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()
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

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

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

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

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
abs_() → Tensor

In-place version of abs()

absolute() → Tensor

Alias for abs()

absolute_() → Tensor

In-place version of absolute() Alias for abs_()

acos() → Tensor
acos_() → Tensor

In-place version of acos()

arccos() → Tensor
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

add_(other, *, alpha=1) → Tensor

In-place version of add()

addbmm(batch1, batch2, *, beta=1, alpha=1) → Tensor
addbmm_(batch1, batch2, *, beta=1, alpha=1) → Tensor

In-place version of addbmm()

addcdiv(tensor1, tensor2, *, value=1) → Tensor
addcdiv_(tensor1, tensor2, *, value=1) → Tensor

In-place version of addcdiv()

addcmul(tensor1, tensor2, *, value=1) → Tensor
addcmul_(tensor1, tensor2, *, value=1) → Tensor

In-place version of addcmul()

addmm(mat1, mat2, *, beta=1, alpha=1) → Tensor
addmm_(mat1, mat2, *, beta=1, alpha=1) → Tensor

In-place version of addmm()

sspaddmm(mat1, mat2, *, beta=1, alpha=1) → Tensor
addmv(mat, vec, *, beta=1, alpha=1) → Tensor
addmv_(mat, vec, *, beta=1, alpha=1) → Tensor

In-place version of addmv()

addr(vec1, vec2, *, beta=1, alpha=1) → Tensor
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
amax(dim=None, keepdim=False) → Tensor
amin(dim=None, keepdim=False) → Tensor
angle() → Tensor
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
argmin(dim=None, keepdim=False) → LongTensor
argsort(dim=-1, descending=False) → LongTensor
asin() → Tensor
asin_() → Tensor

In-place version of asin()

arcsin() → Tensor
arcsin_() → Tensor

In-place version of arcsin()

as_strided(size, stride, storage_offset=0) → Tensor
atan() → Tensor
atan_() → Tensor

In-place version of atan()

arctan() → Tensor
arctan_() → Tensor

In-place version of arctan()

atan2(other) → Tensor
atan2_(other) → Tensor

In-place version of atan2()

all(dim=None, keepdim=False) → Tensor
any(dim=None, keepdim=False) → Tensor
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
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.

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.

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
bitwise_not() → Tensor
bitwise_not_() → Tensor

In-place version of bitwise_not()

bitwise_and() → Tensor
bitwise_and_() → Tensor

In-place version of bitwise_and()

bitwise_or() → Tensor
bitwise_or_() → Tensor

In-place version of bitwise_or()

bitwise_xor() → Tensor
bitwise_xor_() → Tensor

In-place version of bitwise_xor()

bmm(batch2) → Tensor
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
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
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
cholesky_inverse(upper=False) → Tensor
cholesky_solve(input2, upper=False) → Tensor
chunk(chunks, dim=0) → List of Tensors
clamp(min, max) → Tensor
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
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
copysign(other) → Tensor
copysign_(other) → Tensor

In-place version of copysign()

cos() → Tensor
cos_() → Tensor

In-place version of cos()

cosh() → Tensor
cosh_() → Tensor

In-place version of cosh()

count_nonzero(dim=None) → Tensor
acosh() → Tensor
acosh_() → Tensor

In-place version of acosh()

arccosh()

acosh() -> Tensor

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
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
cummax(dim) -> (Tensor, Tensor)
cummin(dim) -> (Tensor, Tensor)
cumprod(dim, dtype=None) → Tensor
cumprod_(dim, dtype=None) → Tensor

In-place version of cumprod()

cumsum(dim, dtype=None) → Tensor
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
dequantize() → Tensor

Given a quantized Tensor, dequantize it and return the dequantized float Tensor.

det() → Tensor
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.

See also Tensor.sparse_dim() and hybrid tensors.

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
diag_embed(offset=0, dim1=-2, dim2=-1) → Tensor
diagflat(offset=0) → Tensor
diagonal(offset=0, dim1=0, dim2=1) → Tensor
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
fmin(other) → Tensor
digamma() → Tensor
digamma_() → Tensor

In-place version of digamma()

dim() → int

Returns the number of dimensions of self tensor.

dist(other, p=2) → Tensor
div(value) → Tensor
div_(value) → Tensor

In-place version of div()

divide(value) → Tensor
divide_(value) → Tensor

In-place version of divide()

dot(other) → Tensor
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)
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
eq_(other) → Tensor

In-place version of eq()

equal(other) → bool
erf() → Tensor
erf_() → Tensor

In-place version of erf()

erfc() → Tensor
erfc_() → Tensor

In-place version of erfc()

erfinv() → Tensor
erfinv_() → Tensor

In-place version of erfinv()

exp() → Tensor
exp_() → Tensor

In-place version of exp()

expm1() → Tensor
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
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
flip(dims) → Tensor
fliplr() → Tensor
flipud() → Tensor
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
float_power_(exponent) → Tensor

In-place version of float_power()

floor() → Tensor
floor_() → Tensor

In-place version of floor()

floor_divide(value) → Tensor
floor_divide_(value) → Tensor

In-place version of floor_divide()

fmod(divisor) → Tensor
fmod_(divisor) → Tensor

In-place version of fmod()

frac() → Tensor
frac_() → Tensor

In-place version of frac()

gather(dim, index) → Tensor
gcd(other) → Tensor
gcd_(other) → Tensor

In-place version of gcd()

ge(other) → Tensor
ge_(other) → Tensor

In-place version of ge().

greater_equal(other) → Tensor
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)
ger(vec2) → Tensor
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
gt_(other) → Tensor

In-place version of gt().

greater(other) → Tensor
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
heaviside(values) → Tensor
histc(bins=100, min=0, max=0) → Tensor
hypot(other) → Tensor
hypot_(other) → Tensor

In-place version of hypot()

i0() → Tensor
i0_() → Tensor

In-place version of i0()

igamma(other) → Tensor
igamma_(other) → Tensor

In-place version of igamma()

igammac(other) → Tensor
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])
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
indices() → Tensor

Return the indices tensor of a sparse COO tensor.

Warning

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

See also Tensor.values().

Note

This method can only be called on a coalesced sparse tensor. See Tensor.coalesce() for details.

inner(other) → Tensor
int(memory_format=torch.preserve_format) → Tensor

self.int() is equivalent to self.to(torch.int32). See to().

Parameters

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

int_repr() → Tensor

Given a quantized Tensor, self.int_repr() returns a CPU Tensor with uint8_t as data type that stores the underlying uint8_t values of the given Tensor.

inverse() → Tensor
isclose(other, rtol=1e-05, atol=1e-08, equal_nan=False) → Tensor
isfinite() → Tensor
isinf() → Tensor
isposinf() → Tensor
isneginf() → Tensor
isnan() → Tensor
is_contiguous(memory_format=torch.contiguous_format) → bool

Returns True if self tensor is contiguous in memory in the order specified by memory format.

Parameters

memory_format (torch.memory_format, optional) – Specifies memory allocation order. Default: torch.contiguous_format.

is_complex() → bool

Returns True if the data type of self is a complex data type.

is_floating_point() → bool

Returns True if the data type of self is a floating point data type.

is_leaf

All Tensors that have requires_grad which is False will be leaf Tensors by convention.

For Tensors that have requires_grad which is True, they will be leaf Tensors if they were created by the user. This means that they are not the result of an operation and so grad_fn is None.

Only leaf Tensors will have their grad populated during a call to backward(). To get grad populated for non-leaf Tensors, you can use retain_grad().

Example:

>>> a = torch.rand(10, requires_grad=True)
>>> a.is_leaf
True
>>> b.is_leaf
False
# b was created by the operation that cast a cpu Tensor into a cuda Tensor
>>> c = torch.rand(10, requires_grad=True) + 2
>>> c.is_leaf
False
# c was created by the addition operation
>>> d = torch.rand(10).cuda()
>>> d.is_leaf
True
# d does not require gradients and so has no operation creating it (that is tracked by the autograd engine)
>>> e.is_leaf
True
# e requires gradients and has no operations creating it
>>> f = torch.rand(10, requires_grad=True, device="cuda")
>>> f.is_leaf
True
# f requires grad, has no operation creating it

is_pinned()

Returns true if this tensor resides in pinned memory.

is_set_to(tensor) → bool

Returns True if both tensors are pointing to the exact same memory (same storage, offset, size and stride).

is_shared()[source]

Checks if tensor is in shared memory.

This is always True for CUDA tensors.

is_signed() → bool

Returns True if the data type of self is a signed data type.

is_sparse

Is True if the Tensor uses sparse storage layout, False otherwise.

istft(n_fft, hop_length=None, win_length=None, window=None, center=True, normalized=False, onesided=None, length=None, return_complex=False)[source]
isreal() → Tensor
item() → number

Returns the value of this tensor as a standard Python number. This only works for tensors with one element. For other cases, see tolist().

This operation is not differentiable.

Example:

>>> x = torch.tensor([1.0])
>>> x.item()
1.0

kthvalue(k, dim=None, keepdim=False) -> (Tensor, LongTensor)
lcm(other) → Tensor
lcm_(other) → Tensor

In-place version of lcm()

ldexp(other) → Tensor
ldexp_(other) → Tensor

In-place version of ldexp()

le(other) → Tensor
le_(other) → Tensor

In-place version of le().

less_equal(other) → Tensor
less_equal_(other) → Tensor

In-place version of less_equal().

lerp(end, weight) → Tensor
lerp_(end, weight) → Tensor

In-place version of lerp()

lgamma() → Tensor
lgamma_() → Tensor

In-place version of lgamma()

log() → Tensor
log_() → Tensor

In-place version of log()

logdet() → Tensor
log10() → Tensor
log10_() → Tensor

In-place version of log10()

log1p() → Tensor
log1p_() → Tensor

In-place version of log1p()

log2() → Tensor
log2_() → Tensor

In-place version of log2()

log_normal_(mean=1, std=2, *, generator=None)

Fills self tensor with numbers samples from the log-normal distribution parameterized by the given mean $\mu$ and standard deviation $\sigma$ . Note that mean and std are the mean and standard deviation of the underlying normal distribution, and not of the returned distribution:

$f(x) = \dfrac{1}{x \sigma \sqrt{2\pi}}\ e^{-\frac{(\ln x - \mu)^2}{2\sigma^2}}$
logaddexp(other) → Tensor
logaddexp2(other) → Tensor
logsumexp(dim, keepdim=False) → Tensor
logical_and() → Tensor
logical_and_() → Tensor

In-place version of logical_and()

logical_not() → Tensor
logical_not_() → Tensor

In-place version of logical_not()

logical_or() → Tensor
logical_or_() → Tensor

In-place version of logical_or()

logical_xor() → Tensor
logical_xor_() → Tensor

In-place version of logical_xor()

logit() → Tensor
logit_() → Tensor

In-place version of logit()

long(memory_format=torch.preserve_format) → Tensor

self.long() is equivalent to self.to(torch.int64). See to().

Parameters

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

lstsq(A) -> (Tensor, Tensor)
lt(other) → Tensor
lt_(other) → Tensor

In-place version of lt().

less()

lt(other) -> Tensor

less_(other) → Tensor

In-place version of less().

lu(pivot=True, get_infos=False)[source]
lu_solve(LU_data, LU_pivots) → Tensor
as_subclass(cls) → Tensor

Makes a cls instance with the same data pointer as self. Changes in the output mirror changes in self, and the output stays attached to the autograd graph. cls must be a subclass of Tensor.

map_(tensor, callable)

Applies callable for each element in self tensor and the given tensor and stores the results in self tensor. self tensor and the given tensor must be broadcastable.

The callable should have the signature:

def callable(a, b) -> number

masked_scatter_(mask, source)

Copies elements from source into self tensor at positions where the mask is True. The shape of mask must be broadcastable with the shape of the underlying tensor. The source should have at least as many elements as the number of ones in mask

Parameters

• source (Tensor) – the tensor to copy from

Note

The mask operates on the self tensor, not on the given source tensor.

masked_scatter(mask, tensor) → Tensor

Out-of-place version of torch.Tensor.masked_scatter_()

masked_fill_(mask, value)

Fills elements of self tensor with value where mask is True. The shape of mask must be broadcastable with the shape of the underlying tensor.

Parameters

• value (float) – the value to fill in with

masked_fill(mask, value) → Tensor

Out-of-place version of torch.Tensor.masked_fill_()

masked_select(mask) → Tensor
matmul(tensor2) → Tensor
matrix_power(n) → Tensor
matrix_exp() → Tensor
max(dim=None, keepdim=False) -> Tensor or (Tensor, Tensor)
maximum(other) → Tensor
mean(dim=None, keepdim=False) -> Tensor or (Tensor, Tensor)
median(dim=None, keepdim=False) -> (Tensor, LongTensor)
nanmedian(dim=None, keepdim=False) -> (Tensor, LongTensor)
min(dim=None, keepdim=False) -> Tensor or (Tensor, Tensor)
minimum(other) → Tensor
mm(mat2) → Tensor
smm(mat) → Tensor
mode(dim=None, keepdim=False) -> (Tensor, LongTensor)
movedim(source, destination) → Tensor
moveaxis(source, destination) → Tensor
msort() → Tensor
mul(value) → Tensor
mul_(value) → Tensor

In-place version of mul().

multiply(value) → Tensor
multiply_(value) → Tensor

In-place version of multiply().

multinomial(num_samples, replacement=False, *, generator=None) → Tensor
mv(vec) → Tensor
mvlgamma(p) → Tensor
mvlgamma_(p) → Tensor

In-place version of mvlgamma()

nansum(dim=None, keepdim=False, dtype=None) → Tensor
narrow(dimension, start, length) → Tensor

Example:

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

narrow_copy(dimension, start, length) → Tensor

Same as Tensor.narrow() except returning a copy rather than shared storage. This is primarily for sparse tensors, which do not have a shared-storage narrow method. Calling narrow_copy with dimemsion > self.sparse_dim() will return a copy with the relevant dense dimension narrowed, and self.shape updated accordingly.

ndimension() → int

Alias for dim()

nan_to_num(nan=0.0, posinf=None, neginf=None) → Tensor
nan_to_num_(nan=0.0, posinf=None, neginf=None) → Tensor

In-place version of nan_to_num().

ne(other) → Tensor
ne_(other) → Tensor

In-place version of ne().

not_equal(other) → Tensor
not_equal_(other) → Tensor

In-place version of not_equal().

neg() → Tensor
neg_() → Tensor

In-place version of neg()

negative() → Tensor
negative_() → Tensor

In-place version of negative()

nelement() → int

Alias for numel()

nextafter(other) → Tensor
nextafter_(other) → Tensor

In-place version of nextafter()

nonzero() → LongTensor
norm(p='fro', dim=None, keepdim=False, dtype=None)[source]
normal_(mean=0, std=1, *, generator=None) → Tensor

Fills self tensor with elements samples from the normal distribution parameterized by mean and std.

numel() → int
numpy() → numpy.ndarray

Returns self tensor as a NumPy ndarray. This tensor and the returned ndarray share the same underlying storage. Changes to self tensor will be reflected in the ndarray and vice versa.

orgqr(input2) → Tensor
ormqr(input2, input3, left=True, transpose=False) → Tensor
outer(vec2) → Tensor
permute(*dims) → Tensor

Returns a view of the original tensor with its dimensions permuted.

Parameters

*dims (int...) – The desired ordering of dimensions

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
polygamma(n) → Tensor
polygamma_(n) → Tensor

In-place version of polygamma()

pow(exponent) → Tensor
pow_(exponent) → Tensor

In-place version of pow()

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

Returns the quantization scheme of a given QTensor.

quantile(q, dim=None, keepdim=False) → Tensor
nanquantile(q, dim=None, keepdim=False) → Tensor
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
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
reciprocal() → Tensor
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)
>>> v.backward(torch.tensor([1., 2., 3.]))

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

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

remainder(divisor) → Tensor
remainder_(divisor) → Tensor

In-place version of remainder()

renorm(p, dim, maxnorm) → Tensor
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
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]
>>> weights = preprocess(loaded_weights)  # some function
>>> weights
tensor([-0.5503,  0.4926, -2.1158, -0.8303])

>>> # Now, start to record operations done to weights
>>> out = weights.pow(2).sum()
>>> out.backward()
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.

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
rot90(k, dims) → Tensor
round() → Tensor
round_() → Tensor

In-place version of round()

rsqrt() → Tensor
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]]),
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
sigmoid_() → Tensor

In-place version of sigmoid()

sign() → Tensor
sign_() → Tensor

In-place version of sign()

signbit() → Tensor
sgn() → Tensor

See torch.sgn()

sgn_() → Tensor

In-place version of sgn()

sin() → Tensor
sin_() → Tensor

In-place version of sin()

sinc() → Tensor
sinc_() → Tensor

In-place version of sinc()

sinh() → Tensor
sinh_() → Tensor

In-place version of sinh()

asinh() → Tensor
asinh_() → Tensor

In-place version of asinh()

arcsinh() → Tensor
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)
solve(A) → Tensor, Tensor
sort(dim=-1, descending=False) -> (Tensor, LongTensor)
split(split_size, dim=0)[source]
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)
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
sqrt_() → Tensor

In-place version of sqrt()

square() → Tensor
square_() → Tensor

In-place version of square()

squeeze(dim=None) → Tensor
squeeze_(dim=None) → Tensor

In-place version of squeeze()

std(dim=None, unbiased=True, keepdim=False) → Tensor
stft(n_fft, hop_length=None, win_length=None, window=None, center=True, pad_mode='reflect', normalized=False, onesided=None, return_complex=None)[source]

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
sub_(other, *, alpha=1) → Tensor

In-place version of sub()

subtract(other, *, alpha=1) → Tensor
subtract_(other, *, alpha=1) → Tensor

In-place version of subtract().

sum(dim=None, keepdim=False, dtype=None) → Tensor
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)
swapaxes(axis0, axis1) → Tensor
swapdims(dim0, dim1) → Tensor
symeig(eigenvectors=False, upper=True) -> (Tensor, Tensor)
t() → Tensor
t_() → Tensor

In-place version of t()

tensor_split(indices_or_sections, dim=0) → List of Tensors
tile(*reps) → Tensor
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
tan() → Tensor
tan_() → Tensor

In-place version of tan()

tanh() → Tensor
tanh_() → Tensor

In-place version of tanh()

atanh() → Tensor
atanh_(other) → Tensor

In-place version of atanh()

arctanh() → Tensor
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)
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
transpose(dim0, dim1) → Tensor
transpose_(dim0, dim1) → Tensor

In-place version of transpose()

triangular_solve(A, upper=True, transpose=False, unitriangular=False) -> (Tensor, Tensor)
tril(k=0) → Tensor
tril_(k=0) → Tensor

In-place version of tril()

triu(k=0) → Tensor
triu_(k=0) → Tensor

In-place version of triu()

true_divide(value) → Tensor
true_divide_(value) → Tensor

In-place version of true_divide_()

trunc() → Tensor
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
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.

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

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

unsqueeze(dim) → Tensor
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.

See also Tensor.indices().

Note

This method can only be called on a coalesced sparse tensor. See Tensor.coalesce() for details.

var(dim=None, unbiased=True, keepdim=False) → Tensor
vdot(other) → Tensor
view(*shape) → Tensor

Returns a new tensor with the same data as the self tensor but of a different shape.

The returned tensor shares the same data and must have the same number of elements, but may have a different size. For a tensor to be viewed, the new view size must be compatible with its original size and stride, i.e., each new view dimension must either be a subspace of an original dimension, or only span across original dimensions $d, d+1, \dots, d+k$ that satisfy the following contiguity-like condition that $\forall i = d, \dots, d+k-1$ ,

$\text{stride}[i] = \text{stride}[i+1] \times \text{size}[i+1]$

Otherwise, it will not be possible to view self tensor as shape without copying it (e.g., via contiguous()). When it is unclear whether a view() can be performed, it is advisable to use reshape(), which returns a view if the shapes are compatible, and copies (equivalent to calling contiguous()) otherwise.

Parameters

shape (torch.Size or int...) – the desired size

Example:

>>> x = torch.randn(4, 4)
>>> x.size()
torch.Size([4, 4])
>>> y = x.view(16)
>>> y.size()
torch.Size([16])
>>> z = x.view(-1, 8)  # the size -1 is inferred from other dimensions
>>> z.size()
torch.Size([2, 8])

>>> a = torch.randn(1, 2, 3, 4)
>>> a.size()
torch.Size([1, 2, 3, 4])
>>> b = a.transpose(1, 2)  # Swaps 2nd and 3rd dimension
>>> b.size()
torch.Size([1, 3, 2, 4])
>>> c = a.view(1, 3, 2, 4)  # Does not change tensor layout in memory
>>> c.size()
torch.Size([1, 3, 2, 4])
>>> torch.equal(b, c)
False

view(dtype) → Tensor

Returns a new tensor with the same data as the self tensor but of a different dtype. dtype must have the same number of bytes per element as self’s dtype.

Warning

This overload is not supported by TorchScript, and using it in a Torchscript program will cause undefined behavior.

Parameters

dtype (torch.dtype) – the desired dtype

Example:

>>> x = torch.randn(4, 4)
>>> x
tensor([[ 0.9482, -0.0310,  1.4999, -0.5316],
[-0.1520,  0.7472,  0.5617, -0.8649],
[-2.4724, -0.0334, -0.2976, -0.8499],
[-0.2109,  1.9913, -0.9607, -0.6123]])
>>> x.dtype
torch.float32

>>> y = x.view(torch.int32)
>>> y
tensor([[ 1064483442, -1124191867,  1069546515, -1089989247],
[-1105482831,  1061112040,  1057999968, -1084397505],
[-1071760287, -1123489973, -1097310419, -1084649136],
[-1101533110,  1073668768, -1082790149, -1088634448]],
dtype=torch.int32)
>>> y[0, 0] = 1000000000
>>> x
tensor([[ 0.0047, -0.0310,  1.4999, -0.5316],
[-0.1520,  0.7472,  0.5617, -0.8649],
[-2.4724, -0.0334, -0.2976, -0.8499],
[-0.2109,  1.9913, -0.9607, -0.6123]])

>>> x.view(torch.int16)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
RuntimeError: Viewing a tensor as a new dtype with a different number of bytes per element is not supported.

view_as(other) → Tensor

View this tensor as the same size as other. self.view_as(other) is equivalent to self.view(other.size()).

Please see view() for more information about view.

Parameters

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

where(condition, y) → Tensor

self.where(condition, y) is equivalent to torch.where(condition, self, y). See torch.where()

xlogy(other) → Tensor
xlogy_(other) → Tensor

In-place version of xlogy()

zero_() → Tensor

Fills self` tensor with zeros.