy the same alignment requirement as this tensor depending on the shape. The caller must check the returned tensor's alignment before calling certain methods that have alignment requirement (e.g., flat(), tensor()).

NOTE: When fed with an N-dimensional tensor, this method returns a tensor also with N dimensions. If you want to select a sub tensor, see SubSlice.

REQUIRES: dims() >= 1 REQUIRES: 0 <= dim0_start <= dim0_limit <= dim_size(0)

SubSlice

Tensor SubSlice(
  int64 index
) const 

Select a subslice from this tensor along the 1st dimension.

When fed with an N-dimensional tensor, this method returns a tensor with N-1 dimensions, where the returned tensor is a subslice of the input tensor along the first dimension. The N-1 dimensions of the returned tensor are the last N-1 dimensions of the input tensor.

NOTE: The returned tensor may not satisfy the same alignment requirement as this tensor depending on the shape. The caller must check the returned tensor's alignment before calling certain methods that have alignment requirement (e.g., flat(), tensor()).

REQUIRES: dims() >= 1 REQUIRES: 0 <= dim0_start < dim_size(0)

SummarizeValue

string SummarizeValue(
  int64 max_entries,
  bool print_v2
) const 

Render the first max_entries values in *this into a string.

Tensor

 Tensor()

Creates a 1-dimensional, 0-element float tensor.

The returned Tensor is not a scalar (shape {}), but is instead an empty one-dimensional Tensor (shape {0}, NumElements() == 0). Since it has no elements, it does not need to be assigned a value and is initialized by default (IsInitialized() is true). If this is undesirable, consider creating a one-element scalar which does require initialization:

Tensor(DT_FLOAT, TensorShape({}))

      

Tensor

 Tensor(
  DataType type,
  const TensorShape & shape
)

Creates a Tensor of the given type and shape.

If LogMemory::IsEnabled() the allocation is logged as coming from an unknown kernel and step. Calling the Tensor constructor directly from within an Op is deprecated: use the OpKernelConstruction/OpKernelContext allocate_* methods to allocate a new tensor, which record the kernel and step.

The underlying buffer is allocated using a CPUAllocator.

Tensor

 Tensor(
  Allocator *a,
  DataType type,
  const TensorShape & shape
)

Creates a tensor with the input type and shape, using the allocator a to allocate the underlying buffer.

If LogMemory::IsEnabled() the allocation is logged as coming from an unknown kernel and step. Calling the Tensor constructor directly from within an Op is deprecated: use the OpKernelConstruction/OpKernelContext allocate_* methods to allocate a new tensor, which record the kernel and step.

a must outlive the lifetime of this Tensor.

Tensor

 Tensor(
  Allocator *a,
  DataType type,
  const TensorShape & shape,
  const AllocationAttributes & allocation_attr
)

Creates a tensor with the input type and shape, using the allocator a and the specified "allocation_attr" to allocate the underlying buffer.

If the kernel and step are known allocation_attr.allocation_will_be_logged should be set to true and LogMemory::RecordTensorAllocation should be called after the tensor is constructed. Calling the Tensor constructor directly from within an Op is deprecated: use the OpKernelConstruction/OpKernelContext allocate_* methods to allocate a new tensor, which record the kernel and step.

a must outlive the lifetime of this Tensor.

Tensor

 Tensor(
  DataType type
)

Creates an empty Tensor of the given data type.

Like Tensor(), returns a 1-dimensional, 0-element Tensor with IsInitialized() returning True. See the Tensor() documentation for details.

Tensor

 Tensor(
  float scalar_value
)

Tensor

 Tensor(
  double scalar_value
)

Tensor

 Tensor(
  int32 scalar_value
)

Tensor

 Tensor(
  uint32 scalar_value
)

Tensor

 Tensor(
  uint16 scalar_value
)

Tensor

 Tensor(
  uint8 scalar_value
)

Tensor

 Tensor(
  int16 scalar_value
)

Tensor

 Tensor(
  int8 scalar_value
)

Tensor

 Tensor(
  string scalar_value
)

Tensor

 Tensor(
  complex64 scalar_value
)

Tensor

 Tensor(
  complex128 scalar_value
)

Tensor

 Tensor(
  int64 scalar_value
)

Tensor

 Tensor(
  uint64 scalar_value
)

Tensor

 Tensor(
  bool scalar_value
)

Tensor

 Tensor(
  qint8 scalar_value
)

Tensor

 Tensor(
  quint8 scalar_value
)

Tensor

 Tensor(
  qint16 scalar_value
)

Tensor

 Tensor(
  quint16 scalar_value
)

Tensor

 Tensor(
  qint32 scalar_value
)

Tensor

 Tensor(
  bfloat16 scalar_value
)

Tensor

 Tensor(
  Eigen::half scalar_value
)

Tensor

 Tensor(
  ResourceHandle scalar_value
)

Tensor

 Tensor(
  const char *scalar_value
)

Tensor

 Tensor(
  const Tensor & other
)

Copy constructor.

Tensor

 Tensor(
  Tensor && other
)

Move constructor.

After this call, is safely destructible and can be assigned to, but other calls on it (e.g. shape manipulation) are not valid.

TotalBytes

size_t TotalBytes() const 

Returns the estimated memory usage of this tensor.

UnsafeCopyFromInternal

void UnsafeCopyFromInternal(
  const Tensor & other,
  DataType dtype,
  const TensorShape & shape
)

Like BitcastFrom, but CHECK fails if any preconditions are not met.

Deprecated. Use BitcastFrom instead and check the returned Status.

bit_casted_shaped

TTypes< T, NDIMS >::Tensor bit_casted_shaped(
  gtl::ArraySlice< int64 > new_sizes
)

Return the tensor data to an Eigen::Tensor with the new shape specified in new_sizes and cast to a new dtype T.

Using a bitcast is useful for move and copy operations. The allowed bitcast is the only difference from shaped().

bit_casted_shaped

TTypes< T, NDIMS >::ConstTensor bit_casted_shaped(
  gtl::ArraySlice< int64 > new_sizes
) const 

Return the tensor data to an Eigen::Tensor with the new shape specified in new_sizes and cast to a new dtype T.

Using a bitcast is useful for move and copy operations. The allowed bitcast is the only difference from shaped().

bit_casted_tensor

TTypes< T, NDIMS >::Tensor bit_casted_tensor()

Return the tensor data to an Eigen::Tensor with the same size but a bitwise cast to the specified dtype T.

Using a bitcast is useful for move and copy operations. NOTE: this is the same as tensor() except a bitcast is allowed.

bit_casted_tensor

TTypes< T, NDIMS >::ConstTensor bit_casted_tensor() const 

Return the tensor data to an Eigen::Tensor with the same size but a bitwise cast to the specified dtype T.

Using a bitcast is useful for move and copy operations. NOTE: this is the same as tensor() except a bitcast is allowed.

dim_size

int64 dim_size(
  int d
) const 

Convenience accessor for the tensor shape.

dims

int dims() const 

Convenience accessor for the tensor shape.

For all shape accessors, see comments for relevant methods of TensorShape in tensor_shape.h.

dtype

DataType dtype() const 

Returns the data type.

flat

TTypes< T >::Flat flat()

Return the tensor data as an Eigen::Tensor of the data type and a specified shape.

These methods allow you to access the data with the dimensions and sizes of your choice. You do not need to know the number of dimensions of the Tensor to call them. However, they CHECK that the type matches and the dimensions requested creates an Eigen::Tensor with the same number of elements as the tensor.

Example:

  
    typedef float T;
    Tensor my_ten(...built with Shape{planes: 4, rows: 3, cols: 5}...);
    // 1D Eigen::Tensor, size 60:
    auto flat = my_ten.flat();
    // 2D Eigen::Tensor 12 x 5:
    auto inner = my_ten.flat_inner_dims();
    // 2D Eigen::Tensor 4 x 15:
    auto outer = my_ten.shaped({4, 15});
    // CHECK fails, bad num elements:
    auto outer = my_ten.shaped({4, 8});
    // 3D Eigen::Tensor 6 x 5 x 2:
    auto weird = my_ten.shaped({6, 5, 2});
    // CHECK fails, type mismatch:
    auto bad   = my_ten.flat();

      

flat

TTypes< T >::ConstFlat flat() const 

flat_inner_dims

TTypes< T, NDIMS >::Tensor flat_inner_dims()

Returns the data as an Eigen::Tensor with NDIMS dimensions, collapsing all Tensor dimensions but the last NDIMS-1 into the first dimension of the result.

If NDIMS > dims() then leading dimensions of size 1 will be added to make the output rank NDIMS.

flat_inner_dims

TTypes< T, NDIMS >::ConstTensor flat_inner_dims() const 

flat_inner_outer_dims

TTypes< T, NDIMS >::Tensor flat_inner_outer_dims(
  int64 begin
)

Returns the data as an Eigen::Tensor with NDIMS dimensions, collapsing the first 'begin' Tensor dimensions into the first dimension of the result and the Tensor dimensions of the last dims() - 'begin' - NDIMS into the last dimension of the result.

If 'begin' < 0 then the |'begin'| leading dimensions of size 1 will be added. If 'begin' + NDIMS > dims() then 'begin' + NDIMS - dims() trailing dimensions of size 1 will be added.

flat_inner_outer_dims

TTypes< T, NDIMS >::ConstTensor flat_inner_outer_dims(
  int64 begin
) const 

flat_outer_dims

TTypes< T, NDIMS >::Tensor flat_outer_dims()

Returns the data as an Eigen::Tensor with NDIMS dimensions, collapsing all Tensor dimensions but the first NDIMS-1 into the last dimension of the result.

If NDIMS > dims() then trailing dimensions of size 1 will be added to make the output rank NDIMS.

flat_outer_dims

TTypes< T, NDIMS >::ConstTensor flat_outer_dims() const 

matrix

TTypes< T >::Matrix matrix()

matrix

TTypes< T >::ConstMatrix matrix() const 

operator=

Tensor & operator=(
  const Tensor & other
)

Assign operator. This tensor shares other's underlying storage.

operator=

Tensor & operator=(
  Tensor && other
)

Move operator. See move constructor for details.

reinterpret_last_dimension

TTypes< T, NDIMS >::Tensor reinterpret_last_dimension()

Return the tensor data to an Eigen::Tensor with the last dimension elements converted into single elements of a larger type.

For example, this is useful for kernels that can treat NCHW_VECT_C int8 tensors as NCHW int32 tensors. The sizeof(T) should equal the size of the original element type * num elements in the original last dimension. NDIMS should be 1 less than the original number of dimensions.

reinterpret_last_dimension

TTypes< T, NDIMS >::ConstTensor reinterpret_last_dimension() const 

Return the tensor data to an Eigen::Tensor with the last dimension elements converted into single elements of a larger type.

For example, this is useful for kernels that can treat NCHW_VECT_C int8 tensors as NCHW int32 tensors. The sizeof(T) should equal the size of the original element type * num elements in the original last dimension. NDIMS should be 1 less than the original number of dimensions.

scalar

TTypes< T >::Scalar scalar()

Return the Tensor data as a TensorMap of fixed size 1: TensorMap>.

Using scalar() allows the compiler to perform optimizations as the size of the tensor is known at compile time.

scalar

TTypes< T >::ConstScalar scalar() const 

shape

const TensorShape & shape() const 

Returns the shape of the tensor.

shaped

TTypes< T, NDIMS >::Tensor shaped(
  gtl::ArraySlice< int64 > new_sizes
)

shaped

TTypes< T, NDIMS >::ConstTensor shaped(
  gtl::ArraySlice< int64 > new_sizes
) const 

tensor

TTypes< T, NDIMS >::Tensor tensor()

tensor

TTypes< T, NDIMS >::ConstTensor tensor() const 

tensor_data

StringPiece tensor_data() const 

Returns a StringPiece mapping the current tensor's buffer.

The returned StringPiece may point to memory location on devices that the CPU cannot address directly.

NOTE: The underlying tensor buffer is refcounted, so the lifetime of the contents mapped by the StringPiece matches the lifetime of the buffer; callers should arrange to make sure the buffer does not get destroyed while the StringPiece is still used.

REQUIRES: DataTypeCanUseMemcpy(dtype()).

unaligned_flat

TTypes< T >::UnalignedFlat unaligned_flat()

unaligned_flat

TTypes< T >::UnalignedConstFlat unaligned_flat() const 

unaligned_shaped

TTypes< T, NDIMS >::UnalignedTensor unaligned_shaped(
  gtl::ArraySlice< int64 > new_sizes
)

unaligned_shaped

TTypes< T, NDIMS >::UnalignedConstTensor unaligned_shaped(
  gtl::ArraySlice< int64 > new_sizes
) const 

vec

TTypes< T >::Vec vec()

Return the tensor data as an Eigen::Tensor with the type and sizes of this Tensor.

Use these methods when you know the data type and the number of dimensions of the Tensor and you want an Eigen::Tensor automatically sized to the Tensor sizes. The implementation check fails if either type or sizes mismatch.

Example:

  
    typedef float T;
    Tensor my_mat(...built with Shape{rows: 3, cols: 5}...);
    auto mat = my_mat.matrix();    // 2D Eigen::Tensor, 3 x 5.
    auto mat = my_mat.tensor(); // 2D Eigen::Tensor, 3 x 5.
    auto vec = my_mat.vec();       // CHECK fails as my_mat is 2D.
    auto vec = my_mat.tensor(); // CHECK fails as my_mat is 2D.
    auto mat = my_mat.matrix();// CHECK fails as type mismatch.

      

vec

TTypes< T >::ConstVec vec() const 

Const versions of all the methods above.

~Tensor

 ~Tensor()