if center:
signal_dim = input.dim()
extended_shape = [1] * (3 - signal_dim) + list(input.size())
pad = int(n_fft // 2)
input = F.pad(input.view(extended_shape), (pad, pad), pad_mode)
input = input.view(input.shape[-signal_dim:])
return torch._C._VariableFunctions.stft(input, n_fft, hop_length, win_length, window, normalized, onesided)
isnan = _add_docstr(torch.isnan, r"""
Returns a new tensor with boolean elements representing if each element is `NaN` or not.
Arguments:
tensor (Tensor): A tensor to check
Returns:
Tensor: A ``torch.ByteTensor`` containing a 1 at each location of `NaN` elements.
Example::
>>> torch.isnan(torch.tensor([1, float('nan'), 2]))
tensor([ 0, 1, 0], dtype=torch.uint8)
""")
def unique(input, sorted=True, return_inverse=False, dim=None):
r"""Returns the unique scalar elements of the input tensor as a 1-D tensor.
Arguments:
input (Tensor): the input tensor
sorted (bool): Whether to sort the unique elements in ascending order
before returning as output.
import sys
import torch
from torch._C import _add_docstr, _fft # type: ignore
Tensor = torch.Tensor
# Note: This not only adds the doc strings for the spectral ops, but
# connects the torch.fft Python namespace to the torch._C._fft builtins.
fft = _add_docstr(
_fft.fft_fft, r"""
fft(input) -> Tensor
Computes the one dimensional discrete Fourier transform of :attr:`input`.
Args:
input (Tensor): the input tensor
Example::
>>> t = torch.randn(4, dtype=torch.complex128)
>>> t
tensor([-1.1364-0.5694j, -0.6637+0.9987j, -1.0102-0.4383j, 0.3017+0.9371j], dtype=torch.complex128)
>>> torch.fft.fft(t)
tensor([-2.5086+0.9281j, -0.0646+0.8343j, -1.7846-2.9435j, -0.1878-1.0965j], dtype=torch.complex128)
""")
import sys
import torch
from torch._C import _add_docstr, _linalg # type: ignore
Tensor = torch.Tensor
# Note: This not only adds doc strings for functions in the linalg namespace, but
# also connects the torch.linalg Python namespace to the torch._C._linalg builtins.
det = _add_docstr(
_linalg.linalg_det, r"""
linalg.det(input) -> Tensor
Alias of :func:`torch.det`.
""")
norm = _add_docstr(
_linalg.linalg_norm, r"""
linalg.norm(input, ord=None, dim=None, keepdim=False, *, out=None, dtype=None) -> Tensor
Returns the matrix norm or vector norm of a given tensor.
This function can calculate one of eight different types of matrix norms, or one
of an infinite number of vector norms, depending on both the number of reduction
dimensions and the value of the `ord` parameter.
Args:
input (Tensor): The input tensor. If dim is None, x must be 1-D or 2-D, unless :attr:`ord`
is None. If both :attr:`dim` and :attr:`ord` are None, the 2-norm of the input flattened to 1-D
will be returned.
class Tensor(torch._C._TensorBase):
def __deepcopy__(self, memo):
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__deepcopy__, (self, ), self,
memo)
if not self.is_leaf:
raise RuntimeError(
"Only Tensors created explicitly by the user "
"(graph leaves) support the deepcopy protocol at the moment")
if id(self) in memo:
return memo[id(self)]
with torch.no_grad():
# TODO: skipping storage copy is wrong for meta, as meta
# does accurate alias tracking; however, the code below
# doesn't work because of
# https://github.com/pytorch/pytorch/issues/47442
if self.is_sparse or self.device.type in ['xla', 'mlc', 'meta']:
new_tensor = self.clone()
else:
new_storage = self.storage().__deepcopy__(memo)
if self.is_quantized:
# quantizer_params can be different type based on torch attribute
quantizer_params: Union[Tuple[torch.qscheme, float, int],
Tuple[torch.qscheme, Tensor,
Tensor, int]]
if self.qscheme() == torch.per_tensor_affine:
quantizer_params = self.qscheme(), self.q_scale(
), self.q_zero_point()
elif self.qscheme() in (
torch.per_channel_affine,
torch.per_channel_affine_float_qparams):
quantizer_params = self.qscheme(), \
self.q_per_channel_scales(), \
self.q_per_channel_zero_points(), \
self.q_per_channel_axis()
else:
raise RuntimeError(
f"Unsupported qscheme {self.qscheme()} in deepcopy"
)
new_tensor = torch._utils._rebuild_qtensor(
new_storage, self.storage_offset(), self.size(),
self.stride(), quantizer_params, self.requires_grad,
self._backward_hooks)
else:
new_tensor = self.new()
new_tensor.set_(new_storage, self.storage_offset(),
self.size(), self.stride())
new_tensor.requires_grad = self.requires_grad
if self.grad is not None:
new_tensor.grad = self.grad.__deepcopy__(memo)
memo[id(self)] = new_tensor
return new_tensor
def __reduce_ex__(self, proto):
if type(self) is Tensor:
return self._reduce_ex_internal(proto)
relevant_args = (self, )
from torch.overrides import has_torch_function, handle_torch_function
if type(self) is not Tensor and has_torch_function(relevant_args):
return handle_torch_function(Tensor.__reduce_ex__, relevant_args,
self, proto)
func, args = self._reduce_ex_internal(proto)
return (_rebuild_from_type, (func, type(self), args, self.__dict__))
def _reduce_ex_internal(self, proto):
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__reduce_ex__, (self, ), self,
proto)
check_serializing_named_tensor(self)
# See Note [Don't serialize hooks]
torch.utils.hooks.warn_if_has_hooks(self)
backward_hooks: Dict[Any, Any] = OrderedDict()
# Note: Numpy array is chosen to be the rebuild component for XLA Tensor.
# We considered a few options:
# 1. CPU tensor can't be used here.
# Otherwise in torch.load CPU storage is reconstructed with randomly
# initialized data, moved onto XLA device, and then storage is updated
# to the serialized content. This works perfectly for CPU/CUDA but not XLA.
# XLA tensor is disconnected with storage so it doesn't get the update.
# 2. Python list is not a good fit due to performance reason.
# `tolist()` converts every single element in the tensor into python objects
# and serialize them one by one.
if self.device.type == 'xla':
arg_xla = (self.cpu().numpy(), self.dtype, str(self.device),
self.requires_grad)
return (torch._utils._rebuild_xla_tensor, arg_xla)
if self.device.type == 'mlc':
arg_mlc = (self.cpu().numpy(), self.dtype, str(self.device),
self.requires_grad)
return (torch._utils._rebuild_mlc_tensor, arg_mlc)
if self.is_quantized:
# quantizer_params can be different type based on torch attribute
quantizer_params: Union[Tuple[torch.qscheme, float, int],
Tuple[Any, Tensor, Tensor, int]]
if self.qscheme() == torch.per_tensor_affine:
quantizer_params = (torch.per_tensor_affine, self.q_scale(),
self.q_zero_point())
elif self.qscheme() in (torch.per_channel_affine,
torch.per_channel_affine_float_qparams):
# convert scales and zero points to tuple to avoid recursive calls
# when/if we get multi-axis quantized tensors in the future, the shape
# is recoverable from the main tensor shape
quantizer_params = (torch.per_channel_affine,
self.q_per_channel_scales(),
self.q_per_channel_zero_points(),
self.q_per_channel_axis())
else:
raise RuntimeError(
f"Serialization is not supported for tensors of type {self.qscheme()}"
)
args_qtensor = (self.storage(), self.storage_offset(),
tuple(self.size()), self.stride(),
quantizer_params, self.requires_grad,
backward_hooks)
return (torch._utils._rebuild_qtensor, args_qtensor)
elif self.is_sparse:
if self.layout == torch.sparse_coo:
args_sparse = (self.layout, (self._indices(), self._values(),
self.size()))
else:
raise NotImplementedError(
'sparse tensor __reduce_ex__ for layout `%s`' %
(self.layout))
return (torch._utils._rebuild_sparse_tensor, args_sparse)
else:
args = (self.storage(), self.storage_offset(), tuple(self.size()),
self.stride(), self.requires_grad, backward_hooks
) # previously was self._backward_hooks
return (torch._utils._rebuild_tensor_v2, args)
def __setstate__(self, state):
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__setstate__, (self, ), self,
state)
# Warning: this method is NOT called when you torch.load() a tensor;
# that is managed by _rebuild_tensor_v2
if not self.is_leaf:
raise RuntimeError(
'__setstate__ can be only called on leaf Tensors')
if len(state) == 4:
# legacy serialization of Tensor
self.set_(*state)
return
elif len(state) == 5:
# legacy serialization of Variable
self.data = state[0]
state = (state[3], state[4], state[2])
# The setting of _backward_hooks is expected to be a no-op.
# See Note [Don't serialize hooks]
self.requires_grad, _, self._backward_hooks = state
def __repr__(self):
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__repr__, (self, ), self)
# All strings are unicode in Python 3.
return torch._tensor_str._str(self)
def backward(self,
gradient=None,
retain_graph=None,
create_graph=False,
inputs=None):
r"""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 :ref:`Default gradient layouts<default-grad-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
:ref:`Stream semantics of backward passes<bwd-cuda-stream-semantics>`.
Args:
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.
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.backward, (self, ),
self,
gradient=gradient,
retain_graph=retain_graph,
create_graph=create_graph,
inputs=inputs)
torch.autograd.backward(self,
gradient,
retain_graph,
create_graph,
inputs=inputs)
def register_hook(self, hook):
r"""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 :attr:`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
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.register_hook, (self, ), self,
hook)
if not self.requires_grad:
raise RuntimeError("cannot register a hook on a tensor that "
"doesn't require gradient")
if self._backward_hooks is None:
self._backward_hooks = OrderedDict()
if self.grad_fn is not None:
self.grad_fn._register_hook_dict(self)
handle = hooks.RemovableHandle(self._backward_hooks)
self._backward_hooks[handle.id] = hook
return handle
def reinforce(self, reward):
def trim(str):
return '\n'.join([line.strip() for line in str.split('\n')])
raise RuntimeError(
trim(r"""reinforce() was removed.
Use torch.distributions instead.
See https://pytorch.org/docs/master/distributions.html
Instead of:
probs = policy_network(state)
action = probs.multinomial()
next_state, reward = env.step(action)
action.reinforce(reward)
action.backward()
Use:
probs = policy_network(state)
# NOTE: categorical is equivalent to what used to be called multinomial
m = torch.distributions.Categorical(probs)
action = m.sample()
next_state, reward = env.step(action)
loss = -m.log_prob(action) * reward
loss.backward()
"""))
detach = _C._add_docstr(
_C._TensorBase.detach, r"""
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_ = _C._add_docstr(
_C._TensorBase.detach_, r"""
Detaches the Tensor from the graph that created it, making it a leaf.
Views cannot be detached in-place.
""")
def retain_grad(self):
r"""Enables .grad attribute for non-leaf Tensors."""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.retain_grad, (self, ), self)
if not self.requires_grad:
raise RuntimeError(
"can't retain_grad on Tensor that has requires_grad=False")
if self.is_leaf: # no-op for leaves
return
if hasattr(self, 'retains_grad'):
return
weak_self = weakref.ref(self)
def retain_grad_hook(grad):
var = weak_self()
if var is None:
return
if var._grad is None:
if grad.is_sparse:
var._grad = grad.clone()
else:
var._grad = grad.clone(
memory_format=torch.contiguous_format)
else:
var._grad = var._grad + grad
self.register_hook(retain_grad_hook)
self.retains_grad = True
def is_shared(self):
r"""Checks if tensor is in shared memory.
This is always ``True`` for CUDA tensors.
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.is_shared, (self, ), self)
return self.storage().is_shared()
def share_memory_(self):
r"""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.
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.share_memory_, (self, ), self)
self.storage().share_memory_()
return self
def __reversed__(self):
r"""Reverses the tensor along dimension 0."""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__reversed__, (self, ), self)
if self.dim() == 0:
return self
else:
return self.flip(0)
def norm(self, p="fro", dim=None, keepdim=False, dtype=None):
r"""See :func:`torch.norm`"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.norm, (self, ),
self,
p=p,
dim=dim,
keepdim=keepdim,
dtype=dtype)
return torch.norm(self, p, dim, keepdim, dtype=dtype)
def lu(self, pivot=True, get_infos=False):
r"""See :func:`torch.lu`"""
# If get_infos is True, then we don't need to check for errors and vice versa
if has_torch_function_unary(self):
return handle_torch_function(Tensor.lu, (self, ),
self,
pivot=pivot,
get_infos=get_infos)
if not torch._jit_internal.is_scripting():
if self.requires_grad:
if not (self.size(-2) == self.size(-1) and
(self.dtype.is_floating_point) or self.is_complex):
raise ValueError(
'lu.backward works only with batches of squared full-rank matrices'
' of floating or complex types.')
from torch._autograd_functions import _LU
LU, pivots, infos = _LU.apply(self, pivot, get_infos)
if get_infos:
return LU, pivots, infos
else:
return LU, pivots
else:
if self.requires_grad:
raise RuntimeError(
'Script and require gradients is not supported at the moment.'
'If you just want to do the forward, use .detach()'
'on the input before calling the function.')
LU, pivots, infos = torch._lu_with_info(self,
pivot=pivot,
check_errors=(not get_infos))
if get_infos:
return LU, pivots, infos
else:
return LU, pivots
def stft(self,
n_fft: int,
hop_length: Optional[int] = None,
win_length: Optional[int] = None,
window: 'Optional[Tensor]' = None,
center: bool = True,
pad_mode: str = 'reflect',
normalized: bool = False,
onesided: Optional[bool] = None,
return_complex: Optional[bool] = None):
r"""See :func:`torch.stft`
.. warning::
This function changed signature at version 0.4.1. Calling with
the previous signature may cause error or return incorrect result.
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.stft, (self, ),
self,
n_fft,
hop_length=hop_length,
win_length=win_length,
window=window,
center=center,
pad_mode=pad_mode,
normalized=normalized,
onesided=onesided,
return_complex=return_complex)
return torch.stft(self,
n_fft,
hop_length,
win_length,
window,
center,
pad_mode,
normalized,
onesided,
return_complex=return_complex)
def istft(self,
n_fft: int,
hop_length: Optional[int] = None,
win_length: Optional[int] = None,
window: 'Optional[Tensor]' = None,
center: bool = True,
normalized: bool = False,
onesided: Optional[bool] = None,
length: Optional[int] = None,
return_complex: bool = False):
r"""See :func:`torch.istft`"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.istft, (self, ),
self,
n_fft,
hop_length=hop_length,
win_length=win_length,
window=window,
center=center,
normalized=normalized,
onesided=onesided,
length=length,
return_complex=return_complex)
return torch.istft(self,
n_fft,
hop_length,
win_length,
window,
center,
normalized,
onesided,
length,
return_complex=return_complex)
def resize(self, *sizes):
if has_torch_function_unary(self):
return handle_torch_function(Tensor.resize, (self, ), self, *sizes)
warnings.warn("non-inplace resize is deprecated")
from torch.autograd._functions import Resize
return Resize.apply(self, sizes)
def resize_as(self, tensor):
if has_torch_function_variadic(self, tensor):
return handle_torch_function(Tensor.resize_as, (self, tensor),
self, tensor)
warnings.warn("non-inplace resize_as is deprecated")
from torch.autograd._functions import Resize
return Resize.apply(self, tensor.size())
def split(self, split_size, dim=0):
r"""See :func:`torch.split`
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.split, (self, ),
self,
split_size,
dim=dim)
if isinstance(split_size, int):
return super(Tensor, self).split(split_size, dim)
elif isinstance(split_size, Tensor):
try:
split_size = int(split_size)
return super(Tensor, self).split(split_size, dim)
except ValueError:
return super(Tensor, self).split_with_sizes(split_size, dim)
else:
return super(Tensor, self).split_with_sizes(split_size, dim)
def unique(self,
sorted=True,
return_inverse=False,
return_counts=False,
dim=None):
r"""Returns the unique elements of the input tensor.
See :func:`torch.unique`
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.unique, (self, ),
self,
sorted=sorted,
return_inverse=return_inverse,
return_counts=return_counts,
dim=dim)
return torch.unique(self,
sorted=sorted,
return_inverse=return_inverse,
return_counts=return_counts,
dim=dim)
def unique_consecutive(self,
return_inverse=False,
return_counts=False,
dim=None):
r"""Eliminates all but the first element from every consecutive group of equivalent elements.
See :func:`torch.unique_consecutive`
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.unique_consecutive, (self, ),
self,
return_inverse=return_inverse,
return_counts=return_counts,
dim=dim)
return torch.unique_consecutive(self,
return_inverse=return_inverse,
return_counts=return_counts,
dim=dim)
def __rsub__(self, other):
if has_torch_function_variadic(self, other):
return handle_torch_function(Tensor.__rsub__, (self, other), self,
other)
return _C._VariableFunctions.rsub(self, other)
def __rdiv__(self, other):
if has_torch_function_variadic(self, other):
return handle_torch_function(Tensor.__rdiv__, (self, other), self,
other)
return self.reciprocal() * other
__rtruediv__ = __rdiv__
__itruediv__ = _C._TensorBase.__idiv__
__pow__ = _C._TensorBase.pow
def __format__(self, format_spec):
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__format__, (self, ), self,
format_spec)
if self.dim() == 0:
return self.item().__format__(format_spec)
return object.__format__(self, format_spec)
def __ipow__(self, other): # type: ignore[misc]
if has_torch_function_variadic(self, other):
return handle_torch_function(Tensor.__ipow__, (self, other), self,
other)
return NotImplemented
@_wrap_type_error_to_not_implemented
def __rpow__(self, other):
dtype = torch.result_type(other, self)
return torch.tensor(other, dtype=dtype, device=self.device)**self
@_wrap_type_error_to_not_implemented
def __floordiv__(self, other):
return torch.floor_divide(self, other)
@_wrap_type_error_to_not_implemented
def __rfloordiv__(self, other):
return torch.floor_divide(other, self)
__neg__ = _C._TensorBase.neg
__abs__ = _C._TensorBase.abs
def __len__(self):
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__len__, (self, ), self)
if self.dim() == 0:
raise TypeError("len() of a 0-d tensor")
return self.shape[0]
def __iter__(self):
# NB: we use 'imap' and not 'map' here, so that in Python 2 we get a
# generator and don't eagerly perform all the indexes. This could
# save us work, and also helps keep trace ordering deterministic
# (e.g., if you zip(*hiddens), the eager map will force all the
# indexes of hiddens[0] before hiddens[1], while the generator
# map will interleave them.)
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__iter__, (self, ), self)
if self.dim() == 0:
raise TypeError('iteration over a 0-d tensor')
if torch._C._get_tracing_state():
warnings.warn(
'Iterating over a tensor might cause the trace to be incorrect. '
'Passing a tensor of different shape won\'t change the number of '
'iterations executed (and might lead to errors or silently give '
'incorrect results).',
category=torch.jit.TracerWarning,
stacklevel=2)
return iter(self.unbind(0))
def __hash__(self):
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__hash__, (self, ), self)
return id(self)
def __dir__(self):
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__dir__, (self, ), self)
if self.is_quantized:
warnings.warn(
'Only a small subset of methods are supported for quantized tensors.'
)
tensor_methods = dir(self.__class__)
tensor_methods.remove('volatile') # deprecated
attrs = list(self.__dict__.keys())
keys = tensor_methods + attrs
# property only available dense, cuda tensors
if (not self.is_cuda) or self.is_sparse:
keys.remove("__cuda_array_interface__")
return sorted(keys)
# Numpy array interface, to support `numpy.asarray(tensor) -> ndarray`
__array_priority__ = 1000 # prefer Tensor ops over numpy ones
def __array__(self, dtype=None):
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__array__, (self, ),
self,
dtype=dtype)
if dtype is None:
return self.numpy()
else:
return self.numpy().astype(dtype, copy=False)
# Wrap Numpy array again in a suitable tensor when done, to support e.g.
# `numpy.sin(tensor) -> tensor` or `numpy.greater(tensor, 0) -> ByteTensor`
def __array_wrap__(self, array):
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__array_wrap__, (self, ),
self,
array=array)
if array.dtype == bool:
# Workaround, torch has no built-in bool tensor
array = array.astype('uint8')
return torch.from_numpy(array)
def __contains__(self, element):
r"""Check if `element` is present in tensor
Args:
element (Tensor or scalar): element to be checked
for presence in current tensor"
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.__contains__, (self, ), self,
element)
if isinstance(element, (torch.Tensor, Number)):
# type hint doesn't understand the __contains__ result array
return (element == self).any().item() # type: ignore[union-attr]
raise RuntimeError(
"Tensor.__contains__ only supports Tensor or scalar, but you passed in a %s."
% type(element))
@property
def __cuda_array_interface__(self):
"""Array view description for cuda tensors.
See:
https://numba.pydata.org/numba-doc/latest/cuda/cuda_array_interface.html
"""
if has_torch_function_unary(self):
# TODO mypy doesn't support @property, see: https://github.com/python/mypy/issues/6185
return handle_torch_function(
Tensor.__cuda_array_interface__.__get__, (self, ),
self) # type: ignore[attr-defined]
# raise AttributeError for unsupported tensors, so that
# hasattr(cpu_tensor, "__cuda_array_interface__") is False.
if not self.is_cuda:
raise AttributeError(
"Can't get __cuda_array_interface__ on non-CUDA tensor type: %s "
"If CUDA data is required use tensor.cuda() to copy tensor to device memory."
% self.type())
if self.is_sparse:
raise AttributeError(
"Can't get __cuda_array_interface__ on sparse type: %s "
"Use Tensor.to_dense() to convert to a dense tensor first." %
self.type())
# RuntimeError, matching tensor.__array__() behavior.
if self.requires_grad:
raise RuntimeError(
"Can't get __cuda_array_interface__ on Variable that requires grad. "
"If gradients aren't required, use var.detach() to get Variable that doesn't require grad."
)
# CUDA devices are little-endian and tensors are stored in native byte
# order. 1-byte entries are endian-agnostic.
typestr = {
torch.complex64: "<c8",
torch.complex128: "<c16",
torch.float16: "<f2",
torch.float32: "<f4",
torch.float64: "<f8",
torch.uint8: "|u1",
torch.int8: "|i1",
torch.int16: "<i2",
torch.int32: "<i4",
torch.int64: "<i8",
}[self.dtype]
itemsize = self.storage().element_size()
shape = tuple(self.shape)
if self.is_contiguous():
# __cuda_array_interface__ v2 requires the strides to be omitted
# (either not set or set to None) for C-contiguous arrays.
strides = None
else:
strides = tuple(s * itemsize for s in self.stride())
data_ptr = self.data_ptr() if self.numel() > 0 else 0
data = (data_ptr, False) # read-only is false
return dict(typestr=typestr,
shape=shape,
strides=strides,
data=data,
version=2)
def refine_names(self, *names):
r"""Refines the dimension names of :attr:`self` according to :attr:`names`.
Refining is a special case of renaming that "lifts" unnamed dimensions.
A ``None`` dim can be refined to have any name; a named dim can only be
refined to have the same name.
Because named tensors can coexist with unnamed tensors, refining names
gives a nice way to write named-tensor-aware code that works with both
named and unnamed tensors.
:attr:`names` may contain up to one Ellipsis (``...``).
The Ellipsis is expanded greedily; it is expanded in-place to fill
:attr:`names` to the same length as ``self.dim()`` using names from the
corresponding indices of ``self.names``.
Python 2 does not support Ellipsis but one may use a string literal
instead (``'...'``).
Args:
names (iterable of str): The desired names of the output tensor. May
contain up to one Ellipsis.
Examples::
>>> imgs = torch.randn(32, 3, 128, 128)
>>> named_imgs = imgs.refine_names('N', 'C', 'H', 'W')
>>> named_imgs.names
('N', 'C', 'H', 'W')
>>> tensor = torch.randn(2, 3, 5, 7, 11)
>>> tensor = tensor.refine_names('A', ..., 'B', 'C')
>>> tensor.names
('A', None, None, 'B', 'C')
.. warning::
The named tensor API is experimental and subject to change.
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.refine_names, (self, ), self,
*names)
names = resolve_ellipsis(names, self.names, 'refine_names')
return super(Tensor, self).refine_names(names)
def align_to(self, *names):
r"""Permutes the dimensions of the :attr:`self` tensor to match the order
specified in :attr:`names`, adding size-one dims for any new names.
All of the dims of :attr:`self` must be named in order to use this method.
The resulting tensor is a view on the original tensor.
All dimension names of :attr:`self` must be present in :attr:`names`.
:attr:`names` may contain additional names that are not in ``self.names``;
the output tensor has a size-one dimension for each of those new names.
:attr:`names` may contain up to one Ellipsis (``...``).
The Ellipsis is expanded to be equal to all dimension names of :attr:`self`
that are not mentioned in :attr:`names`, in the order that they appear
in :attr:`self`.
Python 2 does not support Ellipsis but one may use a string literal
instead (``'...'``).
Args:
names (iterable of str): The desired dimension ordering of the
output tensor. May contain up to one Ellipsis that is expanded
to all unmentioned dim names of :attr:`self`.
Examples::
>>> tensor = torch.randn(2, 2, 2, 2, 2, 2)
>>> named_tensor = tensor.refine_names('A', 'B', 'C', 'D', 'E', 'F')
# Move the F and E dims to the front while keeping the rest in order
>>> named_tensor.align_to('F', 'E', ...)
.. warning::
The named tensor API is experimental and subject to change.
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.align_to, (self, ), self,
*names)
ellipsis_idx = single_ellipsis_index(names, 'align_to')
if ellipsis_idx is None:
return super(Tensor, self).align_to(names)
return super(Tensor, self).align_to(
[name for name in names if not is_ellipsis(name)], ellipsis_idx)
def unflatten(self, dim, sizes):
r"""Expands the dimension :attr:`dim` of the :attr:`self` tensor over multiple dimensions
of sizes given by :attr:`sizes`.
* :attr:`sizes` is the new shape of the unflattened dimension and it can be a `Tuple[int]` as well
as `torch.Size` if :attr:`self` is a `Tensor`, or `namedshape` (Tuple[(name: str, size: int)])
if :attr:`self` is a `NamedTensor`. The total number of elements in sizes must match the number
of elements in the original dim being unflattened.
Args:
dim (Union[int, str]): Dimension to unflatten
sizes (Union[Tuple[int] or torch.Size, Tuple[Tuple[str, int]]]): New shape of the unflattened dimension
Examples:
>>> torch.randn(3, 4, 1).unflatten(1, (2, 2)).shape
torch.Size([3, 2, 2, 1])
>>> torch.randn(3, 4, 1).unflatten(1, (-1, 2)).shape # the size -1 is inferred from the size of dimension 1
torch.Size([3, 2, 2, 1])
>>> torch.randn(2, 4, names=('A', 'B')).unflatten('B', (('B1', 2), ('B2', 2)))
tensor([[[-1.1772, 0.0180],
[ 0.2412, 0.1431]],
[[-1.1819, -0.8899],
[ 1.5813, 0.2274]]], names=('A', 'B1', 'B2'))
>>> torch.randn(2, names=('A',)).unflatten('A', (('B1', -1), ('B2', 1)))
tensor([[-0.8591],
[ 0.3100]], names=('B1', 'B2'))
.. warning::
The named tensor API is experimental and subject to change.
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.unflatten, (self, ), self, dim,
sizes)
if not sizes:
raise RuntimeError("unflatten: sizes must be non-empty")
names = None
if isinstance(sizes,
OrderedDict) or (isinstance(sizes, (tuple, list))
and isinstance(sizes[0],
(tuple, list))):
names, sizes = unzip_namedshape(sizes)
return super(Tensor, self).unflatten(dim, sizes, names)
def rename_(self, *names, **rename_map):
"""In-place version of :meth:`~Tensor.rename`."""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.rename_, (self, ), self,
*names, **rename_map)
# Note [rename_ / rename API]
# The Python API for these is different from the C++ API. In Python:
# 1) tensor.rename(*names) takes a vararglist of names
# 2) tensor.rename(**rename_map) takes a map of names to rename.
# C++ is static, making it difficult to implement similar behavior.
return update_names(self, names, rename_map, inplace=True)
def rename(self, *names, **rename_map):
"""Renames dimension names of :attr:`self`.
There are two main usages:
``self.rename(**rename_map)`` returns a view on tensor that has dims
renamed as specified in the mapping :attr:`rename_map`.
``self.rename(*names)`` returns a view on tensor, renaming all
dimensions positionally using :attr:`names`.
Use ``self.rename(None)`` to drop names on a tensor.
One cannot specify both positional args :attr:`names` and keyword args
:attr:`rename_map`.
Examples::
>>> imgs = torch.rand(2, 3, 5, 7, names=('N', 'C', 'H', 'W'))
>>> renamed_imgs = imgs.rename(N='batch', C='channels')
>>> renamed_imgs.names
('batch', 'channels', 'H', 'W')
>>> renamed_imgs = imgs.rename(None)
>>> renamed_imgs.names
(None,)
>>> renamed_imgs = imgs.rename('batch', 'channel', 'height', 'width')
>>> renamed_imgs.names
('batch', 'channel', 'height', 'width')
.. warning::
The named tensor API is experimental and subject to change.
"""
if has_torch_function_unary(self):
return handle_torch_function(Tensor.rename, (self, ), self, *names,
**rename_map)
# See Note [rename_ / rename API]
return update_names(self, names, rename_map, inplace=False)
def _update_names(self, names, inplace):
if has_torch_function_unary(self):
return handle_torch_function(Tensor._update_names, (self, ), self,
names, inplace)
# See Note [rename_ / rename API]
if inplace:
return super(Tensor, self).rename_(names)
else:
return super(Tensor, self).rename(names)
@property
def grad(self):
"""
This attribute is ``None`` by default and becomes a Tensor the first time a call to
:func:`backward` computes gradients for ``self``.
The attribute will then contain the gradients computed and future calls to
:func:`backward` will accumulate (add) gradients into it.
"""
if has_torch_function_unary(self):
# TODO mypy doesn't support @property, see: https://github.com/python/mypy/issues/6185
return handle_torch_function(Tensor.grad.__get__, (self, ),
self) # type: ignore[attr-defined]
if self.requires_grad and not hasattr(
self,
"retains_grad") and not self.is_leaf and self._grad is None:
warnings.warn(
"The .grad attribute of a Tensor that is not a leaf Tensor is being accessed. Its .grad "
"attribute won't be populated during autograd.backward(). If you indeed want the gradient "
"for a non-leaf Tensor, use .retain_grad() on the non-leaf Tensor. If you access the "
"non-leaf Tensor by mistake, make sure you access the leaf Tensor instead. See "
"github.com/pytorch/pytorch/pull/30531 for more information.",
stacklevel=2)
return self._grad
@grad.setter
def grad(self, new_grad):
if has_torch_function_unary(self):
# TODO mypy doesn't support @property, see: https://github.com/python/mypy/issues/6185
return handle_torch_function(
Tensor.grad.__set__, (self, ), self,
new_grad) # type: ignore[attr-defined]
self._grad = new_grad
@grad.deleter
def grad(self):
if has_torch_function_unary(self):
# TODO mypy doesn't support @property, see: https://github.com/python/mypy/issues/6185
return handle_torch_function(Tensor.grad.__delete__, (self, ),
self) # type: ignore[attr-defined]
del self._grad
@classmethod
def __torch_function__(cls, func, types, args=(), kwargs=None):
"""
This __torch_function__ implementation wraps subclasses such that
methods called on subclasses return a subclass instance instead of
a ``torch.Tensor`` instance.
One corollary to this is that you need coverage for torch.Tensor
methods if implementing __torch_function__ for subclasses.
We recommend always calling ``super().__torch_function__`` as the base
case when doing the above.
While not mandatory, we recommend making `__torch_function__` a classmethod.
"""
if kwargs is None:
kwargs = {}
if not all(issubclass(cls, t) for t in types):
return NotImplemented
with _C.DisableTorchFunction():
ret = func(*args, **kwargs)
return _convert(ret, cls)
__module__ = 'torch'
cholesky = _add_docstr(
_linalg.linalg_cholesky, r"""
linalg.cholesky(input, *, out=None) -> Tensor
Computes the Cholesky decomposition of a Hermitian (or symmetric for real-valued matrices)
positive-definite matrix or the Cholesky decompositions for a batch of such matrices.
Each decomposition has the form:
.. math::
\text{input} = LL^H
where :math:`L` is a lower-triangular matrix and :math:`L^H` is the conjugate transpose of :math:`L`,
which is just a transpose for the case of real-valued input matrices.
In code it translates to ``input = L @ L.t()` if :attr:`input` is real-valued and
``input = L @ L.conj().t()`` if :attr:`input` is complex-valued.
The batch of :math:`L` matrices is returned.
Supports real-valued and complex-valued inputs.
.. note:: If :attr:`input` is not a Hermitian positive-definite matrix, or if it's a batch of matrices
and one or more of them is not a Hermitian positive-definite matrix, then a RuntimeError will be thrown.
If :attr:`input` is a batch of matrices, then the error message will include the batch index
of the first matrix that is not Hermitian positive-definite.
.. warning:: This function always checks whether :attr:`input` is a Hermitian positive-definite matrix
using `info` argument to LAPACK/MAGMA call. For CUDA this causes cross-device memory synchronization.
Args:
input (Tensor): the input tensor of size :math:`(*, n, n)` consisting of Hermitian positive-definite
:math:`n \times n` matrices, where `*` is zero or more batch dimensions.
Keyword args:
out (Tensor, optional): The output tensor. Ignored if ``None``. Default: ``None``
Examples::
>>> a = torch.randn(2, 2, dtype=torch.complex128)
>>> a = torch.mm(a, a.t().conj()) # creates a Hermitian positive-definite matrix
>>> l = torch.linalg.cholesky(a)
>>> a
tensor([[2.5266+0.0000j, 1.9586-2.0626j],
[1.9586+2.0626j, 9.4160+0.0000j]], dtype=torch.complex128)
>>> l
tensor([[1.5895+0.0000j, 0.0000+0.0000j],
[1.2322+1.2976j, 2.4928+0.0000j]], dtype=torch.complex128)
>>> torch.mm(l, l.t().conj())
tensor([[2.5266+0.0000j, 1.9586-2.0626j],
[1.9586+2.0626j, 9.4160+0.0000j]], dtype=torch.complex128)
>>> a = torch.randn(3, 2, 2, dtype=torch.float64)
>>> a = torch.matmul(a, a.transpose(-2, -1)) # creates a symmetric positive-definite matrix
>>> l = torch.linalg.cholesky(a)
>>> a
tensor([[[ 1.1629, 2.0237],
[ 2.0237, 6.6593]],
[[ 0.4187, 0.1830],
[ 0.1830, 0.1018]],
[[ 1.9348, -2.5744],
[-2.5744, 4.6386]]], dtype=torch.float64)
>>> l
tensor([[[ 1.0784, 0.0000],
[ 1.8766, 1.7713]],
[[ 0.6471, 0.0000],
[ 0.2829, 0.1477]],
[[ 1.3910, 0.0000],
[-1.8509, 1.1014]]], dtype=torch.float64)
>>> torch.allclose(torch.matmul(l, l.transpose(-2, -1)), a)
True
""")
Tensor = torch.Tensor
entr = _add_docstr(_special.special_entr,
r"""
entr(input, *, out=None) -> Tensor
Computes the entropy on :attr:`input` (as defined below), elementwise.
.. math::
\begin{align}
\text{entr(x)} = \begin{cases}
-x * \ln(x) & x > 0 \\
0 & x = 0.0 \\
-\infty & x < 0
\end{cases}
\end{align}
""" + """
Args:
input (Tensor): the input tensor.
Keyword args:
out (Tensor, optional): the output tensor.
Example::
>>> a = torch.arange(-0.5, 1, 0.5)
>>> a
tensor([-0.5000, 0.0000, 0.5000])
>>> torch.special.entr(a)
tensor([ -inf, 0.0000, 0.3466])
""")
psi = _add_docstr(_special.special_psi,
fft = _add_docstr(_fft.fft_fft, r"""
fft(input, n=None, dim=-1, norm=None, *, out=None) -> Tensor
Computes the one dimensional discrete Fourier transform of :attr:`input`.
Note:
The Fourier domain representation of any real signal satisfies the
Hermitian property: `X[i] = conj(X[-i])`. This function always returns both
the positive and negative frequency terms even though, for real inputs, the
negative frequencies are redundant. :func:`~torch.fft.rfft` returns the
more compact one-sided representation where only the positive frequencies
are returned.
Args:
input (Tensor): the input tensor
n (int, optional): Signal length. If given, the input will either be zero-padded
or trimmed to this length before computing the FFT.
dim (int, optional): The dimension along which to take the one dimensional FFT.
norm (str, optional): Normalization mode. For the forward transform
(:func:`~torch.fft.fft`), these correspond to:
* ``"forward"`` - normalize by ``1/n``
* ``"backward"`` - no normalization
* ``"ortho"`` - normalize by ``1/sqrt(n)`` (making the FFT orthonormal)
Calling the backward transform (:func:`~torch.fft.ifft`) with the same
normalization mode will apply an overall normalization of ``1/n`` between
the two transforms. This is required to make :func:`~torch.fft.ifft`
the exact inverse.
Default is ``"backward"`` (no normalization).
Keyword args:
{out}
Example:
>>> t = torch.arange(4)
>>> t
tensor([0, 1, 2, 3])
>>> torch.fft.fft(t)
tensor([ 6.+0.j, -2.+2.j, -2.+0.j, -2.-2.j])
>>> t = tensor([0.+1.j, 2.+3.j, 4.+5.j, 6.+7.j])
>>> torch.fft.fft(t)
tensor([12.+16.j, -8.+0.j, -4.-4.j, 0.-8.j])
""".format(**common_args))
import sys
import torch
from torch._C import _add_docstr, _special # type: ignore
Tensor = torch.Tensor
gammaln = _add_docstr(
_special.special_gammaln, r"""
gammaln(input, *, out=None) -> Tensor
Computes the natural logarithm of the absolute value of the gamma function on :attr:`input`.
.. math::
\text{out}_{i} = \ln \Gamma(|\text{input}_{i}|)
""" + """
Args:
input (Tensor): the input tensor.
Keyword args:
out (Tensor, optional): the output tensor.
Example::
>>> a = torch.arange(0.5, 2, 0.5)
>>> torch.special.gammaln(a)
tensor([ 0.5724, 0.0000, -0.1208])
""")
cholesky = _add_docstr(_linalg.linalg_cholesky, r"""
linalg.cholesky(input, *, out=None) -> Tensor
Computes the Cholesky decomposition of a Hermitian (or symmetric for real-valued matrices)
positive-definite matrix or the Cholesky decompositions for a batch of such matrices.
Each decomposition has the form:
.. math::
\text{input} = LL^H
where :math:`L` is a lower-triangular matrix and :math:`L^H` is the conjugate transpose of :math:`L`,
which is just a transpose for the case of real-valued input matrices.
In code it translates to ``input = L @ L.t()` if :attr:`input` is real-valued and
``input = L @ L.conj().t()`` if :attr:`input` is complex-valued.
The batch of :math:`L` matrices is returned.
Supports real-valued and complex-valued inputs.
.. note:: If :attr:`input` is not a Hermitian positive-definite matrix, or if it's a batch of matrices
and one or more of them is not a Hermitian positive-definite matrix, then a RuntimeError will be thrown.
If :attr:`input` is a batch of matrices, then the error message will include the batch index
of the first matrix that is not Hermitian positive-definite.
.. warning:: This function always checks whether :attr:`input` is a Hermitian positive-definite matrix
using `info` argument to LAPACK/MAGMA call. For CUDA this causes cross-device memory synchronization.
Args:
input (Tensor): the input tensor of size :math:`(*, n, n)` consisting of Hermitian positive-definite
:math:`n \times n` matrices, where `*` is zero or more batch dimensions.
Keyword args:
out (Tensor, optional): The output tensor. Ignored if ``None``. Default: ``None``
Examples::
>>> a = torch.randn(2, 2, dtype=torch.complex128)
>>> a = torch.mm(a, a.t().conj()) # creates a Hermitian positive-definite matrix
>>> l = torch.linalg.cholesky(a)
>>> a
tensor([[2.5266+0.0000j, 1.9586-2.0626j],
[1.9586+2.0626j, 9.4160+0.0000j]], dtype=torch.complex128)
>>> l
tensor([[1.5895+0.0000j, 0.0000+0.0000j],
[1.2322+1.2976j, 2.4928+0.0000j]], dtype=torch.complex128)
>>> torch.mm(l, l.t().conj())
tensor([[2.5266+0.0000j, 1.9586-2.0626j],
[1.9586+2.0626j, 9.4160+0.0000j]], dtype=torch.complex128)
>>> a = torch.randn(3, 2, 2, dtype=torch.float64)
>>> a = torch.matmul(a, a.transpose(-2, -1)) # creates a symmetric positive-definite matrix
>>> l = torch.linalg.cholesky(a)
>>> a
tensor([[[ 1.1629, 2.0237],
[ 2.0237, 6.6593]],
[[ 0.4187, 0.1830],
[ 0.1830, 0.1018]],
[[ 1.9348, -2.5744],
[-2.5744, 4.6386]]], dtype=torch.float64)
>>> l
tensor([[[ 1.0784, 0.0000],
[ 1.8766, 1.7713]],
[[ 0.6471, 0.0000],
[ 0.2829, 0.1477]],
[[ 1.3910, 0.0000],
[-1.8509, 1.1014]]], dtype=torch.float64)
>>> torch.allclose(torch.matmul(l, l.transpose(-2, -1)), a)
True
""")
fft = _add_docstr(
_fft.fft_fft, r"""
fft(input, n=None, dim=-1, norm=None) -> Tensor
Computes the one dimensional discrete Fourier transform of :attr:`input`.
Note:
The Fourier domain representation of any real signal satisfies the
Hermitian property: `X[i] = conj(X[-i])`. This function always returns both
the positive and negative frequency terms even though, for real inputs, the
negative frequencies are redundant. :func:`~torch.fft.rfft` returns the
more compact one-sided representation where only the positive frequencies
are returned.
Args:
input (Tensor): the input tensor
n (int, optional): Signal length. If given, the input will either be zero-padded
or trimmed to this length before computing the FFT.
dim (int, optional): The dimension along which to take the one dimensional FFT.
norm (str, optional): Normalization mode. For the forward transform
(:func:`~torch.fft.fft`), these correspond to:
* ``"forward"`` - normalize by ``1/n``
* ``"backward"`` - no normalization
* ``"ortho"`` - normalize by ``1/sqrt(n)`` (making the FFT orthonormal)
Calling the backward transform (:func:`~torch.fft.ifft`) with the same
normalization mode will apply an overall normalization of ``1/n`` between
the two transforms. This is required to make :func:`~torch.fft.ifft`
the exact inverse.
Default is ``"backward"`` (no normalization).
Example:
>>> import torch.fft
>>> t = torch.arange(4)
>>> t
tensor([0, 1, 2, 3])
>>> torch.fft.fft(t)
tensor([ 6.+0.j, -2.+2.j, -2.+0.j, -2.-2.j])
>>> t = tensor([0.+1.j, 2.+3.j, 4.+5.j, 6.+7.j])
>>> torch.fft.fft(t)
tensor([12.+16.j, -8.+0.j, -4.-4.j, 0.-8.j])
""")
__all__ = [
'addmm',
'mm',
'sum',
'softmax',
'log_softmax',
]
addmm = _add_docstr(
_sparse._sparse_addmm, r"""
sparse.addmm(mat, mat1, mat2, *, beta=1., alpha=1.) -> Tensor
This function does exact same thing as :func:`torch.addmm` in the forward,
except that it supports backward for sparse matrix :attr:`mat1`. :attr:`mat1`
need to have `sparse_dim = 2`. Note that the gradients of :attr:`mat1` is a
coalesced sparse tensor.
Args:
mat (Tensor): a dense matrix to be added
mat1 (Tensor): a sparse matrix to be multiplied
mat2 (Tensor): a dense matrix to be multiplied
beta (Number, optional): multiplier for :attr:`mat` (:math:`\beta`)
alpha (Number, optional): multiplier for :math:`mat1 @ mat2` (:math:`\alpha`)
""")
def mm(mat1: Tensor, mat2: Tensor) -> Tensor:
r"""
Performs a matrix multiplication of the sparse matrix :attr:`mat1`
and the (sparse or strided) matrix :attr:`mat2`. Similar to :func:`torch.mm`, If :attr:`mat1` is a
:math:`(n \times m)` tensor, :attr:`mat2` is a :math:`(m \times p)` tensor, out will be a
:math:`(n \times p)` tensor. :attr:`mat1` need to have `sparse_dim = 2`.
import sys
import torch
from torch._C import _add_docstr, _linalg # type: ignore
Tensor = torch.Tensor
# Note: This not only adds doc strings for functions in the linalg namespace, but
# also connects the torch.linalg Python namespace to the torch._C._linalg builtins.
det = _add_docstr(
_linalg.linalg_det, r"""
linalg.det(input) -> Tensor
Alias of :func:`torch.det`.
""")
norm = _add_docstr(
_linalg.linalg_norm, r"""
linalg.norm(input, ord=None, dim=None, keepdim=False, *, out=None, dtype=None) -> Tensor
Returns the matrix norm or vector norm of a given tensor.
This function can calculate one of eight different types of matrix norms, or one
of an infinite number of vector norms, depending on both the number of reduction
dimensions and the value of the `ord` parameter.
Args:
input (Tensor): The input tensor. If dim is None, x must be 1-D or 2-D, unless :attr:`ord`
is None. If both :attr:`dim` and :attr:`ord` are None, the 2-norm of the input flattened to 1-D
will be returned.
import sys
import torch
from torch._C import _add_docstr, _linalg # type: ignore
Tensor = torch.Tensor
# Note: This not only adds doc strings for the linear algebra ops, but
# also connects the torch.linalg Python namespace to the torch._C._linalg builtins.
outer = _add_docstr(
_linalg.linalg_outer, r"""
linalg.outer(input, vec2, *, out=None) -> Tensor
Alias of :func:`torch.ger`.
""")
from torch._C import _add_docstr, _special # type: ignore
from torch._torch_docs import common_args # type: ignore
Tensor = torch.Tensor
gammaln = _add_docstr(_special.special_gammaln,
r"""
gammaln(input, *, out=None) -> Tensor
Computes the natural logarithm of the absolute value of the gamma function on :attr:`input`.
.. math::
\text{out}_{i} = \ln \Gamma(|\text{input}_{i}|)
""" + """
Args:
{input}
Keyword args:
{out}
Example::
>>> a = torch.arange(0.5, 2, 0.5)
>>> torch.special.gammaln(a)
tensor([ 0.5724, 0.0000, -0.1208])
""".format(**common_args))
erf = _add_docstr(_special.special_erf,
r"""
erf(input, *, out=None) -> Tensor
class Variable(_C._VariableBase):
"""Wraps a tensor and records the operations applied to it.
Variable is a thin wrapper around a Tensor object, that also holds
the gradient w.r.t. to it, and a reference to a function that created it.
This reference allows retracing the whole chain of operations that
created the data. If the Variable has been created by the user, its grad_fn
will be ``None`` and we call such objects *leaf* Variables.
Since autograd only supports scalar valued function differentiation, grad
size always matches the data size. Also, grad is normally only allocated
for leaf variables, and will be always zero otherwise.
Attributes:
data: Wrapped tensor of any type.
grad: Variable holding the gradient of type and location matching
the ``.data``. This attribute is lazily allocated and can't
be reassigned.
requires_grad: Boolean indicating whether the Variable has been
created by a subgraph containing any Variable, that requires it.
See :ref:`excluding-subgraphs` for more details.
Can be changed only on leaf Variables.
is_leaf: Boolean indicating if the Variable is a graph leaf (i.e
if it was created by the user).
grad_fn: Gradient function graph trace.
Parameters:
data (any tensor class): Tensor to wrap.
requires_grad (bool): Value of the requires_grad flag. **Keyword only.**
"""
def __deepcopy__(self, memo):
if not self.is_leaf:
raise RuntimeError(
"Only Variables created explicitly by the user "
"(graph leaves) support the deepcopy protocol at the moment")
result = type(self)(self.data.clone())
result.requires_grad = self.requires_grad
memo[id(self)] = result
return result
def __reduce_ex__(self, proto):
state = (self.requires_grad, False, self._backward_hooks)
if proto > 1:
return type(self), (self.data, ), state
if sys.version_info[0] == 2:
from copy_reg import __newobj__
else:
from copyreg import __newobj__
return __newobj__, (type(self), self.data), state
def __setstate__(self, state):
if len(state) == 5:
# legacy serialization of Variable
self.data = state[0]
state = (state[3], state[4], state[2])
if not self.is_leaf:
raise RuntimeError(
'__setstate__ can be only called on leaf variables')
self.requires_grad, _, self._backward_hooks = state
def __repr__(self):
if self.is_sparse:
data_str = ' \n{} with indices:\n{}and values:\n{}'.format(
torch.typename(self.data),
self._indices().data,
self._values().data)
else:
data_str = torch._tensor_str._str(self.data, False)
strt = 'Variable containing:' + data_str
# let's make our own Variable-specific footer
size_str = '(' + ','.join(
str(size)
for size in self.size()) + (',)' if len(self.size()) == 1 else ')')
device_str = '' if not self.is_cuda else \
' (GPU {})'.format(self.get_device())
strt += '[{} of size {}{}]\n'.format(torch.typename(self.data),
size_str, device_str)
# All strings are unicode in Python 3, while we have to encode unicode
# strings in Python2. If we can't, let python decide the best
# characters to replace unicode characters with.
if sys.version_info > (3, ):
return strt
else:
if hasattr(sys.stdout, 'encoding'):
return strt.encode(sys.stdout.encoding or 'UTF-8', 'replace')
else:
return strt.encode('UTF-8', 'replace')
def backward(self, gradient=None, retain_graph=None, create_graph=False):
"""Computes the gradient of current variable w.r.t. graph leaves.
The graph is differentiated using the chain rule. If the variable 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 them before calling it.
Arguments:
gradient (Tensor, Variable or None): Gradient w.r.t. the
variable. If it is a tensor, it will be automatically converted
to a Variable that does not require grad unless ``create_graph`` is True.
None values can be specified for scalar Variables 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``.
"""
torch.autograd.backward(self, gradient, retain_graph, create_graph)
def register_hook(self, hook):
"""Registers a backward hook.
The hook will be called every time a gradient with respect to the
variable is computed. The hook should have the following signature::
hook(grad) -> Variable or None
The hook should not modify its argument, but it can optionally return
a new gradient which will be used in place of :attr:`grad`.
This function returns a handle with a method ``handle.remove()``
that removes the hook from the module.
Example:
>>> v = Variable(torch.Tensor([0, 0, 0]), requires_grad=True)
>>> h = v.register_hook(lambda grad: grad * 2) # double the gradient
>>> v.backward(torch.Tensor([1, 1, 1]))
>>> v.grad.data
2
2
2
[torch.FloatTensor of size 3]
>>> h.remove() # removes the hook
"""
if not self.requires_grad:
raise RuntimeError("cannot register a hook on a variable that "
"doesn't require gradient")
if self._backward_hooks is None:
self._backward_hooks = OrderedDict()
if self.grad_fn is not None:
self.grad_fn._register_hook_dict(self)
handle = hooks.RemovableHandle(self._backward_hooks)
self._backward_hooks[handle.id] = hook
return handle
def reinforce(self, reward):
def trim(str):
return '\n'.join([line.strip() for line in str.split('\n')])
raise RuntimeError(
trim(r"""reinforce() was removed.
Use torch.distributions instead.
See http://pytorch.org/docs/master/distributions.html
Instead of:
probs = policy_network(state)
action = probs.multinomial()
next_state, reward = env.step(action)
action.reinforce(reward)
action.backward()
Use:
probs = policy_network(state)
# NOTE: categorical is equivalent to what used to be called multinomial
m = torch.distributions.Categorical(probs)
action = m.sample()
next_state, reward = env.step(action)
loss = -m.log_prob(action) * reward
loss.backward()
"""))
detach = _add_docstr(
_C._VariableBase.detach, r"""
Returns a new Variable, detached from the current graph.
The result will never require gradient.
.. note::
Returned Variable uses the same data tensor, as the original one, and
in-place modifications on either of them will be seen, and may trigger
errors in correctness checks.
""")
detach_ = _add_docstr(
_C._VariableBase.detach_, r"""
Detaches the Variable from the graph that created it, making it a leaf.
Views cannot be detached in-place.
""")
def retain_grad(self):
"""Enables .grad attribute for non-leaf Variables."""
if self.grad_fn is None: # no-op for leaves
return
if not self.requires_grad:
raise RuntimeError(
"can't retain_grad on Variable that has requires_grad=False")
if hasattr(self, 'retains_grad'):
return
weak_self = weakref.ref(self)
def retain_grad_hook(grad):
var = weak_self()
if var is None:
return
if var._grad is None:
var._grad = grad.clone()
else:
var._grad = var._grad + grad
self.register_hook(retain_grad_hook)
self.retains_grad = True
def type_as(self, other):
if torch.is_tensor(other):
other = Variable(other)
return super(Variable, self).type_as(other)
def is_pinned(self):
r"""Returns true if this tensor resides in pinned memory"""
storage = self.storage()
return storage.is_pinned() if storage else False
def is_shared(self):
r"""Checks if tensor is in shared memory.
This is always ``True`` for CUDA tensors.
"""
return self.storage().is_shared()
def share_memory_(self):
r"""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.
"""
self.storage().share_memory_()
def view_as(self, tensor):
return self.view(tensor.size())
def repeat(self, *repeats):
if len(repeats) == 1 and isinstance(repeats[0], torch.Size):
repeats = repeats[0]
else:
repeats = torch.Size(repeats)
return Repeat.apply(self, repeats)
def btrifact(self, info=None, pivot=True):
if info is not None:
warnings.warn(
"info option in btrifact is deprecated and will be removed in v0.4, "
"consider using btrifact_with_info instead")
factorization, pivots, _info = super(
Variable, self).btrifact_with_info(pivot=pivot)
if not isinstance(info, Variable) or info.type() != _info.type():
raise ValueError(
'btrifact expects info to be a Variable of IntTenor')
info.data.copy_(_info.data)
return factorization, pivots
else:
return super(Variable, self).btrifact(pivot=pivot)
def resize(self, *sizes):
return Resize.apply(self, sizes)
def resize_as(self, variable):
return Resize.apply(self, variable.size())
def index_add(self, dim, index, tensor):
return self.clone().index_add_(dim, index, tensor)
def index_copy(self, dim, index, tensor):
return self.clone().index_copy_(dim, index, tensor)
def index_fill(self, dim, index, value):
return self.clone().index_fill_(dim, index, value)
def scatter(self, dim, index, source):
return self.clone().scatter_(dim, index, source)
def scatter_add(self, dim, index, source):
return self.clone().scatter_add_(dim, index, source)
def masked_copy(self, mask, variable):
warnings.warn(
"masked_copy is deprecated and renamed to masked_scatter, and will be removed in v0.3"
)
return self.masked_scatter(mask, variable)
def masked_copy_(self, mask, variable):
warnings.warn(
"masked_copy_ is deprecated and renamed to masked_scatter_, and will be removed in v0.3"
)
return self.masked_scatter_(mask, variable)
def masked_scatter(self, mask, variable):
return self.clone().masked_scatter_(mask, variable)
def masked_fill(self, mask, value):
return self.clone().masked_fill_(mask, value)
def expand_as(self, tensor):
return self.expand(tensor.size())
def __rsub__(self, other):
return -self + other
def __rdiv__(self, other):
return self.reciprocal() * other
__rtruediv__ = __rdiv__
__itruediv__ = _C._VariableBase.__div__
__pow__ = _C._VariableBase.pow
def __ipow__(self, other):
raise NotImplementedError("in-place pow not implemented")
def __rpow__(self, other):
return PowConstant.apply(other, self)
__neg__ = _C._VariableBase.neg
__eq__ = _C._VariableBase.eq
__ne__ = _C._VariableBase.ne
__lt__ = _C._VariableBase.lt
__le__ = _C._VariableBase.le
__gt__ = _C._VariableBase.gt
__ge__ = _C._VariableBase.ge
def __len__(self):
return len(self.data)
def __iter__(self):
# NB: we use 'imap' and not 'map' here, so that in Python 2 we get a
# generator and don't eagerly perform all the indexes. This could
# save us work, and also helps keep trace ordering deterministic
# (e.g., if you zip(*hiddens), the eager map will force all the
# indexes of hiddens[0] before hiddens[1], while the generator
# map will interleave them.)
return iter(imap(lambda i: self[i], range(self.size(0))))
def __hash__(self):
return id(self)
def __dir__(self):
variable_methods = dir(self.__class__)
variable_methods.remove('volatile') # deprecated
attrs = list(self.__dict__.keys())
keys = variable_methods + attrs
return sorted(keys)
# Numpy array interface, to support `numpy.asarray(tensor) -> ndarray`
def __array__(self, dtype=None):
if dtype is None:
return self.cpu().numpy()
else:
return self.cpu().numpy().astype(dtype, copy=False)
# Wrap Numpy array again in a suitable tensor when done, to support e.g.
# `numpy.sin(tensor) -> tensor` or `numpy.greater(tensor, 0) -> ByteTensor`
def __array_wrap__(self, array):
if array.dtype == bool:
# Workaround, torch has no built-in bool tensor
array = array.astype('uint8')
return Variable.from_numpy(array)
class _torch(object):
pass
r"""Functional interface""" from __future__ import division import warnings import math import torch from torch._C import _infer_size, _add_docstr from . import _reduction as _Reduction from .modules import utils from .modules.utils import _single, _pair, _triple, _list_with_default from . import grad # noqa: F401 from . import _VF from .._jit_internal import boolean_dispatch, List conv2d = _add_docstr(torch.conv2d, r"")
class Tensor(torch._C._TensorBase):
def __deepcopy__(self, memo):
if not self.is_leaf:
raise RuntimeError(
"Only Tensors created explicitly by the user "
"(graph leaves) support the deepcopy protocol at the moment")
if id(self) in memo:
return memo[id(self)]
with torch.no_grad():
if self.is_sparse or self.device.type == 'xla':
new_tensor = self.clone()
else:
new_storage = self.storage().__deepcopy__(memo)
new_tensor = self.new()
new_tensor.set_(new_storage, self.storage_offset(),
self.size(), self.stride())
memo[id(self)] = new_tensor
new_tensor.requires_grad = self.requires_grad
return new_tensor
def __reduce_ex__(self, proto):
check_serializing_named_tensor(self)
# See Note [Don't serialize hooks]
torch.utils.hooks.warn_if_has_hooks(self)
# Note: Numpy array is chosen to be the rebuild component for XLA Tensor.
# We considered a few options:
# 1. CPU tensor can't be used here.
# Otherwise in torch.load CPU storage is reconstructed with randomly
# initialized data, moved onto XLA device, and then storage is updated
# to the serialized content. This works perfectly for CPU/CUDA but not XLA.
# XLA tensor is disconnected with storage so it doesn't get the update.
# 2. Python list is not a good fit due to performance reason.
# `tolist()` converts every single element in the tensor into python objects
# and serialize them one by one.
if self.device.type == 'xla':
args = (self.cpu().numpy(), self.dtype, str(self.device),
self.requires_grad)
return (torch._utils._rebuild_xla_tensor, args)
if self.is_quantized:
if self.qscheme() == torch.per_tensor_affine:
quantizer_params = (torch.per_tensor_affine, self.q_scale(),
self.q_zero_point())
elif self.qscheme() == torch.per_channel_affine:
# convert scales and zero points to tuple to avoid recursive calls
# when/if we get multi-axis quantized tensors in the future, the shape
# is recoverable from the main tensor shape
quantizer_params = (torch.per_channel_affine, [
e.item() for e in self.q_per_channel_scales().reshape(-1)
], [
e.item()
for e in self.q_per_channel_zero_points().reshape(-1)
], self.q_per_channel_axis())
else:
raise RuntimeError(
"Serialization is not supported for tensors of type {}".
format(self.qscheme()))
args = (self.storage(), self.storage_offset(), tuple(self.size()),
self.stride(), quantizer_params, self.requires_grad,
OrderedDict())
return (torch._utils._rebuild_qtensor, args)
else:
args = (self.storage(), self.storage_offset(), tuple(self.size()),
self.stride(), self.requires_grad, OrderedDict()
) # previously was self._backward_hooks
return (torch._utils._rebuild_tensor_v2, args)
def __setstate__(self, state):
# Warning: this method is NOT called when you torch.load() a tensor;
# that is managed by _rebuild_tensor_v2
if not self.is_leaf:
raise RuntimeError(
'__setstate__ can be only called on leaf Tensors')
if len(state) == 4:
# legacy serialization of Tensor
self.set_(*state)
return
elif len(state) == 5:
# legacy serialization of Variable
self.data = state[0]
state = (state[3], state[4], state[2])
# The setting of _backward_hooks is expected to be a no-op.
# See Note [Don't serialize hooks]
self.requires_grad, _, self._backward_hooks = state
def __repr__(self):
# All strings are unicode in Python 3, while we have to encode unicode
# strings in Python2. If we can't, let python decide the best
# characters to replace unicode characters with.
if sys.version_info > (3, ):
return torch._tensor_str._str(self)
else:
if hasattr(sys.stdout, 'encoding'):
return torch._tensor_str._str(self).encode(
sys.stdout.encoding or 'UTF-8', 'replace')
else:
return torch._tensor_str._str(self).encode('UTF-8', 'replace')
def backward(self, gradient=None, retain_graph=None, create_graph=False):
r"""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 them before calling it.
Arguments:
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``.
"""
torch.autograd.backward(self, gradient, retain_graph, create_graph)
def register_hook(self, hook):
r"""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 :attr:`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
"""
if not self.requires_grad:
raise RuntimeError("cannot register a hook on a tensor that "
"doesn't require gradient")
if self._backward_hooks is None:
self._backward_hooks = OrderedDict()
if self.grad_fn is not None:
self.grad_fn._register_hook_dict(self)
handle = hooks.RemovableHandle(self._backward_hooks)
self._backward_hooks[handle.id] = hook
return handle
def reinforce(self, reward):
def trim(str):
return '\n'.join([line.strip() for line in str.split('\n')])
raise RuntimeError(
trim(r"""reinforce() was removed.
Use torch.distributions instead.
See https://pytorch.org/docs/master/distributions.html
Instead of:
probs = policy_network(state)
action = probs.multinomial()
next_state, reward = env.step(action)
action.reinforce(reward)
action.backward()
Use:
probs = policy_network(state)
# NOTE: categorical is equivalent to what used to be called multinomial
m = torch.distributions.Categorical(probs)
action = m.sample()
next_state, reward = env.step(action)
loss = -m.log_prob(action) * reward
loss.backward()
"""))
detach = _add_docstr(
_C._TensorBase.detach, r"""
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_ = _add_docstr(
_C._TensorBase.detach_, r"""
Detaches the Tensor from the graph that created it, making it a leaf.
Views cannot be detached in-place.
""")
def retain_grad(self):
r"""Enables .grad attribute for non-leaf Tensors."""
if self.grad_fn is None: # no-op for leaves
return
if not self.requires_grad:
raise RuntimeError(
"can't retain_grad on Tensor that has requires_grad=False")
if hasattr(self, 'retains_grad'):
return
weak_self = weakref.ref(self)
def retain_grad_hook(grad):
var = weak_self()
if var is None:
return
if var._grad is None:
var._grad = grad.clone()
else:
var._grad = var._grad + grad
self.register_hook(retain_grad_hook)
self.retains_grad = True
def is_shared(self):
r"""Checks if tensor is in shared memory.
This is always ``True`` for CUDA tensors.
"""
return self.storage().is_shared()
def share_memory_(self):
r"""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.
"""
self.storage().share_memory_()
return self
def __reversed__(self):
r"""Reverses the tensor along dimension 0."""
if self.dim() == 0:
return self
else:
return self.flip(0)
def norm(self, p="fro", dim=None, keepdim=False, dtype=None):
r"""See :func:`torch.norm`"""
return torch.norm(self, p, dim, keepdim, dtype=dtype)
def lu(self, pivot=True, get_infos=False):
r"""See :func:`torch.lu`"""
# If get_infos is True, then we don't need to check for errors and vice versa
LU, pivots, infos = torch._lu_with_info(self,
pivot=pivot,
check_errors=(not get_infos))
if get_infos:
return LU, pivots, infos
else:
return LU, pivots
def stft(self,
n_fft,
hop_length=None,
win_length=None,
window=None,
center=True,
pad_mode='reflect',
normalized=False,
onesided=True):
r"""See :func:`torch.stft`
.. warning::
This function changed signature at version 0.4.1. Calling with
the previous signature may cause error or return incorrect result.
"""
return torch.stft(self, n_fft, hop_length, win_length, window, center,
pad_mode, normalized, onesided)
def resize(self, *sizes):
warnings.warn("non-inplace resize is deprecated")
from torch.autograd._functions import Resize
return Resize.apply(self, sizes)
def resize_as(self, tensor):
warnings.warn("non-inplace resize_as is deprecated")
from torch.autograd._functions import Resize
return Resize.apply(self, tensor.size())
def split(self, split_size, dim=0):
r"""See :func:`torch.split`
"""
if isinstance(split_size, int):
return super(Tensor, self).split(split_size, dim)
else:
return super(Tensor, self).split_with_sizes(split_size, dim)
def unique(self,
sorted=True,
return_inverse=False,
return_counts=False,
dim=None):
r"""Returns the unique elements of the input tensor.
See :func:`torch.unique`
"""
return torch.unique(self,
sorted=sorted,
return_inverse=return_inverse,
return_counts=return_counts,
dim=dim)
def unique_consecutive(self,
return_inverse=False,
return_counts=False,
dim=None):
r"""Eliminates all but the first element from every consecutive group of equivalent elements.
See :func:`torch.unique_consecutive`
"""
return torch.unique_consecutive(self,
return_inverse=return_inverse,
return_counts=return_counts,
dim=dim)
def __rsub__(self, other):
return _C._VariableFunctions.rsub(self, other)
def __rdiv__(self, other):
if self.dtype.is_floating_point:
return self.reciprocal() * other
else:
return (self.double().reciprocal() * other).type_as(self)
__rtruediv__ = __rdiv__
__itruediv__ = _C._TensorBase.__idiv__
__pow__ = _C._TensorBase.pow
def __format__(self, format_spec):
if self.dim() == 0:
return self.item().__format__(format_spec)
return object.__format__(self, format_spec)
def __ipow__(self, other):
raise NotImplementedError("in-place pow not implemented")
def __rpow__(self, other):
return self.new_tensor(other)**self
def __floordiv__(self, other):
result = self / other
if result.dtype.is_floating_point:
result = result.trunc()
return result
def __rfloordiv__(self, other):
result = other / self
if result.dtype.is_floating_point:
result = result.trunc()
return result
__neg__ = _C._TensorBase.neg
__eq__ = _C._TensorBase.eq
__ne__ = _C._TensorBase.ne
__lt__ = _C._TensorBase.lt
__le__ = _C._TensorBase.le
__gt__ = _C._TensorBase.gt
__ge__ = _C._TensorBase.ge
__abs__ = _C._TensorBase.abs
def __len__(self):
if self.dim() == 0:
raise TypeError("len() of a 0-d tensor")
return self.shape[0]
def __iter__(self):
# NB: we use 'imap' and not 'map' here, so that in Python 2 we get a
# generator and don't eagerly perform all the indexes. This could
# save us work, and also helps keep trace ordering deterministic
# (e.g., if you zip(*hiddens), the eager map will force all the
# indexes of hiddens[0] before hiddens[1], while the generator
# map will interleave them.)
if self.dim() == 0:
raise TypeError('iteration over a 0-d tensor')
if torch._C._get_tracing_state():
warnings.warn(
'Iterating over a tensor might cause the trace to be incorrect. '
'Passing a tensor of different shape won\'t change the number of '
'iterations executed (and might lead to errors or silently give '
'incorrect results).',
category=RuntimeWarning)
return iter(imap(lambda i: self[i], range(self.size(0))))
def __hash__(self):
return id(self)
def __dir__(self):
tensor_methods = dir(self.__class__)
tensor_methods.remove('volatile') # deprecated
attrs = list(self.__dict__.keys())
keys = tensor_methods + attrs
# property only available dense, cuda tensors
if (not self.is_cuda) or self.is_sparse:
keys.remove("__cuda_array_interface__")
return sorted(keys)
# Numpy array interface, to support `numpy.asarray(tensor) -> ndarray`
__array_priority__ = 1000 # prefer Tensor ops over numpy ones
def __array__(self, dtype=None):
if dtype is None:
return self.numpy()
else:
return self.numpy().astype(dtype, copy=False)
# Wrap Numpy array again in a suitable tensor when done, to support e.g.
# `numpy.sin(tensor) -> tensor` or `numpy.greater(tensor, 0) -> ByteTensor`
def __array_wrap__(self, array):
if array.dtype == bool:
# Workaround, torch has no built-in bool tensor
array = array.astype('uint8')
return torch.from_numpy(array)
def __contains__(self, element):
r"""Check if `element` is present in tensor
Arguments:
element (Tensor or scalar): element to be checked
for presence in current tensor"
"""
if isinstance(element, (torch.Tensor, Number)):
return (element == self).any().item()
raise RuntimeError(
"Tensor.__contains__ only supports Tensor or scalar, but you passed in a %s."
% type(element))
@property
def __cuda_array_interface__(self):
"""Array view description for cuda tensors.
See:
https://numba.pydata.org/numba-doc/latest/cuda/cuda_array_interface.html
"""
# raise AttributeError for unsupported tensors, so that
# hasattr(cpu_tensor, "__cuda_array_interface__") is False.
if not self.is_cuda:
raise AttributeError(
"Can't get __cuda_array_interface__ on non-CUDA tensor type: %s "
"If CUDA data is required use tensor.cuda() to copy tensor to device memory."
% self.type())
if self.is_sparse:
raise AttributeError(
"Can't get __cuda_array_interface__ on sparse type: %s "
"Use Tensor.to_dense() to convert to a dense tensor first." %
self.type())
# RuntimeError, matching tensor.__array__() behavior.
if self.requires_grad:
raise RuntimeError(
"Can't get __cuda_array_interface__ on Variable that requires grad. "
"If gradients aren't required, use var.detach() to get Variable that doesn't require grad."
)
# CUDA devices are little-endian and tensors are stored in native byte
# order. 1-byte entries are endian-agnostic.
typestr = {
torch.float16: "<f2",
torch.float32: "<f4",
torch.float64: "<f8",
torch.uint8: "|u1",
torch.int8: "|i1",
torch.int16: "<i2",
torch.int32: "<i4",
torch.int64: "<i8",
}[self.dtype]
itemsize = self.storage().element_size()
shape = tuple(self.shape)
strides = tuple(s * itemsize for s in self.stride())
data = (self.data_ptr(), False) # read-only is false
return dict(typestr=typestr,
shape=shape,
strides=strides,
data=data,
version=1)
def refine_names(self, *names):
r"""Refines the dimension names of :attr:`self` according to :attr:`names`.
Refining is a special case of renaming that "lifts" unnamed dimensions.
A ``None`` dim can be refined to have any name; a named dim can only be
refined to have the same name.
Because named tensors can coexist with unnamed tensors, refining names
gives a nice way to write named-tensor-aware code that works with both
named and unnamed tensors.
:attr:`names` may contain up to one Ellipsis (``...``).
The Ellipsis is expanded greedily; it is expanded in-place to fill
:attr:`names` to the same length as ``self.dim()`` using names from the
corresponding indices of ``self.names``.
Python 2 does not support Ellipsis but one may use a string literal
instead (``'...'``).
Arguments:
names (iterable of str): The desired names of the output tensor. May
contain up to one Ellipsis.
Examples::
>>> imgs = torch.randn(32, 3, 128, 128)
>>> named_imgs = imgs.refine_names('N', 'C', 'H', 'W')
>>> named_imgs.names
('N', 'C', 'H', 'W')
>>> tensor = torch.randn(2, 3, 5, 7, 11)
>>> tensor = tensor.refine_names('A', ..., 'B', 'C')
>>> tensor.names
('A', None, None, 'B', 'C')
.. warning::
The named tensor API is experimental and subject to change.
"""
names = resolve_ellipsis(names, self.names, 'refine_names')
return super(Tensor, self).refine_names(names)
def align_to(self, *names):
r"""Permutes the dimensions of the :attr:`self` tensor to match the order
specified in :attr:`names`, adding size-one dims for any new names.
All of the dims of :attr:`self` must be named in order to use this method.
The resulting tensor is a view on the original tensor.
All dimension names of :attr:`self` must be present in :attr:`names`.
:attr:`names` may contain additional names that are not in ``self.names``;
the output tensor has a size-one dimension for each of those new names.
:attr:`names` may contain up to one Ellipsis (``...``).
The Ellipsis is expanded to be equal to all dimension names of :attr:`self`
that are not mentioned in :attr:`names`, in the order that they appear
in :attr:`self`.
Python 2 does not support Ellipsis but one may use a string literal
instead (``'...'``).
Arguments:
names (iterable of str): The desired dimension ordering of the
output tensor. May contain up to one Ellipsis that is expanded
to all unmentioned dim names of :attr:`self`.
Examples::
>>> tensor = torch.randn(2, 2, 2, 2, 2, 2)
>>> named_tensor = tensor.refine_names('A', 'B', 'C', 'D', 'E', 'F')
# Move the F and E dims to the front while keeping the rest in order
>>> named_tensor.align_to('F', 'E', ...)
.. warning::
The named tensor API is experimental and subject to change.
"""
return super(Tensor, self).align_to(
resolve_ellipsis(names,
self.names,
'align_to',
is_positional=False))
def unflatten(self, dim, namedshape):
r"""Unflattens the named dimension :attr:`dim`, viewing it in the shape
specified by :attr:`namedshape`.
Arguments:
namedshape: (iterable of ``(name, size)`` tuples).
Examples::
>>> flat_imgs = torch.rand(32, 3 * 128 * 128, names=('N', 'features'))
>>> imgs = flat_imgs.unflatten('features', (('C', 3), ('H', 128), ('W', 128)))
>>> imgs.names, images.shape
(('N', 'C', 'H', 'W'), torch.Size([32, 3, 128, 128]))
.. warning::
The named tensor API is experimental and subject to change.
"""
names, sizes = unzip_namedshape(namedshape)
return super(Tensor, self).unflatten(dim, sizes, names)
def rename_(self, *names, **rename_map):
"""In-place version of :meth:`~Tensor.rename`."""
# Note [rename_ / rename API]
# The Python API for these is different from the C++ API. In Python:
# 1) tensor.rename(*names) takes a vararglist of names
# 2) tensor.rename(**rename_map) takes a map of names to rename.
# C++ is static, making it difficult to implement similar behavior.
return update_names(self, names, rename_map, inplace=True)
def rename(self, *names, **rename_map):
"""Renames dimension names of :attr:`self`.
There are two main usages:
``self.rename(**rename_map)`` returns a view on tensor that has dims
renamed as specified in the mapping :attr:`rename_map`.
``self.rename(*names)`` returns a view on tensor, renaming all
dimensions positionally using :attr:`names`.
Use ``self.rename(None)`` to drop names on a tensor.
One cannot specify both positional args :attr:`names` and keyword args
:attr:`rename_map`.
Examples::
>>> imgs = torch.rand(2, 3, 5, 7, names=('N', 'C', 'H', 'W'))
>>> renamed_imgs = imgs.rename(N='batch', C='channels')
>>> renamed_imgs.names
('batch', 'channels', 'H', 'W')
>>> renamed_imgs = imgs.rename(None)
>>> renamed_imgs.names
(None,)
>>> renamed_imgs = imgs.rename('batch', 'channel', 'height', 'width')
>>> renamed_imgs.names
('batch', 'channel', 'height', 'width')
.. warning::
The named tensor API is experimental and subject to change.
"""
# See Note [rename_ / rename API]
return update_names(self, names, rename_map, inplace=False)
def _update_names(self, names, inplace):
# See Note [rename_ / rename API]
if inplace:
return super(Tensor, self).rename_(names)
else:
return super(Tensor, self).rename(names)
__module__ = 'torch'
'sum',
'softmax',
'log_softmax',
]
addmm = _add_docstr(
_sparse._sparse_addmm, r"""
sparse.addmm(mat, mat1, mat2, *, beta=1., alpha=1.) -> Tensor
This function does exact same thing as :func:`torch.addmm` in the forward,
except that it supports backward for sparse COO matrix :attr:`mat1`.
When :attr:`mat1` is a COO tensor it must have `sparse_dim = 2`.
When inputs are COO tensors, this function also supports backward for both inputs.
Supports both CSR and COO storage formats.
.. note::
This function doesn't support computing derivaties with respect to CSR matrices.
Args:
mat (Tensor): a dense matrix to be added
mat1 (Tensor): a sparse matrix to be multiplied
mat2 (Tensor): a dense matrix to be multiplied
beta (Number, optional): multiplier for :attr:`mat` (:math:`\beta`)
alpha (Number, optional): multiplier for :math:`mat1 @ mat2` (:math:`\alpha`)
""")
mm = _add_docstr(
_sparse._sparse_mm, r"""
Performs a matrix multiplication of the sparse matrix :attr:`mat1`
and the (sparse or strided) matrix :attr:`mat2`. Similar to :func:`torch.mm`, if :attr:`mat1` is a
class Variable(_C._VariableBase):
"""Wraps a tensor and records the operations applied to it.
Variable is a thin wrapper around a Tensor object, that also holds
the gradient w.r.t. to it, and a reference to a function that created it.
This reference allows retracing the whole chain of operations that
created the data. If the Variable has been created by the user, its grad_fn
will be ``None`` and we call such objects *leaf* Variables.
Since autograd only supports scalar valued function differentiation, grad
size always matches the data size. Also, grad is normally only allocated
for leaf variables, and will be always zero otherwise.
Attributes:
data: Wrapped tensor of any type.
grad: Variable holding the gradient of type and location matching
the ``.data``. This attribute is lazily allocated and can't
be reassigned.
requires_grad: Boolean indicating whether the Variable has been
created by a subgraph containing any Variable, that requires it.
See :ref:`excluding-subgraphs` for more details.
Can be changed only on leaf Variables.
volatile: Boolean indicating that the Variable should be used in
inference mode, i.e. don't save the history. See
:ref:`excluding-subgraphs` for more details.
Can be changed only on leaf Variables.
is_leaf: Boolean indicating if the Variable is a graph leaf (i.e
if it was created by the user).
grad_fn: Gradient function graph trace.
Parameters:
data (any tensor class): Tensor to wrap.
requires_grad (bool): Value of the requires_grad flag. **Keyword only.**
volatile (bool): Value of the volatile flag. **Keyword only.**
"""
def __getitem__(self, key):
if torch.is_tensor(key):
key = Variable(key) # auto-wrap tensors
if isinstance(key, Variable):
if type(key.data).__name__ == 'ByteTensor':
return MaskedSelect.apply(self, key)
elif type(key.data).__name__ == 'LongTensor':
return IndexSelect.apply(self, 0, key)
# else fall through and raise an error in Index
return Index.apply(self, key)
def __setitem__(self, key, value):
if isinstance(key, Variable) and type(
key.data).__name__ == 'ByteTensor':
if isinstance(value, Variable):
return MaskedScatter.apply(self, key, value, True)
else:
return MaskedFill.apply(self, key, value, True)
else:
return SetItem.apply(self, key, value)
def __deepcopy__(self, memo):
if not self.is_leaf:
raise RuntimeError(
"Only Variables created explicitly by the user "
"(graph leaves) support the deepcopy protocol at the moment")
result = type(self)(self.data.clone())
result.requires_grad = self.requires_grad
result.volatile = self.volatile
memo[id(self)] = result
return result
def __reduce_ex__(self, proto):
state = (self.requires_grad, self.volatile, self._backward_hooks)
if proto > 1:
return type(self), (self.data, ), state
if sys.version_info[0] == 2:
from copy_reg import __newobj__
else:
from copyreg import __newobj__
return __newobj__, (type(self), self.data), state
def __setstate__(self, state):
if len(state) == 5:
# legacy serialization of Variable
self.data = state[0]
state = (state[3], state[4], state[2])
if not self.is_leaf:
raise RuntimeError(
'__setstate__ can be only called on leaf variables')
self.requires_grad, self.volatile, self._backward_hooks = state
def __repr__(self):
return 'Variable containing:' + self.data.__repr__()
def __bool__(self):
if self.data.numel() <= 1:
return self.data.__bool__()
raise RuntimeError("bool value of Variable containing " +
torch.typename(self.data) +
" with more than one value is ambiguous")
__nonzero__ = __bool__
def __int__(self):
return int(self.data)
def __long__(self):
return long(self.data)
def __float__(self):
return float(self.data)
def backward(self,
gradient=None,
retain_graph=None,
create_graph=None,
retain_variables=None):
"""Computes the gradient of current variable w.r.t. graph leaves.
The graph is differentiated using the chain rule. If the variable 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 them before calling it.
Arguments:
gradient (Tensor, Variable or None): Gradient w.r.t. the
variable. If it is a tensor, it will be automatically converted
to a Variable that is volatile unless ``create_graph`` is True.
None values can be specified for scalar Variables 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``, unless ``gradient`` is a volatile
Variable.
"""
torch.autograd.backward(self, gradient, retain_graph, create_graph,
retain_variables)
def register_hook(self, hook):
"""Registers a backward hook.
The hook will be called every time a gradient with respect to the
variable is computed. The hook should have the following signature::
hook(grad) -> Variable or None
The hook should not modify its argument, but it can optionally return
a new gradient which will be used in place of :attr:`grad`.
This function returns a handle with a method ``handle.remove()``
that removes the hook from the module.
Example:
>>> v = Variable(torch.Tensor([0, 0, 0]), requires_grad=True)
>>> h = v.register_hook(lambda grad: grad * 2) # double the gradient
>>> v.backward(torch.Tensor([1, 1, 1]))
>>> v.grad.data
2
2
2
[torch.FloatTensor of size 3]
>>> h.remove() # removes the hook
"""
if self.volatile:
raise RuntimeError("cannot register a hook on a volatile variable")
if not self.requires_grad:
raise RuntimeError("cannot register a hook on a variable that "
"doesn't require gradient")
if self._backward_hooks is None:
self._backward_hooks = OrderedDict()
if self.grad_fn is not None:
self.grad_fn._register_hook_dict(self)
handle = hooks.RemovableHandle(self._backward_hooks)
self._backward_hooks[handle.id] = hook
return handle
def reinforce(self, reward):
def trim(str):
return '\n'.join([line.strip() for line in str.split('\n')])
raise RuntimeError(
trim(r"""reinforce() was removed.
Use torch.distributions instead.
See http://pytorch.org/docs/master/distributions.html
Instead of:
probs = policy_network(state)
action = probs.multinomial()
next_state, reward = env.step(action)
action.reinforce(reward)
action.backward()
Use:
probs = policy_network(state)
# NOTE: categorical is equivalent to what used to be called multinomial
m = torch.distributions.Categorical(probs)
action = m.sample()
next_state, reward = env.step(action)
loss = -m.log_prob(action) * reward
loss.backward()
"""))
detach = _add_docstr(
_C._VariableBase.detach, r"""
Returns a new Variable, detached from the current graph.
Result will never require gradient. If the input is volatile, the output
will be volatile too.
.. note::
Returned Variable uses the same data tensor, as the original one, and
in-place modifications on either of them will be seen, and may trigger
errors in correctness checks.
""")
detach_ = _add_docstr(
_C._VariableBase.detach_, r"""
Detaches the Variable from the graph that created it, making it a leaf.
Views cannot be detached in-place.
""")
def retain_grad(self):
"""Enables .grad attribute for non-leaf Variables."""
if self.grad_fn is None: # no-op for leaves
return
if not self.requires_grad:
raise RuntimeError(
"can't retain_grad on Variable that has requires_grad=False")
if hasattr(self, 'retains_grad'):
return
weak_self = weakref.ref(self)
def retain_grad_hook(grad):
var = weak_self()
if var is None:
return
if var._grad is None:
var._grad = grad.clone()
else:
var._grad = var._grad + grad
self.register_hook(retain_grad_hook)
self.retains_grad = True
def type(self, t):
if t != type(self.data):
return Type.apply(self, t)
return self
def type_as(self, t):
if isinstance(t, Variable):
t = t.data
return self.type(type(t))
def _get_type(self, name):
module = torch._import_dotted_name(self.data.__module__)
return getattr(module, name)
def cuda(self, device=None, async=False):
return CudaTransfer.apply(self, device, async)
Tensor = torch.Tensor
entr = _add_docstr(
_special.special_entr, r"""
entr(input, *, out=None) -> Tensor
Computes the entropy on :attr:`input` (as defined below), elementwise.
.. math::
\begin{align}
\text{entr(x)} = \begin{cases}
-x * \ln(x) & x > 0 \\
0 & x = 0.0 \\
-\infty & x < 0
\end{cases}
\end{align}
""" + """
Args:
input (Tensor): the input tensor.
Keyword args:
out (Tensor, optional): the output tensor.
Example::
>>> a = torch.arange(-0.5, 1, 0.5)
>>> a
tensor([-0.5000, 0.0000, 0.5000])
>>> torch.special.entr(a)
tensor([ -inf, 0.0000, 0.3466])
""")
psi = _add_docstr(
Tensor = torch.Tensor
entr = _add_docstr(
_special.special_entr, r"""
entr(input, *, out=None) -> Tensor
Computes the entropy on :attr:`input` (as defined below), elementwise.
.. math::
\text{entr(x)} = \begin{cases}
-x * \ln(x) & x > 0 \\
0 & x = 0.0 \\
-\infty & x < 0
\end{cases}
""" + """
Args:
input (Tensor): the input tensor.
Keyword args:
out (Tensor, optional): the output tensor.
Example::
>>> a = torch.arange(-0.5, 1, 0.5)
>>> a
tensor([-0.5000, 0.0000, 0.5000])
>>> torch.special.entr(a)
tensor([ -inf, 0.0000, 0.3466])
""")
gammaln = _add_docstr(
_special.special_gammaln, r"""
cholesky = _add_docstr(_linalg.linalg_cholesky, r"""
linalg.cholesky(input, *, out=None) -> Tensor
Computes the Cholesky decomposition of a Hermitian (or symmetric for real-valued matrices)
positive-definite matrix or the Cholesky decompositions for a batch of such matrices.
Each decomposition has the form:
.. math::
\text{input} = LL^H
where :math:`L` is a lower-triangular matrix and :math:`L^H` is the conjugate transpose of :math:`L`,
which is just a transpose for the case of real-valued input matrices.
In code it translates to ``input = L @ L.t()` if :attr:`input` is real-valued and
``input = L @ L.conj().t()`` if :attr:`input` is complex-valued.
The batch of :math:`L` matrices is returned.
Supports real-valued and complex-valued inputs.
.. note:: If :attr:`input` is not a Hermitian positive-definite matrix, or if it's a batch of matrices
and one or more of them is not a Hermitian positive-definite matrix, then a RuntimeError will be thrown.
If :attr:`input` is a batch of matrices, then the error message will include the batch index
of the first matrix that is not Hermitian positive-definite.
.. warning:: This function always checks whether :attr:`input` is a Hermitian positive-definite matrix
using `info` argument to LAPACK/MAGMA call. For CUDA this causes cross-device memory synchronization.
Args:
input (Tensor): the input tensor of size :math:`(*, n, n)` consisting of Hermitian positive-definite
:math:`n \times n` matrices, where `*` is zero or more batch dimensions.
Keyword args:
out (Tensor, optional): The output tensor. Ignored if ``None``. Default: ``None``
Examples::
>>> a = torch.randn(2, 2, dtype=torch.complex128)
>>> a = torch.mm(a, a.t().conj()) # creates a Hermitian positive-definite matrix
>>> l = torch.linalg.cholesky(a)
>>> a
tensor([[2.5266+0.0000j, 1.9586-2.0626j],
[1.9586+2.0626j, 9.4160+0.0000j]], dtype=torch.complex128)
>>> l
tensor([[1.5895+0.0000j, 0.0000+0.0000j],
[1.2322+1.2976j, 2.4928+0.0000j]], dtype=torch.complex128)
>>> torch.mm(l, l.t().conj())
tensor([[2.5266+0.0000j, 1.9586-2.0626j],
[1.9586+2.0626j, 9.4160+0.0000j]], dtype=torch.complex128)
>>> a = torch.randn(3, 2, 2, dtype=torch.float64)
>>> a = torch.matmul(a, a.transpose(-2, -1)) # creates a symmetric positive-definite matrix
>>> l = torch.linalg.cholesky(a)
>>> a
tensor([[[ 1.1629, 2.0237],
[ 2.0237, 6.6593]],
[[ 0.4187, 0.1830],
[ 0.1830, 0.1018]],
[[ 1.9348, -2.5744],
[-2.5744, 4.6386]]], dtype=torch.float64)
>>> l
tensor([[[ 1.0784, 0.0000],
[ 1.8766, 1.7713]],
[[ 0.6471, 0.0000],
[ 0.2829, 0.1477]],
[[ 1.3910, 0.0000],
[-1.8509, 1.1014]]], dtype=torch.float64)
>>> torch.allclose(torch.matmul(l, l.transpose(-2, -1)), a)
True
""")
class Tensor(torch._C._TensorBase):
def __deepcopy__(self, memo):
if not self.is_leaf:
raise RuntimeError(
"Only Tensors created explicitly by the user "
"(graph leaves) support the deepcopy protocol at the moment")
if id(self) in memo:
return memo[id(self)]
with torch.no_grad():
if self.is_sparse:
new_tensor = self.clone()
else:
new_storage = self.storage().__deepcopy__(memo)
new_tensor = self.new()
new_tensor.set_(new_storage, self.storage_offset(),
self.size(), self.stride())
memo[id(self)] = new_tensor
new_tensor.requires_grad = self.requires_grad
return new_tensor
def __reduce_ex__(self, proto):
# See Note [Don't serialize hooks]
torch.utils.hooks.warn_if_has_hooks(self)
if self.is_quantized:
args = (self.storage(), self.storage_offset(), tuple(self.size()),
self.stride(), self.q_scale().item(),
self.q_zero_point().item(), self.requires_grad,
OrderedDict()) # TODO: self.qscheme()
return (torch._utils._rebuild_qtensor, args)
else:
args = (self.storage(), self.storage_offset(), tuple(self.size()),
self.stride(), self.requires_grad, OrderedDict()
) # previously was self._backward_hooks
return (torch._utils._rebuild_tensor_v2, args)
def __setstate__(self, state):
# Warning: this method is NOT called when you torch.load() a tensor;
# that is managed by _rebuild_tensor_v2
if not self.is_leaf:
raise RuntimeError(
'__setstate__ can be only called on leaf Tensors')
if len(state) == 4:
# legacy serialization of Tensor
self.set_(*state)
return
elif len(state) == 5:
# legacy serialization of Variable
self.data = state[0]
state = (state[3], state[4], state[2])
# The setting of _backward_hooks is expected to be a no-op.
# See Note [Don't serialize hooks]
self.requires_grad, _, self._backward_hooks = state
def __repr__(self):
# All strings are unicode in Python 3, while we have to encode unicode
# strings in Python2. If we can't, let python decide the best
# characters to replace unicode characters with.
if sys.version_info > (3, ):
return torch._tensor_str._str(self)
else:
if hasattr(sys.stdout, 'encoding'):
return torch._tensor_str._str(self).encode(
sys.stdout.encoding or 'UTF-8', 'replace')
else:
return torch._tensor_str._str(self).encode('UTF-8', 'replace')
def backward(self, gradient=None, retain_graph=None, create_graph=False):
r"""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 them before calling it.
Arguments:
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``.
"""
torch.autograd.backward(self, gradient, retain_graph, create_graph)
def register_hook(self, hook):
r"""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 :attr:`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
"""
if not self.requires_grad:
raise RuntimeError("cannot register a hook on a tensor that "
"doesn't require gradient")
if self._backward_hooks is None:
self._backward_hooks = OrderedDict()
if self.grad_fn is not None:
self.grad_fn._register_hook_dict(self)
handle = hooks.RemovableHandle(self._backward_hooks)
self._backward_hooks[handle.id] = hook
return handle
def reinforce(self, reward):
def trim(str):
return '\n'.join([line.strip() for line in str.split('\n')])
raise RuntimeError(
trim(r"""reinforce() was removed.
Use torch.distributions instead.
See https://pytorch.org/docs/master/distributions.html
Instead of:
probs = policy_network(state)
action = probs.multinomial()
next_state, reward = env.step(action)
action.reinforce(reward)
action.backward()
Use:
probs = policy_network(state)
# NOTE: categorical is equivalent to what used to be called multinomial
m = torch.distributions.Categorical(probs)
action = m.sample()
next_state, reward = env.step(action)
loss = -m.log_prob(action) * reward
loss.backward()
"""))
detach = _add_docstr(
_C._TensorBase.detach, r"""
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_ = _add_docstr(
_C._TensorBase.detach_, r"""
Detaches the Tensor from the graph that created it, making it a leaf.
Views cannot be detached in-place.
""")
def retain_grad(self):
r"""Enables .grad attribute for non-leaf Tensors."""
if self.grad_fn is None: # no-op for leaves
return
if not self.requires_grad:
raise RuntimeError(
"can't retain_grad on Tensor that has requires_grad=False")
if hasattr(self, 'retains_grad'):
return
weak_self = weakref.ref(self)
def retain_grad_hook(grad):
var = weak_self()
if var is None:
return
if var._grad is None:
var._grad = grad.clone()
else:
var._grad = var._grad + grad
self.register_hook(retain_grad_hook)
self.retains_grad = True
def is_pinned(self):
r"""Returns true if this tensor resides in pinned memory"""
storage = self.storage()
return storage.is_pinned() if storage else False
def is_shared(self):
r"""Checks if tensor is in shared memory.
This is always ``True`` for CUDA tensors.
"""
return self.storage().is_shared()
def share_memory_(self):
r"""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.
"""
self.storage().share_memory_()
return self
def __reversed__(self):
r"""Reverses the tensor along dimension 0."""
if self.dim() == 0:
return self
else:
return self.flip(0)
def norm(self, p="fro", dim=None, keepdim=False, dtype=None):
r"""See :func:`torch.norm`"""
return torch.norm(self, p, dim, keepdim, dtype=dtype)
def pstrf(self, upper=True):
r"""See :func:`torch.pstrf`"""
warnings.warn(
"torch.pstrf is deprecated in favour of torch.cholesky and will be removed "
"in the next release.",
stacklevel=2)
return super(Tensor, self).pstrf(upper=upper)
def potrf(self, upper=True):
r"""See :func:`torch.cholesky`"""
warnings.warn(
"torch.potrf is deprecated in favour of torch.cholesky and will be removed "
"in the next release. Please use torch.cholesky instead and note that the "
":attr:`upper` argument in torch.cholesky defaults to ``False``.",
stacklevel=2)
return super(Tensor, self).cholesky(upper=upper)
def potri(self, upper=True):
r"""See :func:`torch.cholesky_inverse`"""
warnings.warn(
"torch.potri is deprecated in favour of torch.cholesky_inverse and will be "
"removed in the next release. Please use torch.cholesky_inverse instead and "
"note that the :attr:`upper` argument in torch.cholesky_inverse defaults to "
"``False``.",
stacklevel=2)
return super(Tensor, self).cholesky_inverse(upper=upper)
def potrs(self, u, upper=True):
r"""See :func:`torch.cholesky_solve`"""
warnings.warn(
"torch.potrs is deprecated in favour of torch.cholesky_solve and "
"will be removed in the next release. Please use torch.cholesky_solve instead "
"and note that the :attr:`upper` argument in torch.cholesky_solve defaults "
"to ``False``.",
stacklevel=2)
return super(Tensor, self).cholesky_solve(u, upper=upper)
def gesv(self, A):
r"""See :func:`torch.solve`"""
warnings.warn(
"torch.gesv is deprecated in favour of torch.solve and will be removed in the "
"next release. Please use torch.solve instead.",
stacklevel=2)
return super(Tensor, self).solve(A)
def trtrs(self, A, upper=True, transpose=False, unitriangular=False):
r"""See :func:`torch.triangular_solve`"""
warnings.warn(
"torch.trtrs is deprecated in favour of torch.triangular_solve and will be "
"removed in the next release. Please use torch.triangular_solve instead.",
stacklevel=2)
return super(Tensor,
self).triangular_solve(A,
upper=upper,
transpose=transpose,
unitriangular=unitriangular)
def btrifact(self, pivot=True):
r"""See :func:`torch.lu`"""
warnings.warn(
"torch.btrifact is deprecated in favour of torch.lu and will be removed in "
"the next release. Please use torch.lu instead.",
stacklevel=2)
return torch._lu_with_info(self, pivot=pivot, check_errors=True)
def btrifact_with_info(self, pivot=True):
r"""See :func:`torch.lu`"""
warnings.warn(
"torch.btrifact_with_info is deprecated in favour of torch.lu with the "
"get_infos argument and will be removed in the next release. Please use "
"torch.lu with the get_infos argument set to True instead.",
stacklevel=2)
return torch._lu_with_info(self, pivot=pivot, check_errors=False)
def btrisolve(self, LU_data, LU_pivots):
r"""See :func:`torch.lu_solve`"""
warnings.warn(
"torch.btrisolve is deprecated in favour of torch.lu_solve and will be "
"removed in the next release. Please use torch.lu_solve instead.",
stacklevel=2)
return super(Tensor, self).lu_solve(LU_data=LU_data,
LU_pivots=LU_pivots)
def lu(self, pivot=True, get_infos=False):
r"""See :func:`torch.lu`"""
# If get_infos is True, then we don't need to check for errors and vice versa
LU, pivots, infos = torch._lu_with_info(self,
pivot=pivot,
check_errors=(not get_infos))
if get_infos:
return LU, pivots, infos
else:
return LU, pivots
def stft(self,
n_fft,
hop_length=None,
win_length=None,
window=None,
center=True,
pad_mode='reflect',
normalized=False,
onesided=True):
r"""See :func:`torch.stft`
.. warning::
This function changed signature at version 0.4.1. Calling with
the previous signature may cause error or return incorrect result.
"""
return torch.stft(self, n_fft, hop_length, win_length, window, center,
pad_mode, normalized, onesided)
def resize(self, *sizes):
warnings.warn("non-inplace resize is deprecated")
from torch.autograd._functions import Resize
return Resize.apply(self, sizes)
def resize_as(self, tensor):
warnings.warn("non-inplace resize_as is deprecated")
from torch.autograd._functions import Resize
return Resize.apply(self, tensor.size())
def split(self, split_size, dim=0):
r"""See :func:`torch.split`
"""
if isinstance(split_size, int):
return super(Tensor, self).split(split_size, dim)
else:
return super(Tensor, self).split_with_sizes(split_size, dim)
def unique(self,
sorted=True,
return_inverse=False,
return_counts=False,
dim=None):
r"""Returns the unique elements of the input tensor.
See :func:`torch.unique`
"""
return torch.unique(self,
sorted=sorted,
return_inverse=return_inverse,
return_counts=return_counts,
dim=dim)
def unique_consecutive(self,
return_inverse=False,
return_counts=False,
dim=None):
r"""Eliminates all but the first element from every consecutive group of equivalent elements.
See :func:`torch.unique_consecutive`
"""
return torch.unique_consecutive(self,
return_inverse=return_inverse,
return_counts=return_counts,
dim=dim)
def __rsub__(self, other):
return _C._VariableFunctions.rsub(self, other)
def __rdiv__(self, other):
if self.dtype.is_floating_point:
return self.reciprocal() * other
else:
return (self.double().reciprocal() * other).type_as(self)
__rtruediv__ = __rdiv__
__itruediv__ = _C._TensorBase.__idiv__
__pow__ = _C._TensorBase.pow
def __format__(self, format_spec):
if self.dim() == 0:
return self.item().__format__(format_spec)
return object.__format__(self, format_spec)
def __ipow__(self, other):
raise NotImplementedError("in-place pow not implemented")
def __rpow__(self, other):
return self.new_tensor(other)**self
def __floordiv__(self, other):
result = self / other
if result.dtype.is_floating_point:
result = result.trunc()
return result
def __rfloordiv__(self, other):
result = other / self
if result.dtype.is_floating_point:
result = result.trunc()
return result
__neg__ = _C._TensorBase.neg
__eq__ = _C._TensorBase.eq
__ne__ = _C._TensorBase.ne
__lt__ = _C._TensorBase.lt
__le__ = _C._TensorBase.le
__gt__ = _C._TensorBase.gt
__ge__ = _C._TensorBase.ge
__abs__ = _C._TensorBase.abs
def __std_mean__(self, dim=None, unbiased=True, keepdim=False):
if dim is None:
return _C._VariableFunctions.std_mean(self, unbiased)
else:
return _C._VariableFunctions.std_mean(self, dim, unbiased, keepdim)
def __var_mean__(self, dim=None, unbiased=True, keepdim=False):
if dim is None:
return _C._VariableFunctions.var_mean(self, unbiased)
else:
return _C._VariableFunctions.var_mean(self, dim, unbiased, keepdim)
def __len__(self):
if self.dim() == 0:
raise TypeError("len() of a 0-d tensor")
return self.shape[0]
def __iter__(self):
# NB: we use 'imap' and not 'map' here, so that in Python 2 we get a
# generator and don't eagerly perform all the indexes. This could
# save us work, and also helps keep trace ordering deterministic
# (e.g., if you zip(*hiddens), the eager map will force all the
# indexes of hiddens[0] before hiddens[1], while the generator
# map will interleave them.)
if self.dim() == 0:
raise TypeError('iteration over a 0-d tensor')
if torch._C._get_tracing_state():
warnings.warn(
'Iterating over a tensor might cause the trace to be incorrect. '
'Passing a tensor of different shape won\'t change the number of '
'iterations executed (and might lead to errors or silently give '
'incorrect results).',
category=RuntimeWarning)
return iter(imap(lambda i: self[i], range(self.size(0))))
def __hash__(self):
return id(self)
def __dir__(self):
tensor_methods = dir(self.__class__)
tensor_methods.remove('volatile') # deprecated
attrs = list(self.__dict__.keys())
keys = tensor_methods + attrs
# property only available dense, cuda tensors
if (not self.is_cuda) or self.is_sparse:
keys.remove("__cuda_array_interface__")
return sorted(keys)
# Numpy array interface, to support `numpy.asarray(tensor) -> ndarray`
__array_priority__ = 1000 # prefer Tensor ops over numpy ones
def __array__(self, dtype=None):
if dtype is None:
return self.numpy()
else:
return self.numpy().astype(dtype, copy=False)
# Wrap Numpy array again in a suitable tensor when done, to support e.g.
# `numpy.sin(tensor) -> tensor` or `numpy.greater(tensor, 0) -> ByteTensor`
def __array_wrap__(self, array):
if array.dtype == bool:
# Workaround, torch has no built-in bool tensor
array = array.astype('uint8')
return torch.from_numpy(array)
def __contains__(self, element):
r"""Check if `element` is present in tensor
Arguments:
element (Tensor or scalar): element to be checked
for presence in current tensor"
"""
if isinstance(element, (torch.Tensor, Number)):
return (element == self).any().item()
return NotImplemented
@property
def __cuda_array_interface__(self):
"""Array view description for cuda tensors.
See:
https://numba.pydata.org/numba-doc/latest/cuda/cuda_array_interface.html
"""
# raise AttributeError for unsupported tensors, so that
# hasattr(cpu_tensor, "__cuda_array_interface__") is False.
if not self.is_cuda:
raise AttributeError(
"Can't get __cuda_array_interface__ on non-CUDA tensor type: %s "
"If CUDA data is required use tensor.cuda() to copy tensor to device memory."
% self.type())
if self.is_sparse:
raise AttributeError(
"Can't get __cuda_array_interface__ on sparse type: %s "
"Use Tensor.to_dense() to convert to a dense tensor first." %
self.type())
# RuntimeError, matching tensor.__array__() behavior.
if self.requires_grad:
raise RuntimeError(
"Can't get __cuda_array_interface__ on Variable that requires grad. "
"If gradients aren't required, use var.detach() to get Variable that doesn't require grad."
)
# CUDA devices are little-endian and tensors are stored in native byte
# order. 1-byte entries are endian-agnostic.
typestr = {
torch.float16: "<f2",
torch.float32: "<f4",
torch.float64: "<f8",
torch.uint8: "|u1",
torch.int8: "|i1",
torch.int16: "<i2",
torch.int32: "<i4",
torch.int64: "<i8",
}[self.dtype]
itemsize = self.storage().element_size()
shape = self.shape
strides = tuple(s * itemsize for s in self.stride())
data = (self.data_ptr(), False) # read-only is false
return dict(typestr=typestr,
shape=shape,
strides=strides,
data=data,
version=0)
__module__ = 'torch'
Tensor = torch.Tensor
entr = _add_docstr(
_special.special_entr, r"""
entr(input, *, out=None) -> Tensor
Computes the entropy on :attr:`input` (as defined below), elementwise.
.. math::
\begin{align}
\text{entr(x)} = \begin{cases}
-x * \ln(x) & x > 0 \\
0 & x = 0.0 \\
-\infty & x < 0
\end{cases}
\end{align}
""" + """
Args:
input (Tensor): the input tensor.
Keyword args:
out (Tensor, optional): the output tensor.
Example::
>>> a = torch.arange(-0.5, 1, 0.5)
>>> a
tensor([-0.5000, 0.0000, 0.5000])
>>> torch.special.entr(a)
tensor([ -inf, 0.0000, 0.3466])
""")
gammaln = _add_docstr(
class Tensor(torch._C._TensorBase):
def __deepcopy__(self, memo):
if not self.is_leaf:
raise RuntimeError(
"Only Tensors created explicitly by the user "
"(graph leaves) support the deepcopy protocol at the moment")
if id(self) in memo:
return memo[id(self)]
with torch.no_grad():
if self.is_sparse:
new_tensor = self.clone()
else:
new_storage = self.storage().__deepcopy__(memo)
new_tensor = self.new()
new_tensor.set_(new_storage, self.storage_offset(),
self.size(), self.stride())
memo[id(self)] = new_tensor
new_tensor.requires_grad = self.requires_grad
return new_tensor
def __reduce_ex__(self, proto):
args = (self.storage(), self.storage_offset(), tuple(self.size()),
self.stride(), self.requires_grad, self._backward_hooks)
return (torch._utils._rebuild_tensor_v2, args)
def __setstate__(self, state):
if not self.is_leaf:
raise RuntimeError(
'__setstate__ can be only called on leaf Tensors')
if len(state) == 4:
# legacy serialization of Tensor
self.set_(*state)
return
elif len(state) == 5:
# legacy serialization of Variable
self.data = state[0]
state = (state[3], state[4], state[2])
self.requires_grad, _, self._backward_hooks = state
def __repr__(self):
# All strings are unicode in Python 3, while we have to encode unicode
# strings in Python2. If we can't, let python decide the best
# characters to replace unicode characters with.
if sys.version_info > (3, ):
return torch._tensor_str._str(self)
else:
if hasattr(sys.stdout, 'encoding'):
return torch._tensor_str._str(self).encode(
sys.stdout.encoding or 'UTF-8', 'replace')
else:
return torch._tensor_str._str(self).encode('UTF-8', 'replace')
def backward(self, gradient=None, retain_graph=None, create_graph=False):
r"""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 them before calling it.
Arguments:
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``.
"""
torch.autograd.backward(self, gradient, retain_graph, create_graph)
def register_hook(self, hook):
r"""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 :attr:`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
"""
if not self.requires_grad:
raise RuntimeError("cannot register a hook on a tensor that "
"doesn't require gradient")
if self._backward_hooks is None:
self._backward_hooks = OrderedDict()
if self.grad_fn is not None:
self.grad_fn._register_hook_dict(self)
handle = hooks.RemovableHandle(self._backward_hooks)
self._backward_hooks[handle.id] = hook
return handle
def reinforce(self, reward):
def trim(str):
return '\n'.join([line.strip() for line in str.split('\n')])
raise RuntimeError(
trim(r"""reinforce() was removed.
Use torch.distributions instead.
See http://pytorch.org/docs/master/distributions.html
Instead of:
probs = policy_network(state)
action = probs.multinomial()
next_state, reward = env.step(action)
action.reinforce(reward)
action.backward()
Use:
probs = policy_network(state)
# NOTE: categorical is equivalent to what used to be called multinomial
m = torch.distributions.Categorical(probs)
action = m.sample()
next_state, reward = env.step(action)
loss = -m.log_prob(action) * reward
loss.backward()
"""))
detach = _add_docstr(
_C._TensorBase.detach, r"""
Returns a new Tensor, detached from the current graph.
The result will never require gradient.
.. note::
Returned Tensor uses the same data tensor as the original one.
In-place modifications on either of them will be seen, and may trigger
errors in correctness checks.
""")
detach_ = _add_docstr(
_C._TensorBase.detach_, r"""
Detaches the Tensor from the graph that created it, making it a leaf.
Views cannot be detached in-place.
""")
def retain_grad(self):
r"""Enables .grad attribute for non-leaf Tensors."""
if self.grad_fn is None: # no-op for leaves
return
if not self.requires_grad:
raise RuntimeError(
"can't retain_grad on Tensor that has requires_grad=False")
if hasattr(self, 'retains_grad'):
return
weak_self = weakref.ref(self)
def retain_grad_hook(grad):
var = weak_self()
if var is None:
return
if var._grad is None:
var._grad = grad.clone()
else:
var._grad = var._grad + grad
self.register_hook(retain_grad_hook)
self.retains_grad = True
def is_pinned(self):
r"""Returns true if this tensor resides in pinned memory"""
storage = self.storage()
return storage.is_pinned() if storage else False
def is_shared(self):
r"""Checks if tensor is in shared memory.
This is always ``True`` for CUDA tensors.
"""
return self.storage().is_shared()
def share_memory_(self):
r"""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.
"""
self.storage().share_memory_()
return self
def __reversed__(self):
r"""Reverses the tensor along dimension 0."""
if self.dim() == 0:
return self
else:
return self.flip(0)
def argmax(self, dim=None, keepdim=False):
r"""See :func:`torch.argmax`"""
return torch.argmax(self, dim, keepdim)
def argmin(self, dim=None, keepdim=False):
r"""See :func:`torch.argmin`"""
return torch.argmin(self, dim, keepdim)
def btrifact(self, info=None, pivot=True):
r"""See :func:`torch.btrifact`
"""
if info is not None:
warnings.warn(
"info option in btrifact is deprecated and will be removed in v0.4, "
"consider using btrifact_with_info instead",
stacklevel=2)
factorization, pivots, _info = super(
Tensor, self).btrifact_with_info(pivot=pivot)
if info.type() != _info.type():
raise ValueError('btrifact expects info to be an IntTensor')
info.resize_as_(_info).copy_(_info)
return factorization, pivots
else:
return super(Tensor, self).btrifact(pivot=pivot)
def stft(self,
n_fft,
hop_length=None,
win_length=None,
window=None,
center=True,
pad_mode='reflect',
normalized=False,
onesided=True):
r"""See :func:`torch.stft`
.. warning::
This function changed signature at version 0.4.1. Calling with
the previous signature may cause error or return incorrect result.
"""
return torch.stft(self, n_fft, hop_length, win_length, window, center,
pad_mode, normalized, onesided)
def resize(self, *sizes):
warnings.warn("non-inplace resize is deprecated")
from torch.autograd._functions import Resize
return Resize.apply(self, sizes)
def resize_as(self, tensor):
warnings.warn("non-inplace resize_as is deprecated")
from torch.autograd._functions import Resize
return Resize.apply(self, tensor.size())
def split(self, split_size, dim=0):
r"""See :func:`torch.split`
"""
if isinstance(split_size, int):
return super(Tensor, self).split(split_size, dim)
else:
return super(Tensor, self).split_with_sizes(split_size, dim)
def index_add(self, dim, index, tensor):
return self.clone().index_add_(dim, index, tensor)
def index_copy(self, dim, index, tensor):
return self.clone().index_copy_(dim, index, tensor)
def index_fill(self, dim, index, value):
return self.clone().index_fill_(dim, index, value)
def scatter(self, dim, index, source):
return self.clone().scatter_(dim, index, source)
def scatter_add(self, dim, index, source):
return self.clone().scatter_add_(dim, index, source)
def masked_scatter(self, mask, tensor):
return self.clone().masked_scatter_(mask, tensor)
def masked_fill(self, mask, value):
return self.clone().masked_fill_(mask, value)
def unique(self, sorted=False, return_inverse=False):
r"""Returns the unique scalar elements of the tensor as a 1-D tensor.
See :func:`torch.unique`
"""
output, inverse_indices = self._unique(sorted=sorted,
return_inverse=return_inverse)
if return_inverse:
return output, inverse_indices
else:
return output
def __rsub__(self, other):
return -self + other
def __rdiv__(self, other):
if self.dtype.is_floating_point:
return self.reciprocal() * other
else:
return (self.double().reciprocal() * other).type_as(self)
__rtruediv__ = __rdiv__
__itruediv__ = _C._TensorBase.__idiv__
__pow__ = _C._TensorBase.pow
def __format__(self, format_spec):
if self.dim() == 0:
return self.item().__format__(format_spec)
return object.__format__(self, format_spec)
def __ipow__(self, other):
raise NotImplementedError("in-place pow not implemented")
def __rpow__(self, other):
return self.new([other])**self
def __floordiv__(self, other):
result = self / other
if result.dtype.is_floating_point:
result = result.trunc()
return result
def __rfloordiv__(self, other):
result = other / self
if result.dtype.is_floating_point:
result = result.trunc()
return result
__neg__ = _C._TensorBase.neg
__eq__ = _C._TensorBase.eq
__ne__ = _C._TensorBase.ne
__lt__ = _C._TensorBase.lt
__le__ = _C._TensorBase.le
__gt__ = _C._TensorBase.gt
__ge__ = _C._TensorBase.ge
__abs__ = _C._TensorBase.abs
def __len__(self):
if self.dim() == 0:
raise TypeError("len() of a 0-d tensor")
return self.shape[0]
def __iter__(self):
# NB: we use 'imap' and not 'map' here, so that in Python 2 we get a
# generator and don't eagerly perform all the indexes. This could
# save us work, and also helps keep trace ordering deterministic
# (e.g., if you zip(*hiddens), the eager map will force all the
# indexes of hiddens[0] before hiddens[1], while the generator
# map will interleave them.)
if self.dim() == 0:
raise TypeError('iteration over a 0-d tensor')
return iter(imap(lambda i: self[i], range(self.size(0))))
def __hash__(self):
return id(self)
def __dir__(self):
tensor_methods = dir(self.__class__)
tensor_methods.remove('volatile') # deprecated
attrs = list(self.__dict__.keys())
keys = tensor_methods + attrs
return sorted(keys)
# Numpy array interface, to support `numpy.asarray(tensor) -> ndarray`
def __array__(self, dtype=None):
if dtype is None:
return self.cpu().numpy()
else:
return self.cpu().numpy().astype(dtype, copy=False)
# Wrap Numpy array again in a suitable tensor when done, to support e.g.
# `numpy.sin(tensor) -> tensor` or `numpy.greater(tensor, 0) -> ByteTensor`
def __array_wrap__(self, array):
if array.dtype == bool:
# Workaround, torch has no built-in bool tensor
array = array.astype('uint8')
return torch.from_numpy(array)
__module__ = 'torch'
__all__ = [
'addmm',
'mm',
'sum',
'softmax',
'log_softmax',
]
addmm = _add_docstr(
_sparse._sparse_addmm, r"""
sparse.addmm(mat, mat1, mat2, *, beta=1., alpha=1.) -> Tensor
This function does exact same thing as :func:`torch.addmm` in the forward,
except that it supports backward for sparse matrix :attr:`mat1`. :attr:`mat1`
need to have `sparse_dim = 2`. Note that the gradients of :attr:`mat1` is a
coalesced sparse tensor.
Args:
mat (Tensor): a dense matrix to be added
mat1 (Tensor): a sparse matrix to be multiplied
mat2 (Tensor): a dense matrix to be multiplied
beta (Number, optional): multiplier for :attr:`mat` (:math:`\beta`)
alpha (Number, optional): multiplier for :math:`mat1 @ mat2` (:math:`\alpha`)
""")
def mm(mat1: Tensor, mat2: Tensor) -> Tensor:
r"""
Performs a matrix multiplication of the sparse matrix :attr:`mat1`
and the (sparse or strided) matrix :attr:`mat2`. Similar to :func:`torch.mm`, If :attr:`mat1` is a
:math:`(n \times m)` tensor, :attr:`mat2` is a :math:`(m \times p)` tensor, out will be a
:math:`(n \times p)` tensor. :attr:`mat1` need to have `sparse_dim = 2`.
