Improve Tensor type hints (#28578)

Summary:
I've typed some attributes from ee920b92c4/torch/csrc/autograd/python_variable.cpp (L490) that were not included in the stubs so that MyPy will be aware of them. I made sure to only add those attributes that are mentioned somewhere in the documentation. If there are attributes mentioned in the documentation that are not meant to be part of the public API (or the opposite), please let me know. I've also made sure that attributes that can't be set are typed as read-only properties. If setting `dtype`, `shape`, `device` or `names` directly is not part of the public API, let me know and I'll make them properties as well.

I've also added `__len__`, `__iter__` and `__contains__`, which means MyPy will no longer complain about `len(t)`, `t1 in t2` and `for t1 in t2`.

Shameless plug: I have another typing-related PR here that needs review: https://github.com/pytorch/pytorch/pull/27445

Fixes https://github.com/pytorch/pytorch/issues/28457
Pull Request resolved: https://github.com/pytorch/pytorch/pull/28578

Reviewed By: lerks

Differential Revision: D18113954

Pulled By: fmassa

fbshipit-source-id: 0b69a2966d22054d8d87392f19ec5aa3918773bc

tools/pyi/gen_pyi.py

@@ -275,7 +275,7 @@ def generate_type_hints(fname, decls, is_tensor=False):
                 python_args.append('*')
                 render_kw_only_separator = False
             python_args += ["dtype: _dtype=None",
-                            "layout: layout=strided",
+                            "layout: _layout=strided",
                             "device: Union[_device, str, None]=None",
                             "requires_grad:_bool=False"]
 

torch/__init__.pyi.in

@@ -1,6 +1,6 @@
 # ${generated_comment}
 
-from typing import List, Tuple, Optional, Union, Any, ContextManager, Callable, overload
+from typing import List, Tuple, Optional, Union, Any, ContextManager, Callable, overload, Iterator
 from torch._six import inf
 
 import builtins
@@ -66,6 +66,7 @@ _dtype = dtype
 _device = device
 _qscheme = qscheme
 _size = Union[Size, List[_int], Tuple[_int, ...]]
+_layout = layout
 
 # Meta-type for "numeric" things; matches our docs
 Number = Union[builtins.int, builtins.float, builtins.bool]
@@ -83,16 +84,31 @@ class Generator:
 # 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] = ...
     data: Tensor = ...
+    names: List[str] = ...
+    @property
+    def dtype(self) -> _dtype: ...
+    @property
+    def shape(self) -> Size: ...
+    @property
+    def device(self) -> _device: ...
+    @property
+    def T(self) -> Tensor: ...
+    @property
+    def grad_fn(self) -> Optional[Any]: ...
+    @property
+    def ndim(self) -> _int: ...
+    @property
+    def layout(self) -> _layout: ...
 
     ${tensor_method_hints}
 
     # Manually defined methods from torch/tensor.py
+    def __len__(self) -> _int: ...
+    def __iter__(self) -> Iterator[Tensor]: ...
+    def __contains__(self, item: Union[Tensor, Number]) -> _bool: ...
     def register_hook(self, hook: Callable) -> Any: ...
     def retain_grad(self) -> None: ...
     def is_shared(self) -> _bool: ...