torch.Tensor

A torch.Tensor is a multi-dimensional matrix containing elements of a single data type.

Torch defines 10 tensor types with CPU and GPU variants which are as follows:

Data type	dtype	CPU tensor	GPU tensor
32-bit floating point	`torch.float32` or `torch.float`	`torch.FloatTensor`	`torch.cuda.FloatTensor`
64-bit floating point	`torch.float64` or `torch.double`	`torch.DoubleTensor`	`torch.cuda.DoubleTensor`
16-bit floating point 1	`torch.float16` or `torch.half`	`torch.HalfTensor`	`torch.cuda.HalfTensor`
16-bit floating point 2	`torch.bfloat16`	`torch.BFloat16Tensor`	`torch.cuda.BFloat16Tensor`
32-bit complex	`torch.complex32`
64-bit complex	`torch.complex64`
128-bit complex	`torch.complex128` or `torch.cdouble`
8-bit integer (unsigned)	`torch.uint8`	`torch.ByteTensor`	`torch.cuda.ByteTensor`
8-bit integer (signed)	`torch.int8`	`torch.CharTensor`	`torch.cuda.CharTensor`
16-bit integer (signed)	`torch.int16` or `torch.short`	`torch.ShortTensor`	`torch.cuda.ShortTensor`
32-bit integer (signed)	`torch.int32` or `torch.int`	`torch.IntTensor`	`torch.cuda.IntTensor`
64-bit integer (signed)	`torch.int64` or `torch.long`	`torch.LongTensor`	`torch.cuda.LongTensor`
Boolean	`torch.bool`	`torch.BoolTensor`	`torch.cuda.BoolTensor`

1: Sometimes referred to as binary16: uses 1 sign, 5 exponent, and 10 significand bits. Useful when precision is important at the expense of range.
2: Sometimes referred to as Brain Floating Point: uses 1 sign, 8 exponent, and 7 significand bits. Useful when range is important, since it has the same number of exponent bits as float32

torch.Tensor is an alias for the default tensor type (torch.FloatTensor).

A tensor can be constructed from a Python list or sequence using the torch.tensor() constructor:

>>> torch.tensor([[1., -1.], [1., -1.]])
tensor([[ 1.0000, -1.0000],
        [ 1.0000, -1.0000]])
>>> torch.tensor(np.array([[1, 2, 3], [4, 5, 6]]))
tensor([[ 1,  2,  3],
        [ 4,  5,  6]])

Warning

torch.tensor() always copies data. If you have a Tensor data and just want to change its requires_grad flag, use requires_grad_() or detach() to avoid a copy. If you have a numpy array and want to avoid a copy, use torch.as_tensor().

A tensor of specific data type can be constructed by passing a torch.dtype and/or a torch.device to a constructor or tensor creation op:

>>> torch.zeros([2, 4], dtype=torch.int32)
tensor([[ 0,  0,  0,  0],
        [ 0,  0,  0,  0]], dtype=torch.int32)
>>> cuda0 = torch.device('cuda:0')
>>> torch.ones([2, 4], dtype=torch.float64, device=cuda0)
tensor([[ 1.0000,  1.0000,  1.0000,  1.0000],
        [ 1.0000,  1.0000,  1.0000,  1.0000]], dtype=torch.float64, device='cuda:0')

The contents of a tensor can be accessed and modified using Python’s indexing and slicing notation:

>>> x = torch.tensor([[1, 2, 3], [4, 5, 6]])
>>> print(x[1][2])
tensor(6)
>>> x[0][1] = 8
>>> print(x)
tensor([[ 1,  8,  3],
        [ 4,  5,  6]])

Use torch.Tensor.item() to get a Python number from a tensor containing a single value:

>>> x = torch.tensor([[1]])
>>> x
tensor([[ 1]])
>>> x.item()
1
>>> x = torch.tensor(2.5)
>>> x
tensor(2.5000)
>>> x.item()
2.5

A tensor can be created with requires_grad=True so that torch.autograd records operations on them for automatic differentiation.

>>> x = torch.tensor([[1., -1.], [1., 1.]], requires_grad=True)
>>> out = x.pow(2).sum()
>>> out.backward()
>>> x.grad
tensor([[ 2.0000, -2.0000],
        [ 2.0000,  2.0000]])

Each tensor has an associated torch.Storage, which holds its data. The tensor class also provides multi-dimensional, strided view of a storage and defines numeric operations on it.

Note

For more information on tensor views, see Tensor Views.

Note

For more information on the torch.dtype, torch.device, and torch.layout attributes of a torch.Tensor, see Tensor Attributes.

Note

Methods which mutate a tensor are marked with an underscore suffix. For example, torch.FloatTensor.abs_() computes the absolute value in-place and returns the modified tensor, while torch.FloatTensor.abs() computes the result in a new tensor.

Note

To change an existing tensor’s torch.device and/or torch.dtype, consider using to() method on the tensor.

Warning

Current implementation of torch.Tensor introduces memory overhead, thus it might lead to unexpectedly high memory usage in the applications with many tiny tensors. If this is your case, consider using one large structure.

class torch.Tensor

There are a few main ways to create a tensor, depending on your use case.

To create a tensor with pre-existing data, use torch.tensor().
To create a tensor with specific size, use torch.* tensor creation ops (see Creation Ops).
To create a tensor with the same size (and similar types) as another tensor, use torch.*_like tensor creation ops (see Creation Ops).
To create a tensor with similar type but different size as another tensor, use tensor.new_* creation ops.

new_tensor(data, dtype=None, device=None, requires_grad=False) → Tensor

Returns a new Tensor with data as the tensor data. By default, the returned Tensor has the same torch.dtype and torch.device as this tensor.

Warning

new_tensor() always copies data. If you have a Tensor data and want to avoid a copy, use torch.Tensor.requires_grad_() or torch.Tensor.detach(). If you have a numpy array and want to avoid a copy, use torch.from_numpy().

Warning

When data is a tensor x, new_tensor() reads out ‘the data’ from whatever it is passed, and constructs a leaf variable. Therefore tensor.new_tensor(x) is equivalent to x.clone().detach() and tensor.new_tensor(x, requires_grad=True) is equivalent to x.clone().detach().requires_grad_(True). The equivalents using clone() and detach() are recommended.

Parameters

data (array_like) – The returned Tensor copies data.
dtype (torch.dtype, optional) – the desired type of returned tensor. Default: if None, same torch.dtype as this tensor.
device (torch.device, optional) – the desired device of returned tensor. Default: if None, same torch.device as this tensor.
requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

Example:

>>> tensor = torch.ones((2,), dtype=torch.int8)
>>> data = [[0, 1], [2, 3]]
>>> tensor.new_tensor(data)
tensor([[ 0,  1],
        [ 2,  3]], dtype=torch.int8)

new_full(size, fill_value, dtype=None, device=None, requires_grad=False) → Tensor

Returns a Tensor of size size filled with fill_value. By default, the returned Tensor has the same torch.dtype and torch.device as this tensor.

Parameters

fill_value (scalar) – the number to fill the output tensor with.
dtype (torch.dtype, optional) – the desired type of returned tensor. Default: if None, same torch.dtype as this tensor.
device (torch.device, optional) – the desired device of returned tensor. Default: if None, same torch.device as this tensor.
requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

Example:

>>> tensor = torch.ones((2,), dtype=torch.float64)
>>> tensor.new_full((3, 4), 3.141592)
tensor([[ 3.1416,  3.1416,  3.1416,  3.1416],
        [ 3.1416,  3.1416,  3.1416,  3.1416],
        [ 3.1416,  3.1416,  3.1416,  3.1416]], dtype=torch.float64)

new_empty(size, dtype=None, device=None, requires_grad=False) → Tensor

Returns a Tensor of size size filled with uninitialized data. By default, the returned Tensor has the same torch.dtype and torch.device as this tensor.

Parameters

dtype (torch.dtype, optional) – the desired type of returned tensor. Default: if None, same torch.dtype as this tensor.
device (torch.device, optional) – the desired device of returned tensor. Default: if None, same torch.device as this tensor.
requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

Example:

>>> tensor = torch.ones(())
>>> tensor.new_empty((2, 3))
tensor([[ 5.8182e-18,  4.5765e-41, -1.0545e+30],
        [ 3.0949e-41,  4.4842e-44,  0.0000e+00]])

new_ones(size, dtype=None, device=None, requires_grad=False) → Tensor

Returns a Tensor of size size filled with 1. By default, the returned Tensor has the same torch.dtype and torch.device as this tensor.

Parameters

size (int...) – a list, tuple, or torch.Size of integers defining the shape of the output tensor.
dtype (torch.dtype, optional) – the desired type of returned tensor. Default: if None, same torch.dtype as this tensor.
device (torch.device, optional) – the desired device of returned tensor. Default: if None, same torch.device as this tensor.
requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

Example:

>>> tensor = torch.tensor((), dtype=torch.int32)
>>> tensor.new_ones((2, 3))
tensor([[ 1,  1,  1],
        [ 1,  1,  1]], dtype=torch.int32)

new_zeros(size, dtype=None, device=None, requires_grad=False) → Tensor

Returns a Tensor of size size filled with 0. By default, the returned Tensor has the same torch.dtype and torch.device as this tensor.

Parameters

size (int...) – a list, tuple, or torch.Size of integers defining the shape of the output tensor.
dtype (torch.dtype, optional) – the desired type of returned tensor. Default: if None, same torch.dtype as this tensor.
device (torch.device, optional) – the desired device of returned tensor. Default: if None, same torch.device as this tensor.
requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

Example:

>>> tensor = torch.tensor((), dtype=torch.float64)
>>> tensor.new_zeros((2, 3))
tensor([[ 0.,  0.,  0.],
        [ 0.,  0.,  0.]], dtype=torch.float64)

is_cuda: Is True if the Tensor is stored on the GPU, False otherwise.

is_quantized: Is True if the Tensor is quantized, False otherwise.

is_meta: Is True if the Tensor is a meta tensor, False otherwise. Meta tensors are like normal tensors, but they carry no data.

device: Is the torch.device where this Tensor is.

grad: This attribute is None by default and becomes a Tensor the first time a call to backward() computes gradients for self. The attribute will then contain the gradients computed and future calls to backward() will accumulate (add) gradients into it.

ndim: Alias for dim()

T

Is this Tensor with its dimensions reversed.

If n is the number of dimensions in x, x.T is equivalent to x.permute(n-1, n-2, ..., 0).

real

Returns a new tensor containing real values of the self tensor. The returned tensor and self share the same underlying storage.

Warning

real() is only supported for tensors with complex dtypes.

Example::

>>> x=torch.randn(4, dtype=torch.cfloat)
>>> x
tensor([(0.3100+0.3553j), (-0.5445-0.7896j), (-1.6492-0.0633j), (-0.0638-0.8119j)])
>>> x.real
tensor([ 0.3100, -0.5445, -1.6492, -0.0638])

imag

Returns a new tensor containing imaginary values of the self tensor. The returned tensor and self share the same underlying storage.

Warning

imag() is only supported for tensors with complex dtypes.

Example::

>>> x=torch.randn(4, dtype=torch.cfloat)
>>> x
tensor([(0.3100+0.3553j), (-0.5445-0.7896j), (-1.6492-0.0633j), (-0.0638-0.8119j)])
>>> x.imag
tensor([ 0.3553, -0.7896, -0.0633, -0.8119])

abs() → Tensor: See torch.abs()

abs_() → Tensor: In-place version of abs()

absolute() → Tensor: Alias for abs()

absolute_() → Tensor: In-place version of absolute() Alias for abs_()

acos() → Tensor: See torch.acos()

acos_() → Tensor: In-place version of acos()

arccos() → Tensor: See torch.arccos()

arccos_() → Tensor: In-place version of arccos()

add(other, *, alpha=1) → Tensor

Add a scalar or tensor to self tensor. If both alpha and other are specified, each element of other is scaled by alpha before being used.

When other is a tensor, the shape of other must be broadcastable with the shape of the underlying tensor

See torch.add()

add_(other, *, alpha=1) → Tensor: In-place version of add()

addbmm(batch1, batch2, *, beta=1, alpha=1) → Tensor: See torch.addbmm()

addbmm_(batch1, batch2, *, beta=1, alpha=1) → Tensor: In-place version of addbmm()

addcdiv(tensor1, tensor2, *, value=1) → Tensor: See torch.addcdiv()

addcdiv_(tensor1, tensor2, *, value=1) → Tensor: In-place version of addcdiv()

addcmul(tensor1, tensor2, *, value=1) → Tensor: See torch.addcmul()

addcmul_(tensor1, tensor2, *, value=1) → Tensor: In-place version of addcmul()

addmm(mat1, mat2, *, beta=1, alpha=1) → Tensor: See torch.addmm()

addmm_(mat1, mat2, *, beta=1, alpha=1) → Tensor: In-place version of addmm()

sspaddmm(mat1, mat2, *, beta=1, alpha=1) → Tensor: See torch.sspaddmm()

addmv(mat, vec, *, beta=1, alpha=1) → Tensor: See torch.addmv()

addmv_(mat, vec, *, beta=1, alpha=1) → Tensor: In-place version of addmv()

addr(vec1, vec2, *, beta=1, alpha=1) → Tensor: See torch.addr()

addr_(vec1, vec2, *, beta=1, alpha=1) → Tensor: In-place version of addr()

allclose(other, rtol=1e-05, atol=1e-08, equal_nan=False) → Tensor: See torch.allclose()

amax(dim=None, keepdim=False) → Tensor: See torch.amax()

amin(dim=None, keepdim=False) → Tensor: See torch.amin()

angle() → Tensor: See torch.angle()

apply_(callable) → Tensor: Applies the function callable to each element in the tensor, replacing each element with the value returned by callable.

Note

This function only works with CPU tensors and should not be used in code sections that require high performance.

argmax(dim=None, keepdim=False) → LongTensor: See torch.argmax()

argmin(dim=None, keepdim=False) → LongTensor: See torch.argmin()

argsort(dim=-1, descending=False) → LongTensor: See torch.argsort()

asin() → Tensor: See torch.asin()

asin_() → Tensor: In-place version of asin()

arcsin() → Tensor: See torch.arcsin()

arcsin_() → Tensor: In-place version of arcsin()

as_strided(size, stride, storage_offset=0) → Tensor: See torch.as_strided()

atan() → Tensor: See torch.atan()

atan_() → Tensor: In-place version of atan()

arctan() → Tensor: See torch.arctan()

arctan_() → Tensor: In-place version of arctan()

atan2(other) → Tensor: See torch.atan2()

atan2_(other) → Tensor: In-place version of atan2()

all(dim=None, keepdim=False) → Tensor: See torch.all()

any(dim=None, keepdim=False) → Tensor: See torch.any()

backward(gradient=None, retain_graph=None, create_graph=False, inputs=None) [source]

Computes the gradient of current tensor w.r.t. graph leaves.

The graph is differentiated using the chain rule. If the tensor is non-scalar (i.e. its data has more than one element) and requires gradient, the function additionally requires specifying gradient. It should be a tensor of matching type and location, that contains the gradient of the differentiated function w.r.t. self.

This function accumulates gradients in the leaves - you might need to zero .grad attributes or set them to None before calling it. See Default gradient layouts for details on the memory layout of accumulated gradients.

Note

If you run any forward ops, create gradient, and/or call backward in a user-specified CUDA stream context, see Stream semantics of backward passes.

Parameters

gradient (Tensor or None) – Gradient w.r.t. the tensor. If it is a tensor, it will be automatically converted to a Tensor that does not require grad unless create_graph is True. None values can be specified for scalar Tensors or ones that don’t require grad. If a None value would be acceptable then this argument is optional.
retain_graph (bool, optional) – If False, the graph used to compute the grads will be freed. Note that in nearly all cases setting this option to True is not needed and often can be worked around in a much more efficient way. Defaults to the value of create_graph.
create_graph (bool, optional) – If True, graph of the derivative will be constructed, allowing to compute higher order derivative products. Defaults to False.
inputs (sequence of Tensor) – Inputs w.r.t. which the gradient will be accumulated into .grad. All other Tensors will be ignored. If not provided, the gradient is accumulated into all the leaf Tensors that were used to compute the attr::tensors. All the provided inputs must be leaf Tensors.

baddbmm(batch1, batch2, *, beta=1, alpha=1) → Tensor: See torch.baddbmm()

baddbmm_(batch1, batch2, *, beta=1, alpha=1) → Tensor: In-place version of baddbmm()

bernoulli(*, generator=None) → Tensor

Returns a result tensor where each $\texttt{result[i]}$ is independently sampled from $\text{Bernoulli}(\texttt{self[i]})$ . self must have floating point dtype, and the result will have the same dtype.

See torch.bernoulli()

bernoulli_()

bernoulli_(p=0.5, *, generator=None) → Tensor: Fills each location of self with an independent sample from $\text{Bernoulli}(\texttt{p})$ . self can have integral dtype.

bernoulli_(p_tensor, *, generator=None) → Tensor

p_tensor should be a tensor containing probabilities to be used for drawing the binary random number.

The $\text{i}^{th}$ element of self tensor will be set to a value sampled from $\text{Bernoulli}(\texttt{p\_tensor[i]})$ .

self can have integral dtype, but p_tensor must have floating point dtype.

See also bernoulli() and torch.bernoulli()

bfloat16(memory_format=torch.preserve_format) → Tensor

self.bfloat16() is equivalent to self.to(torch.bfloat16). See to().

Parameters: memory_format (torch.memory_format, optional) – the desired memory format of returned Tensor. Default: torch.preserve_format.

bincount(weights=None, minlength=0) → Tensor: See torch.bincount()

bitwise_not() → Tensor: See torch.bitwise_not()

bitwise_not_() → Tensor: In-place version of bitwise_not()

bitwise_and() → Tensor: See torch.bitwise_and()

bitwise_and_() → Tensor: In-place version of bitwise_and()

bitwise_or() → Tensor: See torch.bitwise_or()

bitwise_or_() → Tensor: In-place version of bitwise_or()

bitwise_xor() → Tensor: See torch.bitwise_xor()

bitwise_xor_() → Tensor: In-place version of bitwise_xor()

bmm(batch2) → Tensor: See torch.bmm()

bool(memory_format=torch.preserve_format) → Tensor

self.bool() is equivalent to self.to(torch.bool). See to().

Parameters: memory_format (torch.memory_format, optional) – the desired memory format of returned Tensor. Default: torch.preserve_format.

byte(memory_format=torch.preserve_format) → Tensor

self.byte() is equivalent to self.to(torch.uint8). See to().

Parameters: memory_format (torch.memory_format, optional) – the desired memory format of returned Tensor. Default: torch.preserve_format.

broadcast_to(shape) → Tensor: See torch.broadcast_to().

cauchy_(median=0, sigma=1, *, generator=None) → Tensor: Fills the tensor with numbers drawn from the Cauchy distribution:

$f(x) = \dfrac{1}{\pi} \dfrac{\sigma}{(x - \text{median})^2 + \sigma^2}$

ceil() → Tensor: See torch.ceil()

ceil_() → Tensor: In-place version of ceil()

char(memory_format=torch.preserve_format) → Tensor

self.char() is equivalent to self.to(torch.int8). See to().

Parameters: memory_format (torch.memory_format, optional) – the desired memory format of returned Tensor. Default: torch.preserve_format.

cholesky(upper=False) → Tensor: See torch.cholesky()

cholesky_inverse(upper=False) → Tensor: See torch.cholesky_inverse()

cholesky_solve(input2, upper=False) → Tensor: See torch.cholesky_solve()

chunk(chunks, dim=0) → List of Tensors: See torch.chunk()

clamp(min, max) → Tensor: See torch.clamp()

clamp_(min, max) → Tensor: In-place version of clamp()

clip(min, max) → Tensor: Alias for clamp().

clip_(min, max) → Tensor: Alias for clamp_().

clone(*, memory_format=torch.preserve_format) → Tensor: See torch.clone()

contiguous(memory_format=torch.contiguous_format) → Tensor

Returns a contiguous in memory tensor containing the same data as self tensor. If self tensor is already in the specified memory format, this function returns the self tensor.

Parameters: memory_format (torch.memory_format, optional) – the desired memory format of returned Tensor. Default: torch.contiguous_format.

copy_(src, non_blocking=False) → Tensor

Copies the elements from src into self tensor and returns self.

The src tensor must be broadcastable with the self tensor. It may be of a different data type or reside on a different device.

Parameters

src (Tensor) – the source tensor to copy from
non_blocking (bool) – if True and this copy is between CPU and GPU, the copy may occur asynchronously with respect to the host. For other cases, this argument has no effect.

conj() → Tensor: See torch.conj()

copysign(other) → Tensor: See torch.copysign()

copysign_(other) → Tensor: In-place version of copysign()

cos() → Tensor: See torch.cos()

cos_() → Tensor: In-place version of cos()

cosh() → Tensor: See torch.cosh()

cosh_() → Tensor: In-place version of cosh()

count_nonzero(dim=None) → Tensor: See torch.count_nonzero()

acosh() → Tensor: See torch.acosh()

acosh_() → Tensor: In-place version of acosh()

arccosh()

acosh() -> Tensor

See torch.arccosh()

arccosh_()

acosh_() -> Tensor

In-place version of arccosh()

cpu(memory_format=torch.preserve_format) → Tensor

Returns a copy of this object in CPU memory.

If this object is already in CPU memory and on the correct device, then no copy is performed and the original object is returned.

Parameters: memory_format (torch.memory_format, optional) – the desired memory format of returned Tensor. Default: torch.preserve_format.

cross(other, dim=-1) → Tensor: See torch.cross()

cuda(device=None, non_blocking=False, memory_format=torch.preserve_format) → Tensor

Returns a copy of this object in CUDA memory.

If this object is already in CUDA memory and on the correct device, then no copy is performed and the original object is returned.

Parameters

device (torch.device) – The destination GPU device. Defaults to the current CUDA device.
non_blocking (bool) – If True and the source is in pinned memory, the copy will be asynchronous with respect to the host. Otherwise, the argument has no effect. Default: False.
memory_format (torch.memory_format, optional) – the desired memory format of returned Tensor. Default: torch.preserve_format.

logcumsumexp(dim) → Tensor: See torch.logcumsumexp()

cummax(dim) -> (Tensor, Tensor): See torch.cummax()

cummin(dim) -> (Tensor, Tensor): See torch.cummin()

cumprod(dim, dtype=None) → Tensor: See torch.cumprod()

cumprod_(dim, dtype=None) → Tensor: In-place version of cumprod()

cumsum(dim, dtype=None) → Tensor: See torch.cumsum()

cumsum_(dim, dtype=None) → Tensor: In-place version of cumsum()

data_ptr() → int: Returns the address of the first element of self tensor.

deg2rad() → Tensor: See torch.deg2rad()

dequantize() → Tensor: Given a quantized Tensor, dequantize it and return the dequantized float Tensor.

det() → Tensor: See torch.det()

dense_dim() → int

Return the number of dense dimensions in a sparse tensor self.

Warning

Throws an error if self is not a sparse tensor.

detach()

Returns a new Tensor, detached from the current graph.

The result will never require gradient.

Note

Returned Tensor shares the same storage with the original one. In-place modifications on either of them will be seen, and may trigger errors in correctness checks. IMPORTANT NOTE: Previously, in-place size / stride / storage changes (such as resize_ / resize_as_ / set_ / transpose_) to the returned tensor also update the original tensor. Now, these in-place changes will not update the original tensor anymore, and will instead trigger an error. For sparse tensors: In-place indices / values changes (such as zero_ / copy_ / add_) to the returned tensor will not update the original tensor anymore, and will instead trigger an error.

detach_(): Detaches the Tensor from the graph that created it, making it a leaf. Views cannot be detached in-place.

diag(diagonal=0) → Tensor: See torch.diag()

diag_embed(offset=0, dim1=-2, dim2=-1) → Tensor: See torch.diag_embed()

diagflat(offset=0) → Tensor: See torch.diagflat()

diagonal(offset=0, dim1=0, dim2=1) → Tensor: See torch.diagonal()

fill_diagonal_(fill_value, wrap=False) → Tensor

Fill the main diagonal of a tensor that has at least 2-dimensions. When dims>2, all dimensions of input must be of equal length. This function modifies the input tensor in-place, and returns the input tensor.

Parameters

fill_value (Scalar) – the fill value
wrap (bool) – the diagonal ‘wrapped’ after N columns for tall matrices.

Example:

>>> a = torch.zeros(3, 3)
>>> a.fill_diagonal_(5)
tensor([[5., 0., 0.],
        [0., 5., 0.],
        [0., 0., 5.]])
>>> b = torch.zeros(7, 3)
>>> b.fill_diagonal_(5)
tensor([[5., 0., 0.],
        [0., 5., 0.],
        [0., 0., 5.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]])
>>> c = torch.zeros(7, 3)
>>> c.fill_diagonal_(5, wrap=True)
tensor([[5., 0., 0.],
        [0., 5., 0.],
        [0., 0., 5.],
        [0., 0., 0.],
        [5., 0., 0.],
        [0., 5., 0.],
        [0., 0., 5.]])

fmax(other) → Tensor: See torch.fmax()

fmin(other) → Tensor: See torch.fmin()

diff(n=1, dim=-1, prepend=None, append=None) → Tensor: See torch.diff()

digamma() → Tensor: See torch.digamma()

digamma_() → Tensor: In-place version of digamma()

dim() → int: Returns the number of dimensions of self tensor.

dist(other, p=2) → Tensor: See torch.dist()

div(value, *, rounding_mode=None) → Tensor: See torch.div()

div_(value, *, rounding_mode=None) → Tensor: In-place version of div()

divide(value, *, rounding_mode=None) → Tensor: See torch.divide()

divide_(value, *, rounding_mode=None) → Tensor: In-place version of divide()

dot(other) → Tensor: See torch.dot()

double(memory_format=torch.preserve_format) → Tensor

self.double() is equivalent to self.to(torch.float64). See to().

Parameters: memory_format (torch.memory_format, optional) – the desired memory format of returned Tensor. Default: torch.preserve_format.

eig(eigenvectors=False) -> (Tensor, Tensor): See torch.eig()

element_size() → int

Returns the size in bytes of an individual element.

Example:

>>> torch.tensor([]).element_size()
4
>>> torch.tensor([], dtype=torch.uint8).element_size()
1

eq(other) → Tensor: See torch.eq()

eq_(other) → Tensor: In-place version of eq()

equal(other) → bool: See torch.equal()

erf() → Tensor: See torch.erf()

erf_() → Tensor: In-place version of erf()

erfc() → Tensor: See torch.erfc()

erfc_() → Tensor: In-place version of erfc()

erfinv() → Tensor: See torch.erfinv()

erfinv_() → Tensor: In-place version of erfinv()

exp() → Tensor: See torch.exp()

exp_() → Tensor: In-place version of exp()

expm1() → Tensor: See torch.expm1()

expm1_() → Tensor: In-place version of expm1()

expand(*sizes) → Tensor

Returns a new view of the self tensor with singleton dimensions expanded to a larger size.

Passing -1 as the size for a dimension means not changing the size of that dimension.

Tensor can be also expanded to a larger number of dimensions, and the new ones will be appended at the front. For the new dimensions, the size cannot be set to -1.

Expanding a tensor does not allocate new memory, but only creates a new view on the existing tensor where a dimension of size one is expanded to a larger size by setting the stride to 0. Any dimension of size 1 can be expanded to an arbitrary value without allocating new memory.

Parameters: *sizes (torch.Size or int...) – the desired expanded size

Warning

More than one element of an expanded tensor may refer to a single memory location. As a result, in-place operations (especially ones that are vectorized) may result in incorrect behavior. If you need to write to the tensors, please clone them first.

Example:

>>> x = torch.tensor([[1], [2], [3]])
>>> x.size()
torch.Size([3, 1])
>>> x.expand(3, 4)
tensor([[ 1,  1,  1,  1],
        [ 2,  2,  2,  2],
        [ 3,  3,  3,  3]])
>>> x.expand(-1, 4)   # -1 means not changing the size of that dimension
tensor([[ 1,  1,  1,  1],
        [ 2,  2,  2,  2],
        [ 3,  3,  3,  3]])

expand_as(other) → Tensor

Expand this tensor to the same size as other. self.expand_as(other) is equivalent to self.expand(other.size()).

Please see expand() for more information about expand.

Parameters: other (torch.Tensor) – The result tensor has the same size as other.

exponential_(lambd=1, *, generator=None) → Tensor: Fills self tensor with elements drawn from the exponential distribution:

$f(x) = \lambda e^{-\lambda x}$

fix() → Tensor: See torch.fix().

fix_() → Tensor: In-place version of fix()

fill_(value) → Tensor: Fills self tensor with the specified value.

flatten(input, start_dim=0, end_dim=-1) → Tensor: see torch.flatten()

flip(dims) → Tensor: See torch.flip()

fliplr() → Tensor: See torch.fliplr()

flipud() → Tensor: See torch.flipud()

float(memory_format=torch.preserve_format) → Tensor

self.float() is equivalent to self.to(torch.float32). See to().

Parameters: memory_format (torch.memory_format, optional) – the desired memory format of returned Tensor. Default: torch.preserve_format.

float_power(exponent) → Tensor: See torch.float_power()

float_power_(exponent) → Tensor: In-place version of float_power()

floor() → Tensor: See torch.floor()

floor_() → Tensor: In-place version of floor()

floor_divide(value) → Tensor: See torch.floor_divide()

floor_divide_(value) → Tensor: In-place version of floor_divide()

fmod(divisor) → Tensor: See torch.fmod()

fmod_(divisor) → Tensor: In-place version of fmod()

frac() → Tensor: See torch.frac()

frac_() → Tensor: In-place version of frac()

gather(dim, index) → Tensor: See torch.gather()

gcd(other) → Tensor: See torch.gcd()

gcd_(other) → Tensor: In-place version of gcd()

ge(other) → Tensor: See torch.ge().

ge_(other) → Tensor: In-place version of ge().

greater_equal(other) → Tensor: See torch.greater_equal().

greater_equal_(other) → Tensor: In-place version of greater_equal().

geometric_(p, *, generator=None) → Tensor: Fills self tensor with elements drawn from the geometric distribution:

$f(X=k) = p^{k - 1} (1 - p)$

geqrf() -> (Tensor, Tensor): See torch.geqrf()

ger(vec2) → Tensor: See torch.ger()

get_device() -> Device ordinal (Integer)

For CUDA tensors, this function returns the device ordinal of the GPU on which the tensor resides. For CPU tensors, an error is thrown.

Example:

>>> x = torch.randn(3, 4, 5, device='cuda:0')
>>> x.get_device()
0
>>> x.cpu().get_device()  # RuntimeError: get_device is not implemented for type torch.FloatTensor

gt(other) → Tensor: See torch.gt().

gt_(other) → Tensor: In-place version of gt().

greater(other) → Tensor: See torch.greater().

greater_(other) → Tensor: In-place version of greater().

half(memory_format=torch.preserve_format) → Tensor

self.half() is equivalent to self.to(torch.float16). See to().

Parameters: memory_format (torch.memory_format, optional) – the desired memory format of returned Tensor. Default: torch.preserve_format.

hardshrink(lambd=0.5) → Tensor: See torch.nn.functional.hardshrink()

heaviside(values) → Tensor: See torch.heaviside()

histc(bins=100, min=0, max=0) → Tensor: See torch.histc()

hypot(other) → Tensor: See torch.hypot()

hypot_(other) → Tensor: In-place version of hypot()

i0() → Tensor: See torch.i0()

i0_() → Tensor: In-place version of i0()

igamma(other) → Tensor: See torch.igamma()

igamma_(other) → Tensor: In-place version of igamma()

igammac(other) → Tensor: See torch.igammac()

igammac_(other) → Tensor: In-place version of igammac()

index_add_(dim, index, tensor) → Tensor

Accumulate the elements of tensor into the self tensor by adding to the indices in the order given in index. For example, if dim == 0 and index[i] == j, then the ith row of tensor is added to the jth row of self.

The dimth dimension of tensor must have the same size as the length of index (which must be a vector), and all other dimensions must match self, or an error will be raised.

Note

This operation may behave nondeterministically when given tensors on a CUDA device. See Reproducibility for more information.

Parameters

dim (int) – dimension along which to index
index (IntTensor or LongTensor) – indices of tensor to select from
tensor (Tensor) – the tensor containing values to add

Example:

>>> x = torch.ones(5, 3)
>>> t = torch.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype=torch.float)
>>> index = torch.tensor([0, 4, 2])
>>> x.index_add_(0, index, t)
tensor([[  2.,   3.,   4.],
        [  1.,   1.,   1.],
        [  8.,   9.,  10.],
        [  1.,   1.,   1.],
        [  5.,   6.,   7.]])

index_add(tensor1, dim, index, tensor2) → Tensor: Out-of-place version of torch.Tensor.index_add_(). tensor1 corresponds to self in torch.Tensor.index_add_().

index_copy_(dim, index, tensor) → Tensor

Copies the elements of tensor into the self tensor by selecting the indices in the order given in index. For example, if dim == 0 and index[i] == j, then the ith row of tensor is copied to the jth row of self.

The dimth dimension of tensor must have the same size as the length of index (which must be a vector), and all other dimensions must match self, or an error will be raised.

Note

If index contains duplicate entries, multiple elements from tensor will be copied to the same index of self. The result is nondeterministic since it depends on which copy occurs last.

Parameters

dim (int) – dimension along which to index
index (LongTensor) – indices of tensor to select from
tensor (Tensor) – the tensor containing values to copy

Example:

>>> x = torch.zeros(5, 3)
>>> t = torch.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype=torch.float)
>>> index = torch.tensor([0, 4, 2])
>>> x.index_copy_(0, index, t)
tensor([[ 1.,  2.,  3.],
        [ 0.,  0.,  0.],
        [ 7.,  8.,  9.],
        [ 0.,  0.,  0.],
        [ 4.,  5.,  6.]])

index_copy(tensor1, dim, index, tensor2) → Tensor: Out-of-place version of torch.Tensor.index_copy_(). tensor1 corresponds to self in torch.Tensor.index_copy_().

index_fill_(dim, index, val) → Tensor

Fills the elements of the self tensor with value val by selecting the indices in the order given in index.

Parameters

dim (int) – dimension along which to index
index (LongTensor) – indices of self tensor to fill in
val (float) – the value to fill with

Example::

>>> x = torch.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype=torch.float)
>>> index = torch.tensor([0, 2])
>>> x.index_fill_(1, index, -1)
tensor([[-1.,  2., -1.],
        [-1.,  5., -1.],
        [-1.,  8., -1.]])

index_fill(tensor1, dim, index, value) → Tensor: Out-of-place version of torch.Tensor.index_fill_(). tensor1 corresponds to self in torch.Tensor.index_fill_().

index_put_(indices, values, accumulate=False) → Tensor

Puts values from the tensor values into the tensor self using the indices specified in indices (which is a tuple of Tensors). The expression tensor.index_put_(indices, values) is equivalent to tensor[indices] = values. Returns self.

If accumulate is True, the elements in values are added to self. If accumulate is False, the behavior is undefined if indices contain duplicate elements.

Parameters

indices (tuple of LongTensor) – tensors used to index into self.
values (Tensor) – tensor of same dtype as self.
accumulate (bool) – whether to accumulate into self

index_put(tensor1, indices, values, accumulate=False) → Tensor: Out-place version of index_put_(). tensor1 corresponds to self in torch.Tensor.index_put_().

index_select(dim, index) → Tensor: See torch.index_select()

indices() → Tensor

Return the indices tensor of a sparse COO tensor.

Warning

Throws an error if self is not a sparse COO tensor.

Example

>>> x = torch.randn(2, 3, 5)
>>> x.size()
torch.Size([2, 3, 5])
>>> x.permute(2, 0, 1).size()
torch.Size([5, 2, 3])

pin_memory() → Tensor: Copies the tensor to pinned memory, if it’s not already pinned.

pinverse() → Tensor: See torch.pinverse()

polygamma(n) → Tensor: See torch.polygamma()

polygamma_(n) → Tensor: In-place version of polygamma()

pow(exponent) → Tensor: See torch.pow()

pow_(exponent) → Tensor: In-place version of pow()

prod(dim=None, keepdim=False, dtype=None) → Tensor: See torch.prod()

put_(indices, tensor, accumulate=False) → Tensor

Copies the elements from tensor into the positions specified by indices. For the purpose of indexing, the self tensor is treated as if it were a 1-D tensor.

If accumulate is True, the elements in tensor are added to self. If accumulate is False, the behavior is undefined if indices contain duplicate elements.

Parameters

indices (LongTensor) – the indices into self
tensor (Tensor) – the tensor containing values to copy from
accumulate (bool) – whether to accumulate into self

Example:

>>> src = torch.tensor([[4, 3, 5],
...                     [6, 7, 8]])
>>> src.put_(torch.tensor([1, 3]), torch.tensor([9, 10]))
tensor([[  4,   9,   5],
        [ 10,   7,   8]])

qr(some=True) -> (Tensor, Tensor): See torch.qr()

qscheme() → torch.qscheme: Returns the quantization scheme of a given QTensor.

quantile(q, dim=None, keepdim=False) → Tensor: See torch.quantile()

nanquantile(q, dim=None, keepdim=False) → Tensor: See torch.nanquantile()

q_scale() → float: Given a Tensor quantized by linear(affine) quantization, returns the scale of the underlying quantizer().

q_zero_point() → int: Given a Tensor quantized by linear(affine) quantization, returns the zero_point of the underlying quantizer().

q_per_channel_scales() → Tensor: Given a Tensor quantized by linear (affine) per-channel quantization, returns a Tensor of scales of the underlying quantizer. It has the number of elements that matches the corresponding dimensions (from q_per_channel_axis) of the tensor.

q_per_channel_zero_points() → Tensor: Given a Tensor quantized by linear (affine) per-channel quantization, returns a tensor of zero_points of the underlying quantizer. It has the number of elements that matches the corresponding dimensions (from q_per_channel_axis) of the tensor.

q_per_channel_axis() → int: Given a Tensor quantized by linear (affine) per-channel quantization, returns the index of dimension on which per-channel quantization is applied.

rad2deg() → Tensor: See torch.rad2deg()

random_(from=0, to=None, *, generator=None) → Tensor: Fills self tensor with numbers sampled from the discrete uniform distribution over [from, to - 1]. If not specified, the values are usually only bounded by self tensor’s data type. However, for floating point types, if unspecified, range will be [0, 2^mantissa] to ensure that every value is representable. For example, torch.tensor(1, dtype=torch.double).random_() will be uniform in [0, 2^53].

ravel(input) → Tensor: see torch.ravel()

reciprocal() → Tensor: See torch.reciprocal()

reciprocal_() → Tensor: In-place version of reciprocal()

record_stream(stream): Ensures that the tensor memory is not reused for another tensor until all current work queued on stream are complete.

Note

The caching allocator is aware of only the stream where a tensor was allocated. Due to the awareness, it already correctly manages the life cycle of tensors on only one stream. But if a tensor is used on a stream different from the stream of origin, the allocator might reuse the memory unexpectedly. Calling this method lets the allocator know which streams have used the tensor.

register_hook(hook) [source]

Registers a backward hook.

The hook will be called every time a gradient with respect to the Tensor is computed. The hook should have the following signature:

hook(grad) -> Tensor or None

The hook should not modify its argument, but it can optionally return a new gradient which will be used in place of grad.

This function returns a handle with a method handle.remove() that removes the hook from the module.

Example:

>>> v = torch.tensor([0., 0., 0.], requires_grad=True)
>>> h = v.register_hook(lambda grad: grad * 2)  # double the gradient
>>> v.backward(torch.tensor([1., 2., 3.]))
>>> v.grad

 2
 4
 6
[torch.FloatTensor of size (3,)]

>>> h.remove()  # removes the hook

remainder(divisor) → Tensor: See torch.remainder()

remainder_(divisor) → Tensor: In-place version of remainder()

renorm(p, dim, maxnorm) → Tensor: See torch.renorm()

renorm_(p, dim, maxnorm) → Tensor: In-place version of renorm()

repeat(*sizes) → Tensor

Repeats this tensor along the specified dimensions.

Unlike expand(), this function copies the tensor’s data.

Warning

repeat() behaves differently from numpy.repeat, but is more similar to numpy.tile. For the operator similar to numpy.repeat, see torch.repeat_interleave().

Parameters: sizes (torch.Size or int...) – The number of times to repeat this tensor along each dimension

Example:

>>> x = torch.tensor([1, 2, 3])
>>> x.repeat(4, 2)
tensor([[ 1,  2,  3,  1,  2,  3],
        [ 1,  2,  3,  1,  2,  3],
        [ 1,  2,  3,  1,  2,  3],
        [ 1,  2,  3,  1,  2,  3]])
>>> x.repeat(4, 2, 1).size()
torch.Size([4, 2, 3])

repeat_interleave(repeats, dim=None) → Tensor: See torch.repeat_interleave().

requires_grad: Is True if gradients need to be computed for this Tensor, False otherwise.

Note

The fact that gradients need to be computed for a Tensor do not mean that the grad attribute will be populated, see is_leaf for more details.

requires_grad_(requires_grad=True) → Tensor

Change if autograd should record operations on this tensor: sets this tensor’s requires_grad attribute in-place. Returns this tensor.

requires_grad_()’s main use case is to tell autograd to begin recording operations on a Tensor tensor. If tensor has requires_grad=False (because it was obtained through a DataLoader, or required preprocessing or initialization), tensor.requires_grad_() makes it so that autograd will begin to record operations on tensor.

Parameters: requires_grad (bool) – If autograd should record operations on this tensor. Default: True.

Example:

>>> # Let's say we want to preprocess some saved weights and use
>>> # the result as new weights.
>>> saved_weights = [0.1, 0.2, 0.3, 0.25]
>>> loaded_weights = torch.tensor(saved_weights)
>>> weights = preprocess(loaded_weights)  # some function
>>> weights
tensor([-0.5503,  0.4926, -2.1158, -0.8303])

>>> # Now, start to record operations done to weights
>>> weights.requires_grad_()
>>> out = weights.pow(2).sum()
>>> out.backward()
>>> weights.grad
tensor([-1.1007,  0.9853, -4.2316, -1.6606])

reshape(*shape) → Tensor

Returns a tensor with the same data and number of elements as self but with the specified shape. This method returns a view if shape is compatible with the current shape. See torch.Tensor.view() on when it is possible to return a view.

See torch.reshape()

Parameters: shape (tuple of python:ints or int...) – the desired shape

reshape_as(other) → Tensor

Returns this tensor as the same shape as other. self.reshape_as(other) is equivalent to self.reshape(other.sizes()). This method returns a view if other.sizes() is compatible with the current shape. See torch.Tensor.view() on when it is possible to return a view.

Please see reshape() for more information about reshape.

Parameters: other (torch.Tensor) – The result tensor has the same shape as other.

resize_(*sizes, memory_format=torch.contiguous_format) → Tensor

Resizes self tensor to the specified size. If the number of elements is larger than the current storage size, then the underlying storage is resized to fit the new number of elements. If the number of elements is smaller, the underlying storage is not changed. Existing elements are preserved but any new memory is uninitialized.

Warning

This is a low-level method. The storage is reinterpreted as C-contiguous, ignoring the current strides (unless the target size equals the current size, in which case the tensor is left unchanged). For most purposes, you will instead want to use view(), which checks for contiguity, or reshape(), which copies data if needed. To change the size in-place with custom strides, see set_().

Parameters

sizes (torch.Size or int...) – the desired size
memory_format (torch.memory_format, optional) – the desired memory format of Tensor. Default: torch.contiguous_format. Note that memory format of self is going to be unaffected if self.size() matches sizes.

Example:

>>> x = torch.tensor([[1, 2], [3, 4], [5, 6]])
>>> x.resize_(2, 2)
tensor([[ 1,  2],
        [ 3,  4]])

resize_as_(tensor, memory_format=torch.contiguous_format) → Tensor

Resizes the self tensor to be the same size as the specified tensor. This is equivalent to self.resize_(tensor.size()).

Parameters: memory_format (torch.memory_format, optional) – the desired memory format of Tensor. Default: torch.contiguous_format. Note that memory format of self is going to be unaffected if self.size() matches tensor.size().

retain_grad() [source]: Enables .grad attribute for non-leaf Tensors.

roll(shifts, dims) → Tensor: See torch.roll()

rot90(k, dims) → Tensor: See torch.rot90()

round() → Tensor: See torch.round()

round_() → Tensor: In-place version of round()

rsqrt() → Tensor: See torch.rsqrt()

rsqrt_() → Tensor: In-place version of rsqrt()

scatter(dim, index, src) → Tensor: Out-of-place version of torch.Tensor.scatter_()

scatter_(dim, index, src, reduce=None) → Tensor

Writes all values from the tensor src into self at the indices specified in the index tensor. For each value in src, its output index is specified by its index in src for dimension != dim and by the corresponding value in index for dimension = dim.

For a 3-D tensor, self is updated as:

self[index[i][j][k]][j][k] = src[i][j][k]  # if dim == 0
self[i][index[i][j][k]][k] = src[i][j][k]  # if dim == 1
self[i][j][index[i][j][k]] = src[i][j][k]  # if dim == 2

This is the reverse operation of the manner described in gather().

self, index and src (if it is a Tensor) should all have the same number of dimensions. It is also required that index.size(d) <= src.size(d) for all dimensions d, and that index.size(d) <= self.size(d) for all dimensions d != dim. Note that index and src do not broadcast.

Moreover, as for gather(), the values of index must be between 0 and self.size(dim) - 1 inclusive.

Warning

When indices are not unique, the behavior is non-deterministic (one of the values from src will be picked arbitrarily) and the gradient will be incorrect (it will be propagated to all locations in the source that correspond to the same index)!

Note

The backward pass is implemented only for src.shape == index.shape.

Additionally accepts an optional reduce argument that allows specification of an optional reduction operation, which is applied to all values in the tensor src into self at the indicies specified in the index. For each value in src, the reduction operation is applied to an index in self which is specified by its index in src for dimension != dim and by the corresponding value in index for dimension = dim.

Given a 3-D tensor and reduction using the multiplication operation, self is updated as:

self[index[i][j][k]][j][k] *= src[i][j][k]  # if dim == 0
self[i][index[i][j][k]][k] *= src[i][j][k]  # if dim == 1
self[i][j][index[i][j][k]] *= src[i][j][k]  # if dim == 2

Reducing with the addition operation is the same as using scatter_add_().

Parameters

dim (int) – the axis along which to index
index (LongTensor) – the indices of elements to scatter, can be either empty or of the same dimensionality as src. When empty, the operation returns self unchanged.
src (Tensor or float) – the source element(s) to scatter.
reduce (str, optional) – reduction operation to apply, can be either 'add' or 'multiply'.

Example:

>>> src = torch.arange(1, 11).reshape((2, 5))
>>> src
tensor([[ 1,  2,  3,  4,  5],
        [ 6,  7,  8,  9, 10]])
>>> index = torch.tensor([[0, 1, 2, 0]])
>>> torch.zeros(3, 5, dtype=src.dtype).scatter_(0, index, src)
tensor([[1, 0, 0, 4, 0],
        [0, 2, 0, 0, 0],
        [0, 0, 3, 0, 0]])
>>> index = torch.tensor([[0, 1, 2], [0, 1, 4]])
>>> torch.zeros(3, 5, dtype=src.dtype).scatter_(1, index, src)
tensor([[1, 2, 3, 0, 0],
        [6, 7, 0, 0, 8],
        [0, 0, 0, 0, 0]])

>>> torch.full((2, 4), 2.).scatter_(1, torch.tensor([[2], [3]]),
...            1.23, reduce='multiply')
tensor([[2.0000, 2.0000, 2.4600, 2.0000],
        [2.0000, 2.0000, 2.0000, 2.4600]])
>>> torch.full((2, 4), 2.).scatter_(1, torch.tensor([[2], [3]]),
...            1.23, reduce='add')
tensor([[2.0000, 2.0000, 3.2300, 2.0000],
        [2.0000, 2.0000, 2.0000, 3.2300]])

scatter_add_(dim, index, src) → Tensor

Adds all values from the tensor other into self at the indices specified in the index tensor in a similar fashion as scatter_(). For each value in src, it is added to an index in self which is specified by its index in src for dimension != dim and by the corresponding value in index for dimension = dim.

For a 3-D tensor, self is updated as:

self[index[i][j][k]][j][k] += src[i][j][k]  # if dim == 0
self[i][index[i][j][k]][k] += src[i][j][k]  # if dim == 1
self[i][j][index[i][j][k]] += src[i][j][k]  # if dim == 2

self, index and src should have same number of dimensions. It is also required that index.size(d) <= src.size(d) for all dimensions d, and that index.size(d) <= self.size(d) for all dimensions d != dim. Note that index and src do not broadcast.

Note

This operation may behave nondeterministically when given tensors on a CUDA device. See Reproducibility for more information.

Note

The backward pass is implemented only for src.shape == index.shape.

Parameters

dim (int) – the axis along which to index
index (LongTensor) – the indices of elements to scatter and add, can be either empty or of the same dimensionality as src. When empty, the operation returns self unchanged.
src (Tensor) – the source elements to scatter and add

Example:

>>> src = torch.ones((2, 5))
>>> index = torch.tensor([[0, 1, 2, 0, 0]])
>>> torch.zeros(3, 5, dtype=src.dtype).scatter_add_(0, index, src)
tensor([[1., 0., 0., 1., 1.],
        [0., 1., 0., 0., 0.],
        [0., 0., 1., 0., 0.]])
>>> index = torch.tensor([[0, 1, 2, 0, 0], [0, 1, 2, 2, 2]])
>>> torch.zeros(3, 5, dtype=src.dtype).scatter_add_(0, index, src)
tensor([[2., 0., 0., 1., 1.],
        [0., 2., 0., 0., 0.],
        [0., 0., 2., 1., 1.]])

scatter_add(dim, index, src) → Tensor: Out-of-place version of torch.Tensor.scatter_add_()

select(dim, index) → Tensor

Slices the self tensor along the selected dimension at the given index. This function returns a view of the original tensor with the given dimension removed.

Parameters

dim (int) – the dimension to slice
index (int) – the index to select with

Note

select() is equivalent to slicing. For example, tensor.select(0, index) is equivalent to tensor[index] and tensor.select(2, index) is equivalent to tensor[:,:,index].

set_(source=None, storage_offset=0, size=None, stride=None) → Tensor

Sets the underlying storage, size, and strides. If source is a tensor, self tensor will share the same storage and have the same size and strides as source. Changes to elements in one tensor will be reflected in the other.

If source is a Storage, the method sets the underlying storage, offset, size, and stride.

Parameters

source (Tensor or Storage) – the tensor or storage to use
storage_offset (int, optional) – the offset in the storage
size (torch.Size, optional) – the desired size. Defaults to the size of the source.
stride (tuple, optional) – the desired stride. Defaults to C-contiguous strides.

share_memory_() [source]

Moves the underlying storage to shared memory.

This is a no-op if the underlying storage is already in shared memory and for CUDA tensors. Tensors in shared memory cannot be resized.

short(memory_format=torch.preserve_format) → Tensor

self.short() is equivalent to self.to(torch.int16). See to().

Parameters: memory_format (torch.memory_format, optional) – the desired memory format of returned Tensor. Default: torch.preserve_format.

sigmoid() → Tensor: See torch.sigmoid()

sigmoid_() → Tensor: In-place version of sigmoid()

sign() → Tensor: See torch.sign()

sign_() → Tensor: In-place version of sign()

signbit() → Tensor: See torch.signbit()

sgn() → Tensor: See torch.sgn()

sgn_() → Tensor: In-place version of sgn()

sin() → Tensor: See torch.sin()

sin_() → Tensor: In-place version of sin()

sinc() → Tensor: See torch.sinc()

sinc_() → Tensor: In-place version of sinc()

sinh() → Tensor: See torch.sinh()

sinh_() → Tensor: In-place version of sinh()

asinh() → Tensor: See torch.asinh()

asinh_() → Tensor: In-place version of asinh()

arcsinh() → Tensor: See torch.arcsinh()

arcsinh_() → Tensor: In-place version of arcsinh()

size() → torch.Size

Returns the size of the self tensor. The returned value is a subclass of tuple.

Example:

>>> torch.empty(3, 4, 5).size()
torch.Size([3, 4, 5])

slogdet() -> (Tensor, Tensor): See torch.slogdet()

solve(A) → Tensor, Tensor: See torch.solve()

sort(dim=-1, descending=False) -> (Tensor, LongTensor): See torch.sort()

split(split_size, dim=0) [source]: See torch.split()

sparse_mask(mask) → Tensor

Returns a new sparse tensor with values from a strided tensor self filtered by the indices of the sparse tensor mask. The values of mask sparse tensor are ignored. self and mask tensors must have the same shape.

Note

The returned sparse tensor has the same indices as the sparse tensor mask, even when the corresponding values in self are zeros.

Parameters: mask (Tensor) – a sparse tensor whose indices are used as a filter

Example:

>>> nse = 5
>>> dims = (5, 5, 2, 2)
>>> I = torch.cat([torch.randint(0, dims[0], size=(nse,)),
...                torch.randint(0, dims[1], size=(nse,))], 0).reshape(2, nse)
>>> V = torch.randn(nse, dims[2], dims[3])
>>> S = torch.sparse_coo_tensor(I, V, dims).coalesce()
>>> D = torch.randn(dims)
>>> D.sparse_mask(S)
tensor(indices=tensor([[0, 0, 0, 2],
                       [0, 1, 4, 3]]),
       values=tensor([[[ 1.6550,  0.2397],
                       [-0.1611, -0.0779]],

                      [[ 0.2326, -1.0558],
                       [ 1.4711,  1.9678]],

                      [[-0.5138, -0.0411],
                       [ 1.9417,  0.5158]],

                      [[ 0.0793,  0.0036],
                       [-0.2569, -0.1055]]]),
       size=(5, 5, 2, 2), nnz=4, layout=torch.sparse_coo)

sparse_dim() → int

Return the number of sparse dimensions in a sparse tensor self.

Warning

Throws an error if self is not a sparse tensor.

See also Tensor.dense_dim() and hybrid tensors.

sqrt() → Tensor: See torch.sqrt()

sqrt_() → Tensor: In-place version of sqrt()

square() → Tensor: See torch.square()

square_() → Tensor: In-place version of square()

squeeze(dim=None) → Tensor: See torch.squeeze()

squeeze_(dim=None) → Tensor: In-place version of squeeze()

std(dim=None, unbiased=True, keepdim=False) → Tensor: See torch.std()

stft(n_fft, hop_length=None, win_length=None, window=None, center=True, pad_mode='reflect', normalized=False, onesided=None, return_complex=None) [source]: See torch.stft()

Warning

This function changed signature at version 0.4.1. Calling with the previous signature may cause error or return incorrect result.

storage() → torch.Storage: Returns the underlying storage.

storage_offset() → int

Returns self tensor’s offset in the underlying storage in terms of number of storage elements (not bytes).

Example:

>>> x = torch.tensor([1, 2, 3, 4, 5])
>>> x.storage_offset()
0
>>> x[3:].storage_offset()
3

storage_type() → type: Returns the type of the underlying storage.

stride(dim) → tuple or int

Returns the stride of self tensor.

Stride is the jump necessary to go from one element to the next one in the specified dimension dim. A tuple of all strides is returned when no argument is passed in. Otherwise, an integer value is returned as the stride in the particular dimension dim.

Parameters: dim (int, optional) – the desired dimension in which stride is required

Example:

>>> x = torch.tensor([[1, 2, 3, 4, 5], [6, 7, 8, 9, 10]])
>>> x.stride()
(5, 1)
>>> x.stride(0)
5
>>> x.stride(-1)
1

sub(other, *, alpha=1) → Tensor: See torch.sub().

sub_(other, *, alpha=1) → Tensor: In-place version of sub()

subtract(other, *, alpha=1) → Tensor: See torch.subtract().

subtract_(other, *, alpha=1) → Tensor: In-place version of subtract().

sum(dim=None, keepdim=False, dtype=None) → Tensor: See torch.sum()

sum_to_size(*size) → Tensor

Sum this tensor to size. size must be broadcastable to this tensor size.

Parameters: size (int...) – a sequence of integers defining the shape of the output tensor.

svd(some=True, compute_uv=True) -> (Tensor, Tensor, Tensor): See torch.svd()

swapaxes(axis0, axis1) → Tensor: See torch.swapaxes()

swapdims(dim0, dim1) → Tensor: See torch.swapdims()

symeig(eigenvectors=False, upper=True) -> (Tensor, Tensor): See torch.symeig()

t() → Tensor: See torch.t()

t_() → Tensor: In-place version of t()

tensor_split(indices_or_sections, dim=0) → List of Tensors: See torch.tensor_split()

tile(*reps) → Tensor: See torch.tile()

to(*args, **kwargs) → Tensor

Performs Tensor dtype and/or device conversion. A torch.dtype and torch.device are inferred from the arguments of self.to(*args, **kwargs).

Note

If the self Tensor already has the correct torch.dtype and torch.device, then self is returned. Otherwise, the returned tensor is a copy of self with the desired torch.dtype and torch.device.

Here are the ways to call to:

to(dtype, non_blocking=False, copy=False, memory_format=torch.preserve_format) → Tensor

Returns a Tensor with the specified dtype

Args:: memory_format (torch.memory_format, optional): the desired memory format of returned Tensor. Default: torch.preserve_format.

to(device=None, dtype=None, non_blocking=False, copy=False, memory_format=torch.preserve_format) → Tensor

Returns a Tensor with the specified device and (optional) dtype. If dtype is None it is inferred to be self.dtype. When non_blocking, tries to convert asynchronously with respect to the host if possible, e.g., converting a CPU Tensor with pinned memory to a CUDA Tensor. When copy is set, a new Tensor is created even when the Tensor already matches the desired conversion.

Args:: memory_format (torch.memory_format, optional): the desired memory format of returned Tensor. Default: torch.preserve_format.

to(other, non_blocking=False, copy=False) → Tensor: Returns a Tensor with same torch.dtype and torch.device as the Tensor other. When non_blocking, tries to convert asynchronously with respect to the host if possible, e.g., converting a CPU Tensor with pinned memory to a CUDA Tensor. When copy is set, a new Tensor is created even when the Tensor already matches the desired conversion.

Example:

>>> tensor = torch.randn(2, 2)  # Initially dtype=float32, device=cpu
>>> tensor.to(torch.float64)
tensor([[-0.5044,  0.0005],
        [ 0.3310, -0.0584]], dtype=torch.float64)

>>> cuda0 = torch.device('cuda:0')
>>> tensor.to(cuda0)
tensor([[-0.5044,  0.0005],
        [ 0.3310, -0.0584]], device='cuda:0')

>>> tensor.to(cuda0, dtype=torch.float64)
tensor([[-0.5044,  0.0005],
        [ 0.3310, -0.0584]], dtype=torch.float64, device='cuda:0')

>>> other = torch.randn((), dtype=torch.float64, device=cuda0)
>>> tensor.to(other, non_blocking=True)
tensor([[-0.5044,  0.0005],
        [ 0.3310, -0.0584]], dtype=torch.float64, device='cuda:0')

to_mkldnn() → Tensor: Returns a copy of the tensor in torch.mkldnn layout.

take(indices) → Tensor: See torch.take()

tan() → Tensor: See torch.tan()

tan_() → Tensor: In-place version of tan()

tanh() → Tensor: See torch.tanh()

tanh_() → Tensor: In-place version of tanh()

atanh() → Tensor: See torch.atanh()

atanh_(other) → Tensor: In-place version of atanh()

arctanh() → Tensor: See torch.arctanh()

arctanh_(other) → Tensor: In-place version of arctanh()

tolist() → list or number

Returns the tensor as a (nested) list. For scalars, a standard Python number is returned, just like with item(). Tensors are automatically moved to the CPU first if necessary.

This operation is not differentiable.

Examples:

>>> a = torch.randn(2, 2)
>>> a.tolist()
[[0.012766935862600803, 0.5415473580360413],
 [-0.08909505605697632, 0.7729271650314331]]
>>> a[0,0].tolist()
0.012766935862600803

topk(k, dim=None, largest=True, sorted=True) -> (Tensor, LongTensor): See torch.topk()

to_sparse(sparseDims) → Tensor

Returns a sparse copy of the tensor. PyTorch supports sparse tensors in coordinate format.

Parameters: sparseDims (int, optional) – the number of sparse dimensions to include in the new sparse tensor

Example:

>>> d = torch.tensor([[0, 0, 0], [9, 0, 10], [0, 0, 0]])
>>> d
tensor([[ 0,  0,  0],
        [ 9,  0, 10],
        [ 0,  0,  0]])
>>> d.to_sparse()
tensor(indices=tensor([[1, 1],
                       [0, 2]]),
       values=tensor([ 9, 10]),
       size=(3, 3), nnz=2, layout=torch.sparse_coo)
>>> d.to_sparse(1)
tensor(indices=tensor([[1]]),
       values=tensor([[ 9,  0, 10]]),
       size=(3, 3), nnz=1, layout=torch.sparse_coo)

trace() → Tensor: See torch.trace()

transpose(dim0, dim1) → Tensor: See torch.transpose()

transpose_(dim0, dim1) → Tensor: In-place version of transpose()

triangular_solve(A, upper=True, transpose=False, unitriangular=False) -> (Tensor, Tensor): See torch.triangular_solve()

tril(k=0) → Tensor: See torch.tril()

tril_(k=0) → Tensor: In-place version of tril()

triu(k=0) → Tensor: See torch.triu()

triu_(k=0) → Tensor: In-place version of triu()

true_divide(value) → Tensor: See torch.true_divide()

true_divide_(value) → Tensor: In-place version of true_divide_()

trunc() → Tensor: See torch.trunc()

trunc_() → Tensor: In-place version of trunc()

type(dtype=None, non_blocking=False, **kwargs) → str or Tensor

Returns the type if dtype is not provided, else casts this object to the specified type.

If this is already of the correct type, no copy is performed and the original object is returned.

Parameters

dtype (type or string) – The desired type
non_blocking (bool) – If True, and the source is in pinned memory and destination is on the GPU or vice versa, the copy is performed asynchronously with respect to the host. Otherwise, the argument has no effect.
**kwargs – For compatibility, may contain the key async in place of the non_blocking argument. The async arg is deprecated.

type_as(tensor) → Tensor

Returns this tensor cast to the type of the given tensor.

This is a no-op if the tensor is already of the correct type. This is equivalent to self.type(tensor.type())

Parameters: tensor (Tensor) – the tensor which has the desired type

unbind(dim=0) → seq: See torch.unbind()

unfold(dimension, size, step) → Tensor

Returns a view of the original tensor which contains all slices of size size from self tensor in the dimension dimension.

Step between two slices is given by step.

If sizedim is the size of dimension dimension for self, the size of dimension dimension in the returned tensor will be (sizedim - size) / step + 1.

An additional dimension of size size is appended in the returned tensor.

Parameters

dimension (int) – dimension in which unfolding happens
size (int) – the size of each slice that is unfolded
step (int) – the step between each slice

Example:

>>> x = torch.arange(1., 8)
>>> x
tensor([ 1.,  2.,  3.,  4.,  5.,  6.,  7.])
>>> x.unfold(0, 2, 1)
tensor([[ 1.,  2.],
        [ 2.,  3.],
        [ 3.,  4.],
        [ 4.,  5.],
        [ 5.,  6.],
        [ 6.,  7.]])
>>> x.unfold(0, 2, 2)
tensor([[ 1.,  2.],
        [ 3.,  4.],
        [ 5.,  6.]])

uniform_(from=0, to=1) → Tensor: Fills self tensor with numbers sampled from the continuous uniform distribution:

$P(x) = \dfrac{1}{\text{to} - \text{from}}$

unique(sorted=True, return_inverse=False, return_counts=False, dim=None) [source]

Returns the unique elements of the input tensor.

See torch.unique()

unique_consecutive(return_inverse=False, return_counts=False, dim=None) [source]

Eliminates all but the first element from every consecutive group of equivalent elements.

See torch.unique_consecutive()

unsqueeze(dim) → Tensor: See torch.unsqueeze()

unsqueeze_(dim) → Tensor: In-place version of unsqueeze()

values() → Tensor

Return the values tensor of a sparse COO tensor.

Warning

Throws an error if self is not a sparse COO tensor.