torch/__init__.pyi.in

# ${generated_comment}

from typing import List, Tuple, Optional, Union, Any, ContextManager, Callable, overload
from torch._six import inf

import builtins

# These identifiers are reexported from other modules.  These modules
# are not mypy-clean yet, so in order to use this stub file usefully
# from mypy you will need to specify --follow-imports=silent.
# Not all is lost: these imports still enable IDEs like PyCharm to offer
# autocomplete.
#
# Note: Why does the syntax here look so strange?  Import visibility
# rules in stubs are different from normal Python files!  You must use
# 'from ... import ... as ...' syntax to cause an identifier to be
# exposed (or use a wildcard); regular syntax is not exposed.
from .random import set_rng_state as set_rng_state, get_rng_state as get_rng_state, \
    manual_seed as manual_seed, initial_seed as initial_seed, seed as seed
from ._tensor_str import set_printoptions as set_printoptions
from .functional import *
from .serialization import save as save, load as load
from .autograd import no_grad as no_grad, enable_grad as enable_grad, \
    set_grad_enabled as set_grad_enabled
from . import cuda as cuda
from . import optim as optim

class dtype: ...

class layout: ...

strided : layout = ...

class memory_format: ...

contiguous_format: memory_format = ...

class qscheme: ...

per_tensor_affine: qscheme = ...

# See https://github.com/python/mypy/issues/4146 for why these workarounds
# is necessary
_int = builtins.int
_float = builtins.float

class device:
    type: str
    index: _int

    @overload
    def __init__(self, device: Union[_int, str]) -> None: ...

    @overload
    def __init__(self, type: str, index: _int) -> None: ...

class Size(tuple): ...

class Storage: ...

# See https://github.com/python/mypy/issues/4146 for why these workarounds
# is necessary
_dtype = dtype
_device = device
_qscheme = qscheme
_size = Union[Size, List[_int], Tuple[_int, ...]]

# Meta-type for "numeric" things; matches our docs
Number = Union[builtins.int, builtins.float]

class Generator:
    device: _device = ...

    @overload
    def __init__(self, device: Optional[_device]=None) -> None: ...

    @overload
    def __init__(self, device: Union[_int, str]) -> None: ...

# TODO: One downside of doing it this way, is direct use of
# torch.tensor.Tensor doesn't get type annotations.  Nobody
# should really do that, so maybe this is not so bad.
class Tensor:
    dtype: _dtype = ...
    shape: Size = ...
    device: _device = ...
    requires_grad: bool = ...
    grad: Optional[Tensor] = ...

    ${tensor_method_hints}

    # Manually defined methods from torch/tensor.py
    def backward(self, gradient: Optional[Tensor]=None, retain_graph: Optional[bool]=None, create_graph: bool=False) -> None: ...
    def register_hook(self, hook: Callable) -> Any: ...
    def retain_grad(self) -> None: ...
    def is_pinned(self) -> bool: ...
    def is_shared(self) -> bool: ...
    def share_memory_(self) -> None: ...
    # TODO: fill in the types for these, or otherwise figure out some
    # way to not have to write these out again...
    def nonzero(self, *, as_tuple=True): ...
    def norm(self, p="fro", dim=None, keepdim=False): ...
    def stft(self, n_fft, hop_length=None, win_length=None, window=None,
             center=True, pad_mode='reflect', normalized=False, onesided=True): ...
    def split(self, split_size, dim=0): ...
    def unique(self, sorted=True, return_inverse=False, dim=None): ...
    def unique_consecutive(self, sorted=True, return_inverse=False, return_counts=False, dim=None): ...
    def lu(self, pivot=True, get_infos=False): ...

${function_hints}

${legacy_class_hints}

${dtype_class_hints}

# Pure Python functions defined in torch/__init__.py

def typename(obj) -> str: ...
def is_tensor(obj) -> bool: ...
def is_storage(obj) -> bool: ...
def set_default_tensor_type(type) -> None: ...  # ick, what a bad legacy API
def set_default_dtype(d : _dtype) -> None: ...
def manager_path() -> str: ...
def compiled_with_cxx11_abi() -> bool: ...

# The return value of this function depends on the value of `as_tuple`,
# (similar to `unique`, `lu`, etc.); as such, it is not
# possible to type correctly
def nonzero(input: Tensor, *, out: Optional[Tensor]=None, as_tuple: Optional[bool]=None): ...