Once your updates are saved, run the program again with go run:

go run main.go
Now, save your main.go file and use go run to run your program again:

go run main.go
Then, once you’ve saved the new Unwrap method, run your program:

go run main.go
Shortcuts

"""TorchScript

This module contains functionality to support the JIT's scripting frontend, notably:
    - torch.jit.script

This is not intended to be imported directly; please use the exposed
functionalities in `torch.jit`.
"""
import functools
import collections
import enum
import inspect
import copy
import pickle
import warnings
from typing import Any, Dict, List, Tuple, Union, Callable


import torch
import torch._jit_internal as _jit_internal
from torch.utils import set_module
from torch.jit._recursive import ScriptMethodStub, wrap_cpp_module, infer_methods_to_compile, _compile_and_register_class
from torch.nn import Module
from torch.jit._state import _enabled
from torch.jit._builtins import _register_builtin
from torch._six import with_metaclass
from torch.jit.frontend import get_jit_def, get_default_args, get_jit_class_def
from torch._jit_internal import _qualified_name
from torch.jit._fuser import _graph_for
from torch.jit._state import (
    _try_get_jit_cached_function,
    _try_get_jit_cached_overloads,
    _set_jit_function_cache,
    _set_jit_overload_cache,
)
from torch.overrides import (
    has_torch_function, has_torch_function_unary, has_torch_function_variadic)
from torch.package import PackageExporter, PackageImporter
from ._serialization import validate_map_location

from torch.jit._monkeytype_config import (
    monkeytype_trace,
    JitTypeTraceConfig ,
    JitTypeTraceStore
)
from torch._classes import classes

type_trace_db = JitTypeTraceStore()  # DB to hold all call traces from MonkeyType

torch._C.ScriptMethod.graph_for = _graph_for  # type: ignore[attr-defined]
torch._C.ScriptFunction.graph_for = _graph_for  # type: ignore[attr-defined]
ScriptFunction = torch._C.ScriptFunction
ScriptFunction.__doc__ = """
Functionally equivalent to a :class:`ScriptModule`, but represents a single
function and does not have any attributes or Parameters.
"""
set_module(ScriptFunction, "torch.jit")


if _enabled:
    Attribute = collections.namedtuple("Attribute", ["value", "type"])
else:

[docs]    def Attribute(value, type):  # type: ignore[no-redef]
        return value

Attribute.__doc__ = """
    This method is a pass-through function that returns `value`, mostly
    used to indicate to the TorchScript compiler that the left-hand side
    expression is a class instance attribute with type of `type`. Note that
    `torch.jit.Attribute` should only be used in `__init__` method of `nn.Module`
    subclasses.

    Though TorchScript can infer correct type for most Python expressions, there are some cases where
    type inference can be wrong, including:

    - Empty containers like `[]` and `{}`, which TorchScript assumes to be container of `Tensor`
    - Optional types like `Optional[T]` but assigned a valid value of type `T`, TorchScript would assume
      it is type `T` rather than `Optional[T]`

    In eager mode, it is simply a pass-through function that returns `value`
    without other implications.

    Example:

    .. testcode::

        import torch
        from typing import Dict

        class AttributeModule(torch.nn.Module):
            def __init__(self):
                super(M, self).__init__()
                self.foo = torch.jit.Attribute(0.1, float)

                # we should be able to use self.foo as a float here
                assert 0.0 < self.foo

                self.names_ages = torch.jit.Attribute({}, Dict[str, int])
                self.names_ages["someone"] = 20
                assert isinstance(self.names_ages["someone"], int)

        m = AttributeModule()
        # m will contain two attributes
        # 1. foo of type float
        # 2. names_ages of type Dict[str, int]

    .. testcleanup::

        del AttributeModule
        del m

    Args:
        value: An initial value to be assigned to attribute.
        type: A Python type

    Returns:
        Returns `value`
"""

def _get_type_trace_db():
    # This is a private API. Use of this for external purposes is discouraged.
    return type_trace_db

# Gets a function from the name of a method on a type
def _get_function_from_type(cls, name):
    return getattr(cls, name, None)


# ScriptClasses must be new-style classes because we construct them using their
# __new__ method.
def _is_new_style_class(cls):
    if hasattr(cls, "__class__"):
        return "__dict__" in dir(cls) or hasattr(cls, "__slots__")


# These OrderedDictWrapper classes replace the actual OrderedDicts in
# module with versions that get/set properties inside of Module.
# This allows us to reuse most of nn.Module while still storing the
# data in C++.
# Each OrderedDict needs to support:
#  x not in view
#  x in view
#  view[name] = ...
#  view.values()
#  del view[name]
#  view.items()
#  view.keys()
#  len(view)


class OrderedDictWrapper(object):
    def __init__(self, _c):
        self._c = _c

    def keys(self):
        return [k for k, v in self.items()]

    def values(self):
        return [v for k, v in self.items()]

    def __len__(self):
        return len(self.values())

    def __delitem__(self, k):
        raise RuntimeError("cannot delete methods or parameters of a script module")

    def items(self):
        return self._c.items()

    def __setitem__(self, k, v):
        if k not in self:
            raise RuntimeError(
                "Can't add a new parameter after ScriptModule construction."
                " Tried to add '{}".format(k)
            )
        self._c.setattr(k, v)

    def __contains__(self, k):
        return self._c.contains(k)

    def __getitem__(self, k):
        if k not in self:
            raise KeyError(k)
        return self._c.getattr(k)


class OrderedModuleDict(OrderedDictWrapper):
    def __init__(self, module, python_dict):
        super(OrderedModuleDict, self).__init__(torch._C.ModuleDict(module))
        # contains _both_ script modules and non-script python-only modules

        # because script modules are subclassed in python and the
        # C++ Module class will not hold references to them,
        # to ensure that you always get the same python value here
        # we store it in the python dict as well
        self._python_modules = python_dict

    def items(self):
        r = self._python_modules.items()
        return r

    def __contains__(self, k):
        return k in self._python_modules

    def __setitem__(self, k, v):
        # Cases where sub-module can be re-assigned after ScriptModule construction
        # 1. If the attr is an module interface type, it's guaranteed that the module is
        #    not inlined in the graph, so it's safe to swap a new ScriptModule in.
        # 2. if the new value if a ScriptModule with the same JIT type, IR won't change
        #    and it's legit to swap a new module in.
        # In these two cases we allow swapping a new scripted module and update the
        # corresponding python module dict to keep sync.
        # Note: the value to be swapped in has to be ScriptModule instead of nn.Module,
        # otherwise it's illegal and we throw error.
        if isinstance(v, ScriptModule):
            self._c.setattr(k, v)
            self._python_modules[k] = v
        else:
            raise RuntimeError(
                "Cannot re-assign modules in a ScriptModule with non-scripted "
                "module, tried to replace existing module '{}': {}".format(k, v)
            )

    def __getitem__(self, k):
        return self._python_modules[k]


# For each user-defined class that subclasses ScriptModule, this meta-class:
# (1) finds all the methods annotated with @script_method in a ScriptModule and
#     removes them from the class attributes
# (2) puts a wrapper around the class's __init__ method to recursively compile
#     all of the script_methods with the module after the original __init__ has
#     run. This has to occur after the user-defined __init__ so that submodules and
#     parameters are initialized _before_ the script compiler resolve references to
#     `self.param` or `self.module`.
class ScriptMeta(type):
    def __init__(cls, name, bases, attrs):  # noqa: B902
        # Aggregate all the ScriptMethods and constants from superclasses
        cls._methods: Dict[str, Any] = {}
        cls._constants_set = set(getattr(cls, "__constants__", ()))
        for base in reversed(bases):
            for k, v in getattr(base, "_methods", {}).items():
                cls._methods[k] = v
            base_constants = getattr(base, "_constants_set", set())
            cls._constants_set = cls._constants_set.union(base_constants)

        # find all the script methods of the current class
        for k, v in sorted(attrs.items()):
            if isinstance(v, ScriptMethodStub):
                delattr(cls, k)
                cls._methods[v.original_method.__name__] = v

        if getattr(cls, "_disable_script_meta", False):
            # We leave built-in ScriptModule types alone, since this metaclass
            # is only for compiling user classes that inherit from
            # ScriptModule.
            return super(ScriptMeta, cls).__init__(name, bases, attrs)

        original_init = getattr(cls, "__init__", lambda self: None)

        @functools.wraps(original_init)
        def init_then_script(self, *args, **kwargs):
            num_methods = len(cls._methods)
            original_init(self, *args, **kwargs)
            added_methods_in_init = len(cls._methods) > num_methods

            if type(self) == cls:

                def make_stubs(module):
                    cls = type(module)
                    if hasattr(cls, "_methods"):
                        return [v for k, v in sorted(cls._methods.items())]
                    else:
                        return infer_methods_to_compile(module)

                self.__dict__[
                    "_actual_script_module"
                ] = torch.jit._recursive.create_script_module(self, make_stubs, share_types=not added_methods_in_init)

                # Delete the Python attributes that now shadow the ScriptModule
                # ones, so that __getattr__ and __setattr__ will properly find
                # the scripted versions.
                concrete_type = self._actual_script_module._concrete_type
                for name in concrete_type.get_attributes():
                    delattr(self, name)
                for name, _ in concrete_type.get_modules():
                    delattr(self, name)
                for name in ("_parameters", "_buffers", "_modules"):
                    delattr(self, name)

        cls.__init__ = init_then_script  # type: ignore[misc]
        super(ScriptMeta, cls).__init__(name, bases, attrs)


class _CachedForward(object):
    def __get__(self, obj, cls):
        return self.__getattr__("forward")  # type: ignore[attr-defined]


class ScriptWarning(Warning):
    pass


def script_method(fn):
    if not _enabled:
        return fn
    # NOTE: we need to traverse two frames here because the meta-class frame
    # for ScriptModule will be present, as opposed to invoking @script on a
    # a function or invoking define() on a CompilationUnit.
    # The stack will look like:
    #
    # 0. createResolutionCallback()
    # 1. script_method()
    # 2. ScriptModule metaclass frame
    # 3. Surrounding scope
    #
    # createResolutionCallback internally adds 1 to get us to the scope of this
    # function (the calling function). Adding 2 gets us to the proper surrounding scope.
    _rcb = _jit_internal.createResolutionCallbackFromFrame(frames_up=2)
    ast = get_jit_def(fn, fn.__name__, self_name="ScriptModule")
    return ScriptMethodStub(_rcb, ast, fn)


class ConstMap:
    def __init__(self, const_mapping):
        self.const_mapping = const_mapping

    def __getattr__(self, attr):
        return self.const_mapping[attr]


def unpackage_script_module(importer: PackageImporter, script_module_id: str) -> torch.nn.Module:
    """
    Called by ``torch.package.PackageImporter``'s Pickler's ``persistent_load`` function.
    Performs work of loading and returning a ScriptModule from a ``torch.package`` archive.
    """
    if not isinstance(importer.zip_reader, torch._C.PyTorchFileReader):
        raise RuntimeError(
            "Loading ScriptObjects from a PackageImporter created from a "
            "directory is not supported. Use a package archive file instead."
        )
    cu = torch._C.CompilationUnit()
    cpp_module = torch._C._import_ir_module_from_package(
        cu,
        importer.zip_reader,
        importer.storage_context,
        validate_map_location(importer.last_map_location),
        script_module_id,
    )
    return wrap_cpp_module(cpp_module)


if _enabled:
    _magic_methods = [
        "__iter__",
        "__len__",
        "__neg__",
        "__mul__",
        "__contains__",
        "__add__",
        "__sub__",
        "__pow__",
        "__truediv__",
        "__mod__",
        "__ne__",
        "__eq__",
        "__lt__",
        "__gt__",
        "__le__",
        "__ge__",
        "__and__",
        "__or__",
        "__xor__",
        "__getitem__",
        "__setitem__",
        "__call__",
        "__int__",
        "__float__",
        "__bool__",
        "__str__",
        "__enter__",
        "__exit__",
    ]

    class RecursiveScriptClass(object):
        """
        An analogue of RecursiveScriptModule for regular objects that are not modules.
        This class is a wrapper around a torch._C.ScriptObject that represents an instance
        of a TorchScript class and allows it to be used in Python.

        Attributes:
            _c [torch._C.ScriptObject]: The C++ object to which attribute lookups and method
                calls are forwarded.
            _props [Dict[str, property]]: A dictionary of properties fetched from self._c and
                exposed on this wrppaer.
        """
        def __init__(self, cpp_class):
            super(RecursiveScriptClass, self).__init__()
            self.__dict__["_initializing"] = True
            self._c = cpp_class

            # Add wrapped object's properties to this class instance.
            self._props = {prop.name: property(prop.getter, prop.setter) for prop in self._c._properties()}

            self.__dict__["_initializing"] = False

        def __getattr__(self, attr):
            if "_initializing" in self.__dict__ and self.__dict__["_initializing"]:
                return super(RecursiveScriptClass, self).__getattr__(attr)  # type: ignore[misc]

            if attr in self._props:
                return self._props[attr].fget()

            return getattr(self._c, attr)

        def __setattr__(self, attr, value):
            if "_initializing" in self.__dict__ and self.__dict__["_initializing"]:
                return super(RecursiveScriptClass, self).__setattr__(attr, value)

            if attr in self._props:
                return self._props[attr].fset(value)

            setattr(self._c, attr, value)

        # Delegate calls to magic methods like __len__ to the C++ module backing the
        # RecursiveScriptClass.
        def forward_magic_method(self, method_name, *args, **kwargs):
            if not self._c._has_method(method_name):
                raise TypeError()

            self_method = self.__getattr__(method_name)
            return self_method(*args, **kwargs)

        def __getstate__(self):
            raise pickle.PickleError("ScriptClasses cannot be pickled")

        def __iadd__(self, other):
            if self._c._has_method("__iadd__"):
                return self.forward_magic_method("__iadd__", other)
            else:
                return self.forward_magic_method("__add__", other)


    for method_name in _magic_methods:
        def method_template(self, *args, **kwargs):
            return self.forward_magic_method(method_name, *args, **kwargs)

        setattr(RecursiveScriptClass, method_name, method_template)

    # this is a Python 'non-data descriptor' that causes the first access
    # to ScriptModule's forward to look up the forward method and stash
    # it in the objects dict. Due to the standard rules for attribute lookup,
    # subsequent lookups will just directly return the previously looked up method.
    # This is necessary because nn.Module defines forward as a method. If we
    # did nothing, __getattr__ would not be called. Instead we'd get nn.Module.forward
    # which always throws an exception.

    class ScriptModule(with_metaclass(ScriptMeta, Module)):  # type: ignore[misc]
        r"""
        A wrapper around C++ ``torch::jit::Module``. ``ScriptModule``\s
        contain methods, attributes, parameters, and
        constants. These can be accessed the same way as on a normal ``nn.Module``.
        """
        __jit_unused_properties__ = ['code', 'code_with_constants', 'graph', 'inlined_graph', 'original_name']

        def __init__(self):
            super(ScriptModule, self).__init__()

        forward = _CachedForward()

        def __getattr__(self, attr):
            if "_actual_script_module" not in self.__dict__:
                return super(ScriptModule, self).__getattr__(attr)
            return getattr(self._actual_script_module, attr)

        def __setattr__(self, attr, value):
            if "_actual_script_module" not in self.__dict__:
                # Unwrap torch.jit.Attribute into a regular setattr + record
                # the provided type in __annotations__.
                #
                # This ensures that if we use the attr again in `__init__`, it
                # will look like the actual value, not an instance of Attribute.
                if isinstance(value, Attribute):
                    # NB: Ensure that we set __annotations__ on the specific
                    # class in question, and not on a superclass (which would
                    # be wrong wrong wrong!).
                    # See also https://github.com/pytorch/pytorch/issues/39463
                    if "__annotations__" not in self.__class__.__dict__:
                        self.__class__.__annotations__ = {}
                    self.__annotations__[attr] = value.type
                    value = value.value
                return super(ScriptModule, self).__setattr__(attr, value)

            setattr(self._actual_script_module, attr, value)

        def define(self, src):
            if "_actual_script_module" in self.__dict__:
                # If we have completed initialization, just defer to the
                # backing RecursiveScriptModule to eagerly compile the provided
                # source.
                return self._actual_script_module.define(src)

            # Otherwise, we are still in the object's __init__.
            # In that case, add `src` as a stub to be compiled.
            #
            # We use frames_up=1 to get to the proper surrounding scope. The stack
            # will look like:
            # 0. createResolutionCallback
            # 1. define()
            # 2. surrounding scope.
            #
            # createResolutionCallback internally adds 1 to get us to our frame, then
            # we add 1 to get to the proper surrounding scope.
            rcb = _jit_internal.createResolutionCallbackFromFrame(frames_up=1)
            ast = torch._C._parse_source_def(src)
            self._methods[ast.name().name] = ScriptMethodStub(rcb, ast, None)

        def _replicate_for_data_parallel(self):
            return self._actual_script_module._replicate_for_data_parallel()

        def __reduce_package__(self, exporter: PackageExporter):
            """
            Called by ``torch.package.PackageExporter``'s Pickler's ``persistent_id`` when
            saving TorchScript objects. Performs act of saving a ScriptModule inside of
            a ``torch.package`` archive.

            Returns method to load the ScriptModule from a ``torch.package.PackageImporter``'s
            Pickler's ``persistent_load`` function.
            """
            script_module_id = exporter.get_unique_id()
            exporter.script_module_serializer.serialize(self._c, int(script_module_id))
            return (unpackage_script_module, (script_module_id,))

    class RecursiveScriptModule(ScriptModule):
        # XXX: RecursiveScriptModule inherits from ScriptModule for the sole
        # reason that it retains the existing isinstance(ScriptModule)
        # behavior.
        r"""
        The core data structure in TorchScript is the ``ScriptModule``. It is an
        analogue of torch's ``nn.Module`` and represents an entire model as a tree of
        submodules. Like normal modules, each individual module in a ``ScriptModule`` can
        have submodules, parameters, and methods. In ``nn.Module``\s methods are implemented
        as Python functions, but in ``ScriptModule``\s methods are implemented as
        TorchScript functions, a statically-typed subset of Python that contains all
        of PyTorch's built-in Tensor operations. This difference allows your
        ``ScriptModule``\s code to run without the need for a Python interpreter.

        ``ScriptModule``\s should not be created manually, instead use
        either :func:`tracing <torch.jit.trace>` or :func:`scripting <torch.jit.script>`.
        Tracing and scripting can be applied incrementally and :ref:`composed as necessary <Types>`.

        * Tracing records the tensor operations as executed with a set of example inputs and uses these
          operations to construct a computation graph. You can use the full dynamic behavior of Python with tracing,
          but values other than Tensors and control flow aren't captured in the graph.

        * Scripting inspects the Python code of the model
          and compiles it to TorchScript. Scripting allows the use of many `types`_ of values and supports dynamic control flow.
          Many, but not all features of Python are supported by the compiler, so changes to the source code may be necessary.
        """
        _disable_script_meta = True

        def __init__(self, cpp_module):
            self.__dict__["_initializing"] = True
            self._c = cpp_module
            super(RecursiveScriptModule, self).__init__()
            # Delete the 'training' attribute set up by `Module.__init__`. It
            # will get set on the underlying cpp module, so we delete it here
            # to avoid this version shadowing the cpp module version.
            delattr(self, "training")

        @staticmethod
        def _construct(cpp_module, init_fn):
            """
            Construct a RecursiveScriptModule that's ready for use. PyTorch
            code should use this to construct a RecursiveScriptModule instead
            of instead of calling `__init__` directly, as it makes sure the
            object is properly finalized (and in the future, we may take
            control of how the RecursiveScriptModule instance is created).

            Args:
                cpp_module:  The C++ Module that will hold the actual state of
                             this RecursiveScriptModule instance.
                init_fn:  Lambda that initializes the RecursiveScriptModule passed to it.
            """
            script_module = RecursiveScriptModule(cpp_module)
            init_fn(script_module)

            # Finalize the ScriptModule: replace the nn.Module state with our
            # custom implementations and flip the _initializing bit.
            RecursiveScriptModule._finalize_scriptmodule(script_module)
            return script_module

        @staticmethod
        def _finalize_scriptmodule(script_module):
            script_module._parameters = OrderedDictWrapper(
                torch._C.ParameterDict(script_module._c)
            )
            script_module._buffers = OrderedDictWrapper(
                torch._C.BufferDict(script_module._c)
            )
            script_module._modules = OrderedModuleDict(
                script_module._c, script_module._modules
            )
            script_module._initializing = False

        def _reconstruct(self, cpp_module):
            """
            Re-construct an instance of RecursiveScriptModule using an instance of a C++ module.

            Args:
                cpp_module: The C++ module that this RecursiveScriptModule will be rebuilt around.
            """
            self.__init__(cpp_module)  # type: ignore[misc]

            # Copy the concrete type from the C++ module to this ScriptModule.
            self._concrete_type = torch._C.ConcreteModuleType.from_jit_type(
                self._c._type()
            )

            # Copy submodules from the C++ module to this ScriptModule.
            modules = {}
            for name, cpp_module in torch._C.ModuleDict(self._c).items():
                modules[name] = wrap_cpp_module(cpp_module)
            self._modules = OrderedModuleDict(self._c, modules)

            # Copy parameters and buffers.
            self._parameters = OrderedDictWrapper(torch._C.ParameterDict(self._c))
            self._buffers = OrderedDictWrapper(torch._C.BufferDict(self._c))

            # Get rid of the functions from the old C++ module.
            self.__dict__ = {
                k: v
                for k, v in self.__dict__.items()
                if not isinstance(v, torch._C.ScriptMethod)
            }
            self.__dict__["_initializing"] = False

        @property
        def graph(self):
            r"""
            Returns a string representation of the internal graph for the
            ``forward`` method. See :ref:`interpreting-graphs` for details.
            """
            return self._c._get_method("forward").graph

        @property
        def inlined_graph(self):
            r"""
            Returns a string representation of the internal graph for the
            ``forward`` method. This graph will be preprocessed to inline all function and method calls.
            See :ref:`interpreting-graphs` for details.
            """
            return self.forward.inlined_graph

        @property
        def code(self):
            r"""
            Returns a pretty-printed representation (as valid Python syntax) of
            the internal graph for the ``forward`` method. See
            :ref:`inspecting-code` for details.
            """
            return self.forward.code

        @property
        def code_with_constants(self):
            r"""
            Returns a tuple of:

            [0] a pretty-printed representation (as valid Python syntax) of
            the internal graph for the ``forward`` method. See `code`.
            [1] a ConstMap following the CONSTANT.cN format of the output in [0].
            The indices in the [0] output are keys to the underlying constant's values.

            See :ref:`inspecting-code` for details.
            """
            r = self.forward.code_with_constants
            return (r[0], ConstMap(r[1]))

        def save(self, f, **kwargs):
            r"""
            save(f, _extra_files={})

            See :func:`torch.jit.save <torch.jit.save>` for details.
            """
            return self._c.save(str(f), **kwargs)

        def _save_for_lite_interpreter(self, *args, **kwargs):
            r"""
            _save_for_lite_interpreter(f)

            Add (or update) the bytecode session to the script model. The updated model is used
            in lite interpreter for mobile applications.

            Args:
                f: a string containing a file name.
                _extra_files: Map from filename to contents which will be stored as part of 'f'.

            """
            return self._c._save_for_mobile(*args, **kwargs)

        def _save_to_buffer_for_lite_interpreter(self, *args, **kwargs):
            return self._c._save_to_buffer_for_mobile(*args, **kwargs)

        def save_to_buffer(self, *args, **kwargs):
            return self._c.save_to_buffer(*args, **kwargs)

        def get_debug_state(self, *args, **kwargs):
            return self._c.get_debug_state()

        def extra_repr(self):
            return "original_name={}".format(self.original_name)

        def graph_for(self, *args, **kwargs):
            return self.forward.graph_for(*args, **kwargs)

        @property
        def original_name(self):
            if type(self) == str(self._c._type().name()):
                return ""
            return str(self._c._type().name())

        def define(self, src):
            # We use frames_up=1 to get to the proper surrounding scope. The stack
            # will look like:
            # 0. createResolutionCallback
            # 1. define()
            # 2. surrounding scope.
            #
            # createResolutionCallback internally adds 1 to get us to our frame, then
            # we add 1 to get to the proper surrounding scope.
            rcb = _jit_internal.createResolutionCallbackFromFrame(frames_up=1)
            self._c._define(self._concrete_type, src, rcb)

        def __getattr__(self, attr):
            if "_initializing" not in self.__dict__:
                raise RuntimeError(
                    "ScriptModule has not been initialized, did you forget to call super's init?"
                )

            if self._initializing:
                return super(RecursiveScriptModule, self).__getattr__(attr)

            # _modules check is before hasattr since modules are included as attributes in _c,
            # but we want to get the python wrapper from _modules instead of the raw _c object.
            if attr in self._modules:
                return self._modules[attr]
            elif self._c.hasattr(attr):
                return self._c.getattr(attr)
            elif self._c._has_method(attr):
                script_method = self._c._get_method(attr)
                # cache method so future calls do not go through __getattr__
                # to improve invocation performance
                self.__dict__[attr] = script_method
                return script_method

            return super(RecursiveScriptModule, self).__getattr__(attr)

        def __setattr__(self, attr, value):
            if self._initializing:
                return super(RecursiveScriptModule, self).__setattr__(attr, value)

            if attr in self._modules:
                self._modules[attr] = value
            elif self._c.hasattr(attr):
                self._c.setattr(attr, value)
            elif (
                hasattr(self, "_concrete_type")
                and attr in self._concrete_type.get_constants().keys()
            ):
                # TODO: we don't have _concrete_type set after load(), and in general we lose constant information.
                # We should encode constants as class type attributes (or something) so it persists across save/load.
                raise AttributeError(
                    "Cannot mutate TorchScript constant value: '{}'. Value: '{}'".format(
                        attr, value
                    )
                )
            else:
                # We allow setting Python attributes on the ScriptModule, for
                # when people want to stash some convenience info on it.
                # TODO: it's possible that the following is confusing:
                #   s = torch.jit.script(...)
                #   s.python_attr = ...
                #   s.save()   <--- this doesn't have `python_attr`
                # It's fairly trivial to save enough info to warn in this case.
                return super(RecursiveScriptModule, self).__setattr__(attr, value)

        def __copy__(self):
            return torch.jit._recursive.wrap_cpp_module(copy.copy(self._c))

        def __deepcopy__(self, memo):
            return torch.jit._recursive.wrap_cpp_module(copy.deepcopy(self._c, memo))

        # Python magic methods do method lookups on an object's class type, instead of looking up
        # the method defines on the class instance. In order to continue to expose the magic methods
        # of builtin-containers (ModuleList, Sequential, ModuleDict) to Python, we
        # define magic methods here as a shim to the correct attribute.
        def forward_magic_method(self, method_name, *args, **kwargs):
            self_method = getattr(self, method_name)
            if getattr(self_method, "__func__", None) == getattr(
                RecursiveScriptModule, method_name
            ):
                raise NotImplementedError()
            return self_method(*args, **kwargs)

        def __iter__(self):
            return self.forward_magic_method("__iter__")

        def __getitem__(self, idx):
            return self.forward_magic_method("__getitem__", idx)

        def __len__(self):
            return self.forward_magic_method("__len__")

        def __contains__(self, key):
            return self.forward_magic_method("__contains__", key)

        # dir is defined by the base nn.Module, so instead of throwing if
        # it is not overridden, we call into the nn.Module __dir__ method
        def __dir__(self):
            self_method = self.__dir__
            if self_method.__func__ == _get_function_from_type(  # type: ignore[attr-defined]
                RecursiveScriptModule, "__dir__"
            ):
                return super(RecursiveScriptModule, self).__dir__()
            return self_method()

        # to resolve bool(value), Python looks if __bool__ is defined then __iter__
        # is defined then returns true for classes. Since __iter__() on this
        # class throws if it isn't overridden, we define __bool__ to preserve default behavior
        def __bool__(self):
            self_method = self.__bool__
            if self_method.__func__ == _get_function_from_type(  # type: ignore[attr-defined]
                RecursiveScriptModule, "__bool__"
            ):
                return True
            return self_method()

        def _replicate_for_data_parallel(self):
            # we have to initialize ScriptModule properly so that
            # it works with pybind11
            def init_fn(script_module):
                # Don't do anything here, we'll initialize the ScriptModule below
                return

            return RecursiveScriptModule._construct(
                self._c._replicate_for_data_parallel(), init_fn
            )

    # Need to copy all RecursiveScriptModule methods to ScriptModule.
    #
    # This is because `super(MyScriptModule, self).foo()` does not use
    # `__getattr__` to look up `foo`. So we need to make each method available on
    # the ScriptModule manually.
    for name, item in RecursiveScriptModule.__dict__.items():
        if not callable(item) and not isinstance(item, property):
            continue
        if name.startswith("__") or hasattr(ScriptModule, name):
            continue
        # We can copy over the implementation wholesale because besides the
        # `super()` thing above, ScriptModule behaves exactly like
        # RecursiveScriptModule
        setattr(ScriptModule, name, item)

    def _get_methods(cls):
        import inspect

        # In Python 3 unbound methods are functions, but in Python 2 they are methods
        return inspect.getmembers(
            cls, predicate=lambda x: inspect.isfunction(x) or inspect.ismethod(x)
        )

    _compiled_methods_allowlist = {
        "forward",
        "register_buffer",
        "register_parameter",
        "add_module",
        "_apply",
        "apply",
        "cuda",
        "cpu",
        "to",
        "type",
        "float",
        "double",
        "half",
        "state_dict",
        "_save_to_state_dict",
        "load_state_dict",
        "_load_from_state_dict",
        "_named_members",
        "parameters",
        "named_parameters",
        "buffers",
        "named_buffers",
        "children",
        "named_children",
        "modules",
        "named_modules",
        "zero_grad",
        "share_memory",
        "_get_name",
        "extra_repr",
        "_slow_forward",
        "_tracing_name",
        "eval",
        "train",
        "get_extra_state",
        "set_extra_state"
    }

    def _make_fail(name):
        def fail(self, *args, **kwargs):
            raise RuntimeError(name + " is not supported on ScriptModules")

        return fail

    for name, method in _get_methods(torch.nn.Module):
        if name.startswith("__"):
            continue
        if (
            name not in RecursiveScriptModule.__dict__
            and name not in _compiled_methods_allowlist
        ):
            setattr(RecursiveScriptModule, method.__name__, _make_fail(name))


else:
    # TODO MAKE SURE THAT DISABLING WORKS
    class RecursiveScriptClass(object):  # type: ignore[no-redef]
        def __init__(self):
            super().__init__()

[docs]    class ScriptModule(torch.nn.Module):  # type: ignore[no-redef]
        def __init__(self, arg=None):
            super().__init__()

    class RecursiveScriptModule(ScriptModule):  # type: ignore[no-redef]
        def __init__(self, arg=None):
            super().__init__()

def call_prepare_scriptable_func_impl(obj, memo):
    if not isinstance(obj, torch.nn.Module):
        return obj

    obj_id = id(obj)

    # If obj_id is in memo, obj has already been prepared or is being
    # prepared in another call up the stack.
    if obj_id in memo:
        return memo[id(obj)]

    obj = obj.__prepare_scriptable__() if hasattr(obj, '__prepare_scriptable__') else obj  # type: ignore[operator]
    # Record obj in memo to avoid infinite recursion in the case of cycles in the module
    # hierarchy when recursing below.
    memo[obj_id] = obj

    new_obj_dict = {}

    for name, sub_module in obj.__dict__.items():
        if name == '_modules':
            for k, v in sub_module.items():
                sub_module[k] = call_prepare_scriptable_func_impl(v, memo)
            new_obj_dict[name] = sub_module
        elif isinstance(sub_module, torch.nn.Module) and not isinstance(sub_module, ScriptModule):
            new_obj_dict[name] = call_prepare_scriptable_func_impl(sub_module, memo)
        else:
            new_obj_dict[name] = sub_module

    for k, v in new_obj_dict.items():
        obj.__dict__[name] = v

    return obj


def call_prepare_scriptable_func(obj):
    memo: Dict[int, torch.nn.Module] = {}
    return call_prepare_scriptable_func_impl(obj, memo)

def create_script_dict(obj):
    """
    Create a ``torch._C.ScriptDict`` instance with the data from ``obj``.

    Args:
        obj (dict): The Python dictionary that is used to initialize the ``ScriptDict``
                    returned by this function.

    Returns:
        An instance of ``torch._C.ScriptDict`` that has the same data as ``obj``
        and can be passed between Python and TorchScript with reference semantics and
        zero copy overhead.
    """
    return torch._C.ScriptDict(obj)  # type: ignore[attr-defined]


def create_script_list(obj, type_hint=None):
    """
    Create a ``torch._C.ScriptList`` instance with the data from ``obj``.
    Args:
        obj (dict): The Python list that is used to initialize the ``ScriptList``
                    returned by this function.
    Returns:
        An instance of ``torch._C.ScriptList`` that has the same data as ``obj``
        and can be passed between Python and TorchScript with reference semantics and
        zero copy overhead.
    """
    return torch._C.ScriptList(obj)  # type: ignore[attr-defined]


[docs]def script(obj, optimize=None, _frames_up=0, _rcb=None,
           example_inputs: Union[List[Tuple], Dict[Callable, List[Tuple]], None] = None):
    r"""
    Scripting a function or ``nn.Module`` will inspect the source code, compile
    it as TorchScript code using the TorchScript compiler, and return a :class:`ScriptModule` or
    :class:`ScriptFunction`. TorchScript itself is a subset of the Python language, so not all
    features in Python work, but we provide enough functionality to compute on
    tensors and do control-dependent operations. For a complete guide, see the
    :ref:`language-reference`.

    Scripting a dictionary or list copies the data inside it into a TorchScript instance than can be
    subsequently passed by reference between Python and TorchScript with zero copy overhead.

    ``torch.jit.script`` can be used as a function for modules, functions, dictionaries and lists
     and as a decorator ``@torch.jit.script`` for :ref:`torchscript-classes` and functions.

    Args:
        obj (callable, class, or ``nn.Module``):  The ``nn.Module``, function, class type,
                                                  dictionary, or list to compile.
        example_inputs (Union[List[Tuple], Dict[Callable, List[Tuple]], None]): Provide example inputs
            to annotate the arguments for a function or ``nn.Module``.

    Returns:
        If ``obj`` is ``nn.Module``, ``script`` returns
        a :class:`ScriptModule` object. The returned :class:`ScriptModule` will
        have the same set of sub-modules and parameters as the
        original ``nn.Module``. If ``obj`` is a standalone function,
        a :class:`ScriptFunction` will be returned. If ``obj`` is a ``dict``, then
        ``script`` returns an instance of `torch._C.ScriptDict`. If ``obj`` is a ``list``,
        then ``script`` returns an instance of `torch._C.ScriptList`.

    **Scripting a function**
        The ``@torch.jit.script`` decorator will construct a :class:`ScriptFunction`
        by compiling the body of the function.

        Example (scripting a function):

        .. testcode::

            import torch

            @torch.jit.script
            def foo(x, y):
                if x.max() > y.max():
                    r = x
                else:
                    r = y
                return r

            print(type(foo))  # torch.jit.ScriptFunction

            # See the compiled graph as Python code
            print(foo.code)

            # Call the function using the TorchScript interpreter
            foo(torch.ones(2, 2), torch.ones(2, 2))

        .. testoutput::
            :hide:

            ...

    ****Scripting a function using example_inputs**
        Example inputs can be used to annotate a function arguments.

        Example (annotating a function before scripting):

        .. testcode::

            import torch

            def test_sum(a, b):
                return a + b

            # Annotate the arguments to be int
            scripted_fn = torch.jit.script(test_sum, example_inputs=[(3, 4)])

            print(type(scripted_fn))  # torch.jit.ScriptFunction

            # See the compiled graph as Python code
            print(scripted_fn.code)

            # Call the function using the TorchScript interpreter
            scripted_fn(20, 100)

        .. testoutput::
            :hide:

            ...

    **Scripting an nn.Module**
        Scripting an ``nn.Module`` by default will compile the ``forward`` method and recursively
        compile any methods, submodules, and functions called by ``forward``. If a ``nn.Module`` only uses
        features supported in TorchScript, no changes to the original module code should be necessary. ``script``
        will construct :class:`ScriptModule` that has copies of the attributes, parameters, and methods of
        the original module.

        Example (scripting a simple module with a Parameter):

        .. testcode::

            import torch

            class MyModule(torch.nn.Module):
                def __init__(self, N, M):
                    super(MyModule, self).__init__()
                    # This parameter will be copied to the new ScriptModule
                    self.weight = torch.nn.Parameter(torch.rand(N, M))

                    # When this submodule is used, it will be compiled
                    self.linear = torch.nn.Linear(N, M)

                def forward(self, input):
                    output = self.weight.mv(input)

                    # This calls the `forward` method of the `nn.Linear` module, which will
                    # cause the `self.linear` submodule to be compiled to a `ScriptModule` here
                    output = self.linear(output)
                    return output

            scripted_module = torch.jit.script(MyModule(2, 3))

        Example (scripting a module with traced submodules):

        .. testcode::

            import torch
            import torch.nn as nn
            import torch.nn.functional as F

            class MyModule(nn.Module):
                def __init__(self):
                    super(MyModule, self).__init__()
                    # torch.jit.trace produces a ScriptModule's conv1 and conv2
                    self.conv1 = torch.jit.trace(nn.Conv2d(1, 20, 5), torch.rand(1, 1, 16, 16))
                    self.conv2 = torch.jit.trace(nn.Conv2d(20, 20, 5), torch.rand(1, 20, 16, 16))

                def forward(self, input):
                    input = F.relu(self.conv1(input))
                    input = F.relu(self.conv2(input))
                    return input

            scripted_module = torch.jit.script(MyModule())

        To compile a method other than ``forward`` (and recursively compile anything it calls), add
        the :func:`@torch.jit.export <torch.jit.export>` decorator to the method. To opt out of compilation
        use :func:`@torch.jit.ignore <torch.jit.ignore>` or :func:`@torch.jit.unused <torch.jit.unused>`.

        Example (an exported and ignored method in a module)::

            import torch
            import torch.nn as nn

            class MyModule(nn.Module):
                def __init__(self):
                    super(MyModule, self).__init__()

                @torch.jit.export
                def some_entry_point(self, input):
                    return input + 10

                @torch.jit.ignore
                def python_only_fn(self, input):
                    # This function won't be compiled, so any
                    # Python APIs can be used
                    import pdb
                    pdb.set_trace()

                def forward(self, input):
                    if self.training:
                        self.python_only_fn(input)
                    return input * 99

            scripted_module = torch.jit.script(MyModule())
            print(scripted_module.some_entry_point(torch.randn(2, 2)))
            print(scripted_module(torch.randn(2, 2)))

        Example ( Annotating forward of nn.Module using example_inputs)::

            import torch
            import torch.nn as nn
            from typing import NamedTuple

            class MyModule(NamedTuple):
            result: List[int]

            class TestNNModule(torch.nn.Module):
                def forward(self, a) -> MyModule:
                    result = MyModule(result=a)
                    return result

            pdt_model = TestNNModule()

            # Runs the pdt_model in eager model with the inputs provided and annotates the arguments of forward
            scripted_model = torch.jit.script(pdt_model, example_inputs={pdt_model: [([10, 20, ], ), ], })

            # Run the scripted_model with actual inputs
            print(scripted_model([20]))
    """
    global type_trace_db
    if not _enabled:
        return obj

    if optimize is not None:
        warnings.warn(
            "`optimize` is deprecated and has no effect. Use `with torch.jit.optimized_execution() instead"
        )

    # No-op for modules, functions, class instances that are already scripted
    if isinstance(obj, RecursiveScriptClass):
        return obj
    if isinstance(obj, ScriptModule):
        return obj
    if isinstance(obj, ScriptFunction):
        return obj

    if example_inputs:
        # If MonkeyType is installed, enable profile directed type annotation
        # Check if example_inputs are defined and generate call traces
        # for the method by running eager mode version of the method with
        # the provide example inputs. This logs all the traces in type_trace_db
        type_trace_db = JitTypeTraceStore()
        if monkeytype_trace:
            monkeytype_config = JitTypeTraceConfig(type_trace_db)
            with monkeytype_trace(monkeytype_config):
                if isinstance(example_inputs, Dict):
                    # If the obj is an nn.Module or a class, then each method is
                    # executed with the arguments provided in the example inputs.
                    # example inputs here will be of type Dict(class.method, (arguments))
                    # This is used to infer type annotations for those methods
                    # which are not called directly under the hood of monkeytype.
                    for module, example_input in example_inputs.items():
                        for example in example_input:
                            module(*example)
                elif isinstance(example_inputs, List):
                    for examples in example_inputs:
                        obj(*examples)
                else:
                    raise ValueError("Error: Unable to infer types. Please format the inputs to type `List[Tuple]`"
                                     " or `Dict[Callable, List[Tuple]]` to be run with MonkeyType.")
        else:
            warnings.warn("Warning: monkeytype is not installed. Please install https://github.com/Instagram/MonkeyType "
                          "to enable Profile-Directed Typing in TorchScript. Refer to "
                          "https://github.com/Instagram/MonkeyType/blob/master/README.rst to install MonkeyType. ")

    if isinstance(obj, torch.nn.Module):
        obj = call_prepare_scriptable_func(obj)
        return torch.jit._recursive.create_script_module(
            obj, torch.jit._recursive.infer_methods_to_compile
        )

    if isinstance(obj, dict):
        return create_script_dict(obj)
    if isinstance(obj, list):
        return create_script_list(obj)

    if inspect.isclass(obj):
        qualified_name = _qualified_name(obj)
        # If this type is a `nn.Module` subclass, they probably meant to pass
        # an instance instead of a Module
        if issubclass(obj, torch.nn.Module):
            raise RuntimeError(
                "Type '{}' cannot be compiled since it inherits"
                " from nn.Module,"
                " pass an instance instead".format(obj)
            )

        # Enums are automatically usable in TorchScript, explicitly scripting
        # is not necessary, but not harmful either.
        if issubclass(obj, enum.Enum):
            return obj

        if not _is_new_style_class(obj):
            raise RuntimeError(
                "TorchScript classes must be new-style classes. "
                "Please inherit from 'object'."
            )
        if len(obj.mro()) > 2:
            raise RuntimeError(
                "TorchScript classes does not support inheritance yet. "
                "Please directly inherit from 'object'."
            )
        if _rcb is None:
            _rcb = _jit_internal.createResolutionCallbackFromFrame(_frames_up + 1)
        _compile_and_register_class(obj, _rcb, qualified_name)
        return obj
    elif inspect.isfunction(obj) or inspect.ismethod(obj):
        qualified_name = _qualified_name(obj)
        # this is a decorated fn, and we need to the underlying fn and its rcb
        if hasattr(obj, "__script_if_tracing_wrapper"):
            obj = obj.__original_fn
            _rcb = _jit_internal.createResolutionCallbackFromClosure(obj)

        _check_directly_compile_overloaded(obj)
        maybe_already_compiled_fn = _try_get_jit_cached_function(obj)
        if maybe_already_compiled_fn:
            return maybe_already_compiled_fn
        ast = get_jit_def(obj, obj.__name__)
        if _rcb is None:
            _rcb = _jit_internal.createResolutionCallbackFromClosure(obj)
        fn = torch._C._jit_script_compile(
            qualified_name, ast, _rcb, get_default_args(obj)
        )
        # Forward docstrings
        fn.__doc__ = obj.__doc__
        _set_jit_function_cache(obj, fn)
        return fn
    else:
        return torch.jit._recursive.create_script_class(obj)


# overloads are registered in _jit_internal and compiled here so that _overload
# can be used in nn/functional.py without an import cycle


def _check_overload_defaults(impl_defaults, overload_defaults, loc):
    for name, overload_value in overload_defaults.items():
        if name not in impl_defaults or impl_defaults[name] != overload_value:
            raise torch.jit.frontend.FrontendError(
                loc,
                "Default parameters on overloads do not affect the runtime so they "
                "must equal to the default parameter on the implementation function. Found on "
                "parameter {name}".format(name=name),
            )


def _compile_function_with_overload(overload_fn, qual_name, impl_fn):
    overload_decl = get_jit_def(overload_fn, overload_fn.__name__).decl()
    overload_signature = torch.jit.annotations.get_signature(
        overload_fn, None, None, inspect.ismethod(overload_fn)
    )
    impl_ast = get_jit_def(impl_fn, impl_fn.__name__)
    overload_defaults = get_default_args(overload_fn)
    implementation_defaults = get_default_args(impl_fn)
    _rcb = _jit_internal.createResolutionCallbackFromClosure(impl_fn)
    _check_overload_defaults(
        implementation_defaults, overload_defaults, overload_decl.range()
    )
    fn = torch._C._jit_script_compile_overload(
        qual_name,
        overload_decl,
        impl_ast,
        _rcb,
        implementation_defaults,
        overload_signature,
    )
    return fn


def _get_overloads(obj):
    # check for cached compiled fns
    existing_compiled_fns = _try_get_jit_cached_overloads(obj)
    qual_name = _qualified_name(obj)
    uncompiled_overloads = _jit_internal._get_fn_overloads(qual_name)
    if uncompiled_overloads is None:
        return existing_compiled_fns

    if obj in uncompiled_overloads:
        raise RuntimeError(_jit_internal.get_overload_no_implementation_error_message(
            'function', obj))

    compiled_fns = []
    for overload_fn in uncompiled_overloads:
        compiled_fns.append(
            _compile_function_with_overload(overload_fn, qual_name, obj)
        )

    if existing_compiled_fns:
        compiled_fns = existing_compiled_fns + compiled_fns

    # cache compilation, remove information stored to do compilation
    _set_jit_overload_cache(obj, compiled_fns)
    _jit_internal._clear_fn_overloads(qual_name)
    return compiled_fns


def _check_directly_compile_overloaded(obj):
    qual_name = _qualified_name(obj)
    if _jit_internal._get_fn_overloads(qual_name) or _try_get_jit_cached_overloads(obj):
        raise RuntimeError(
            "Function {} cannot be directly compiled because it"
            " is overloaded. It must be used in a context of a function"
            " where its inputs can determine which overload to call.".format(qual_name)
        )


def interface(obj):
    if not inspect.isclass(obj):
        raise RuntimeError("interface must be applied to a class")
    if not _is_new_style_class(obj):
        raise RuntimeError("TorchScript interfaces must inherit from 'object'")

    # Expected MRO is:
    #   User module
    #   torch.nn.modules.module.Module
    #   object
    is_module_interface = issubclass(obj, torch.nn.Module) and len(obj.mro()) == 3

    if not is_module_interface and len(obj.mro()) > 2:
        raise RuntimeError(
            "TorchScript interface does not support inheritance yet. "
            "Please directly inherit from 'object' or 'nn.Module'."
        )

    qualified_name = _qualified_name(obj)
    rcb = _jit_internal.createResolutionCallbackFromFrame(1)
    # if this type is a `nn.Module` subclass, generate a module interface type
    # instead of a class interface type; a module interface type only compiles
    # the user provided methods as part of the interface
    ast = get_jit_class_def(obj, obj.__name__)
    mangled_classname = torch._C._jit_script_interface_compile(
        qualified_name, ast, rcb, is_module_interface
    )
    obj.__torch_script_interface__ = mangled_classname
    return obj


def _recursive_compile_class(obj, loc):
    _qual_name = _qualified_name(obj)
    # We're starting a new compilation, so update the error call stack in
    # case it fails
    error_stack = torch._C.CallStack(_qual_name, loc)
    rcb = _jit_internal.createResolutionCallbackForClassMethods(obj)
    return _compile_and_register_class(obj, rcb, _qual_name)

CompilationUnit = torch._C.CompilationUnit
set_module(CompilationUnit, "torch.jit")


def pad(s: str, padding: int, offset: int = 0, char: str = ' '):
    if padding >= len(s):
        padding -= len(s)
    return ''.join([char for _ in range(padding + offset)]) + s


class _ScriptProfileColumn:
    def __init__(self, header: str, alignment: int = 4, offset: int = 0):
        self.header = header
        self.alignment = alignment
        self.offset = offset
        self.rows: Dict[int, Any] = {}

    def add_row(self, lineno: int, value: Any):
        self.rows[lineno] = value

    def materialize(self):
        max_length = len(self.header)
        rows: List[Tuple[int, str]] = []
        for (key, value) in self.rows.items():
            cell = str(value)
            rows.append((key, cell))
            max_length = max(len(cell), max_length)

        if self.alignment > 0:
            padding = max_length + self.alignment
            padding -= padding % self.alignment
        else:
            padding = 0

        rows = [(key, pad(cell, padding, self.offset)) for key, cell in rows]
        return pad(self.header, padding, self.offset), rows


class _ScriptProfileTable:
    def __init__(self, cols: List[_ScriptProfileColumn], source_range: List[int]):
        self.cols = cols
        self.source_range = source_range

    def dump_string(self):
        outputs: List[str] = []
        cells: List[Tuple[str, Dict[int, str]]] = []
        header_buffer = ''
        for col in self.cols:
            header, rows = col.materialize()
            header_buffer += header
            cells.append((header, dict(rows)))

        outputs.append(header_buffer)
        outputs.append(pad('', len(header_buffer), 0, '='))
        for line in self.source_range:
            row_buffer = ''
            for header, rows in cells:
                cell = rows.get(line)
                if cell is None:
                    row_buffer += pad('', len(header))
                else:
                    row_buffer += cell
            outputs.append(row_buffer)
        return '\n'.join(outputs)


class _ScriptProfile:
    def __init__(self):
        self.profile = classes.profiling._ScriptProfile()

    def enable(self):
        self.profile.enable()

    def disable(self):
        self.profile.disable()

    def dump_string(self) -> str:
        outputs: List[str] = []
        for source_stats in self.profile._dump_stats():
            source_ref = source_stats.source()
            source_lines = source_ref.text().splitlines()
            dedent = min([len(line) - len(line.lstrip(' ')) for line in source_lines])
            source_lines = [line[dedent:] for line in source_lines]

            start_line = source_ref.starting_lineno()
            end_line = start_line + len(source_lines)
            source_range = range(start_line, end_line)
            lineno = _ScriptProfileColumn("Line #")
            hits = _ScriptProfileColumn("Hits")
            time_ns = _ScriptProfileColumn("Time (ns)")
            line_contents = _ScriptProfileColumn("Line Contents", 0, 1)
            stats = source_stats.line_map()
            for line in source_range:
                lineno.add_row(line, line)
                line_contents.add_row(line, source_lines[line - start_line])
                stat = stats.get(line)
                if stat is not None:
                    hits.add_row(line, stat.count())
                    time_ns.add_row(line, stat.duration_ns())

            table = _ScriptProfileTable([lineno, hits, time_ns, line_contents], list(source_range))
            outputs.append(table.dump_string())
        return '\n\n'.join(outputs)

    def dump(self):
        print(self.dump_string())


def _unwrap_optional(x):
    assert x is not None, "Unwrapping null optional"
    return x


_register_builtin(_unwrap_optional, "aten::_unwrap_optional")
_register_builtin(_jit_internal.is_scripting, "aten::is_scripting")
_register_builtin(has_torch_function, "aten::has_torch_function")
_register_builtin(has_torch_function_unary, "aten::has_torch_function")
_register_builtin(has_torch_function_variadic, "aten::has_torch_function")
Shortcuts

import torch
from torch.types import _TensorOrTensors
import torch.testing
from torch.overrides import is_tensor_like
import collections
from itertools import product
import warnings
from typing import Callable, Union, Optional, Iterable, List, Tuple, Dict
from torch._vmap_internals import vmap
import functools


class GradcheckError(RuntimeError):
    # Custom error so that user errors are not caught in the gradcheck's try-catch
    pass


def _is_float_or_complex_tensor(obj):
    return is_tensor_like(obj) and (obj.is_floating_point() or obj.is_complex())


def _allocate_jacobians_with_inputs(input_tensors: Tuple, numel_output) -> Tuple[torch.Tensor, ...]:
    # Makes zero-filled tensors from inputs. If `numel_output` is not None, for
    # each tensor in `input_tensors`, returns a new zero-filled tensor with height
    # of `t.numel` and width of `numel_output`. Otherwise, for each tensor, returns
    # a 1-d tensor with size `(t.numel,)`. Each new tensor will be strided and have
    # the same dtype and device as those of the corresponding input.
    out: List[torch.Tensor] = []
    for t in input_tensors:
        if _is_float_or_complex_tensor(t) and t.requires_grad:
            out.append(t.new_zeros((t.numel(), numel_output), layout=torch.strided))
    return tuple(out)


def _allocate_jacobians_with_outputs(output_tensors: Tuple, numel_input, dtype=None,
                                     device=None) -> Tuple[torch.Tensor, ...]:
    # Makes zero-filled tensors from outputs. If `dim` is not None, for each tensor
    # in `output_tensors`, returns a new zero-filled tensor with height of `dim` and
    # width of `t.numel`. Otherwise, for each tensor, returns a 1-d tensor with size
    # (t.numel,).
    out: List[torch.Tensor] = []
    options = {"dtype": dtype, "device": device, "layout": torch.strided}
    for t in output_tensors:
        if _is_float_or_complex_tensor(t):
            out.append(t.new_zeros((numel_input, t.numel()), **options))
    return tuple(out)


def _iter_tensors(x: Union[torch.Tensor, Iterable[torch.Tensor]],
                  only_requiring_grad: bool = False) -> Iterable[torch.Tensor]:
    if is_tensor_like(x):
        # mypy doesn't narrow type of `x` to torch.Tensor
        if x.requires_grad or not only_requiring_grad:  # type: ignore[union-attr]
            yield x  # type: ignore[misc]
    elif isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
        for elem in x:
            for result in _iter_tensors(elem, only_requiring_grad):
                yield result


def _iter_tensor(x_tensor):
    # (Only used for slow gradcheck) Returns a generator that yields the following
    # elements at each iteration:
    #  1) a tensor: the same tensor is returned across all iterations. The tensor
    #     is not the same as the original x_tensor as given as input - it is
    #     prepared so that it can be modified in-place. Depending on whether the
    #     input tensor is strided, sparse, or dense, the returned tensor may or may
    #     not share storage with x_tensor.
    #  2) a tuple of indices that can be used with advanced indexing (yielded in
    #     dictionary order)
    #  3) flattened index that will be used to index into the Jacobian tensor
    #
    # For a tensor t with size (2, 2), _iter_tensor yields:
    #     `x, (0, 0), 0`, `x, (0, 1), 1`, `x, (1, 0), 2`, `x, (1, 1), 3`
    #
    # where x is the t.data of the original tensor. Perturbing the entry of x
    # at index (1, 1) yields the 3rd column of the overall Jacobian matrix.
    if x_tensor.is_sparse:
        def get_stride(size):
            dim = len(size)
            tmp = 1
            stride = [0] * dim
            for i in reversed(range(dim)):
                stride[i] = tmp
                tmp *= size[i]
            return stride
        x_nnz = x_tensor._nnz()
        x_size = list(x_tensor.size())
        x_indices = x_tensor._indices().t()
        x_values = x_tensor._values()
        x_stride = get_stride(x_size)
        # Use .data here to get around the version check
        x_values = x_values.data
        for i in range(x_nnz):
            x_value = x_values[i]
            for x_idx in product(*[range(m) for m in x_values.size()[1:]]):
                indices = x_indices[i].tolist() + list(x_idx)
                d_idx = sum(indices[k] * x_stride[k] for k in range(len(x_size)))
                yield x_value, x_idx, d_idx
    elif x_tensor.layout == torch._mkldnn:  # type: ignore[attr-defined]
        for d_idx, x_idx in enumerate(product(*[range(m) for m in x_tensor.size()])):
            # this is really inefficient, but without indexing implemented, there's
            # not really a better way than converting back and forth
            x_tensor_dense = x_tensor.to_dense()
            yield x_tensor_dense, x_idx, d_idx
    else:
        # Use .data here to get around the version check
        x_tensor = x_tensor.data
        for d_idx, x_idx in enumerate(product(*[range(m) for m in x_tensor.size()])):
            yield x_tensor, x_idx, d_idx


def _get_numerical_jacobian(fn, inputs, outputs=None, target=None, eps=1e-3,
                            is_forward_ad=False) -> List[Tuple[torch.Tensor, ...]]:
    """Computes the numerical Jacobian of `fn(inputs)` with respect to `target`. If
    not specified, targets are the input. Returns M * N Jacobians where N is the
    number of tensors in target that require grad and M is the number of non-integral
    outputs.

    Args:
        fn: the function to compute the jacobian for
        inputs: inputs to `fn`
        outputs: provide precomputed outputs to avoid one extra invocation of fn
        target: the Tensors wrt whom Jacobians are calculated (default=`inputs`)
        eps: the magnitude of the perturbation during finite differencing
             (default=`1e-3`)
        is_forward_ad: if this numerical jacobian is computed to be checked wrt
                       forward AD gradients (this is used for error checking only)

    Returns:
        A list of M N-tuples of tensors

    Note that `target` may not even be part of `input` to `fn`, so please be
    **very careful** in this to not clone `target`.
    """
    jacobians: List[Tuple[torch.Tensor, ...]] = []
    if outputs is None:
        outputs = _as_tuple(fn(*_as_tuple(inputs)))
    if not is_forward_ad and any(o.is_complex() for o in outputs):
        raise ValueError("Expected output to be non-complex. get_numerical_jacobian no "
                         "longer supports functions that return complex outputs.")
    if target is None:
        target = inputs
    inp_indices = [i for i, a in enumerate(target) if is_tensor_like(a) and a.requires_grad]
    for i, (inp, inp_idx) in enumerate(zip(_iter_tensors(target, True), inp_indices)):
        jacobians += [get_numerical_jacobian_wrt_specific_input(fn, inp_idx, inputs, outputs, eps,
                                                                input=inp, is_forward_ad=is_forward_ad)]
    return jacobians


def get_numerical_jacobian(fn, inputs, target=None, eps=1e-3, grad_out=1.0):
    """Deprecated API to compute the numerical Jacobian for a given fn and its inputs.

    Args:
        fn: the function to compute the Jacobian for (must take inputs as a tuple)
        input: input to `fn`
        target: the Tensors wrt whom Jacobians are calculated (default=`input`)
        eps: the magnitude of the perturbation during finite differencing
             (default=`1e-3`)

    Returns:
        A list of Jacobians of `fn` (restricted to its first output) with respect to
        each input or target, if provided.

    Note that `target` may not even be part of `input` to `fn`, so please be
    **very careful** in this to not clone `target`.
    """
    warnings.warn("get_numerical_jacobian was part of PyTorch's private API and not "
                  "meant to be exposed. We are deprecating it and it will be removed "
                  "in a future version of PyTorch. If you have a specific use for "
                  "this or feature request for this to be a stable API, please file "
                  "us an issue at https://github.com/pytorch/pytorch/issues/new")
    if grad_out != 1.0:  # grad_out param is only kept for backward compatibility reasons
        raise ValueError("Expected grad_out to be 1.0. get_numerical_jacobian no longer "
                         "supports values of grad_out != 1.0.")

    def fn_pack_inps(*inps):
        return fn(inps)
    jacobians = _get_numerical_jacobian(fn_pack_inps, inputs, None, target, eps)

    return tuple(jacobian_for_each_output[0] for jacobian_for_each_output in jacobians)


def _compute_numerical_gradient(fn, entry, v, norm_v, nbhd_checks_fn):
    # Performs finite differencing by perturbing `entry` in-place by `v` and
    # returns the gradient of each of the outputs wrt to x at idx.
    orig = entry.clone()
    entry.copy_(orig - v)
    outa = fn()
    entry.copy_(orig + v)
    outb = fn()
    entry.copy_(orig)

    def compute(a, b):
        nbhd_checks_fn(a, b)
        ret = (b - a) / (2 * norm_v)
        return ret.detach().reshape(-1)

    return tuple(compute(a, b) for (a, b) in zip(outa, outb))


def _compute_numerical_jvps_wrt_specific_input(jvp_fn, delta, input_is_complex,
                                               is_forward_ad=False) -> List[torch.Tensor]:
    # Computing the jacobian only works for real delta
    # For details on the algorithm used here, refer:
    # Section 3.5.3 https://arxiv.org/pdf/1701.00392.pdf
    # s = fn(z) where z = x for real valued input
    # and z = x + yj for complex valued input
    jvps: List[torch.Tensor] = []
    ds_dx_tup = jvp_fn(delta[0] if isinstance(delta, tuple) else delta)

    if input_is_complex:  # C -> R
        ds_dy_tup = jvp_fn(delta[1] * 1j) if isinstance(delta, tuple) else jvp_fn(delta * 1j)
        for ds_dx, ds_dy in zip(ds_dx_tup, ds_dy_tup):
            assert(not ds_dx.is_complex())
            # conjugate wirtinger derivative
            conj_w_d = ds_dx + ds_dy * 1j
            jvps.append(conj_w_d)
    else:
        for ds_dx in ds_dx_tup:  # R -> R or (R -> C for the forward AD case)
            assert(is_forward_ad or not ds_dx.is_complex())
            jvps.append(ds_dx)
    return jvps


def _combine_jacobian_cols(jacobians_cols: Dict[int, List[torch.Tensor]], outputs, input,
                           numel) -> Tuple[torch.Tensor, ...]:
    # jacobian_cols maps column_idx -> output_idx -> single column of jacobian Tensor
    # we return a list that maps output_idx -> full jacobian Tensor
    jacobians = _allocate_jacobians_with_outputs(outputs, numel, dtype=input.dtype if input.dtype.is_complex else None)
    for i, jacobian in enumerate(jacobians):
        for k, v in jacobians_cols.items():
            jacobian[k] = v[i]
    return jacobians


def _prepare_input(input: torch.Tensor, maybe_perturbed_input: Optional[torch.Tensor],
                   fast_mode=False) -> torch.Tensor:
    # Prepares the inputs to be passed into the function while including the new
    # modified input.
    if input.layout == torch._mkldnn:  # type: ignore[attr-defined] # no attr _mkldnn
        # Convert back to mkldnn
        if maybe_perturbed_input is not None:
            return maybe_perturbed_input.to_mkldnn()
        else:
            return input
    elif input.layout == torch.sparse_coo:
        if fast_mode and maybe_perturbed_input is not None:
            # entry is already a "cloned" version of the original tensor
            # thus changes to entry are not reflected in the input
            return maybe_perturbed_input
        else:
            return input
    else:
        # We cannot use entry (input.data) if we want gradgrad to work because
        # fn (in the gradgrad case) needs to compute grad wrt input
        return input


def check_outputs_same_dtype_and_shape(output1, output2, eps, idx=None) -> None:
    # Check that the returned outputs don't have different dtype or shape when you
    # perturb the input
    on_index = "on index {idx} " if idx is not None else ""
    assert output1.shape == output2.shape, \
        (f"Expected `func` to return outputs with the same shape"
         f" when inputs are perturbed {on_index}by {eps}, but got:"
         f" shapes {output1.shape} and {output2.shape}.")
    assert output1.dtype == output2.dtype, \
        (f"Expected `func` to return outputs with the same dtype"
         f" when inputs are perturbed {on_index}by {eps}, but got:"
         f" dtypes {output1.dtype} and {output2.dtype}.")


def get_numerical_jacobian_wrt_specific_input(fn, input_idx, inputs, outputs, eps,
                                              input=None, is_forward_ad=False) -> Tuple[torch.Tensor, ...]:
    # Computes the numerical jacobians wrt to a single input. Returns N jacobian
    # tensors, where N is the number of outputs. We use a dictionary for
    # jacobian_cols because indices aren't necessarily consecutive for sparse inputs
    # When we perturb only a single element of the input tensor at a time, the jvp
    # is equivalent to a single col of the Jacobian matrix of fn.
    jacobian_cols: Dict[int, List[torch.Tensor]] = {}
    input = inputs[input_idx] if input is None else input
    assert input.requires_grad
    for x, idx, d_idx in _iter_tensor(input):
        wrapped_fn = _with_prepare_inputs(fn, inputs, input_idx, x)
        input_to_perturb = x[idx]
        nbhd_checks_fn = functools.partial(check_outputs_same_dtype_and_shape, idx=idx, eps=eps)
        jvp_fn = _get_numerical_jvp_fn(wrapped_fn, input_to_perturb, eps, nbhd_checks_fn)
        jacobian_cols[d_idx] = _compute_numerical_jvps_wrt_specific_input(jvp_fn, eps, x.is_complex(), is_forward_ad)
    return _combine_jacobian_cols(jacobian_cols, outputs, input, input.numel())

def _get_analytical_jacobian_forward_ad(fn, inputs, outputs, *, check_grad_dtypes=False,
                                        all_u=None) -> Tuple[Tuple[torch.Tensor, ...], ...]:
    """Computes the analytical Jacobian using forward mode AD of `fn(inputs)` using forward mode AD with respect
    to `target`. Returns N * M Jacobians where N is the number of tensors in target that require grad and
    M is the number of non-integral outputs.
    Contrary to other functions here, this function requires "inputs" to actually be used by the function.
    The computed value is expected to be wrong if the function captures the inputs by side effect instead of
    using the passed ones (many torch.nn tests do this).

    Args:
        fn: the function to compute the jacobian for
        inputs: inputs to `fn`
        outputs: provide precomputed outputs to avoid one extra invocation of fn
        check_grad_dtypes: if True, will check that the gradient dtype are valid
        all_u (optional): if provided, the Jacobian will be right multiplied with this vector

    Returns:
        A tuple of M N-tuples of tensors
    """
    # To avoid early import issues
    fwAD = torch.autograd.forward_ad

    tensor_inputs = tuple(i for i in inputs if is_tensor_like(i) and i.requires_grad)

    if any(i.is_complex() for i in tensor_inputs):
        raise ValueError("Expected inputs to be non-complex for _get_analytical_jacobian_forward_ad.")

    if all_u:
        jacobians = tuple(_allocate_jacobians_with_outputs(outputs, 1) for i in tensor_inputs)
    else:
        jacobians = tuple(_allocate_jacobians_with_outputs(outputs, i.numel()) for i in tensor_inputs)

    with fwAD.dual_level():
        fw_grads = []
        dual_inputs = []
        for i, inp in enumerate(inputs):
            if is_tensor_like(inp) and inp.requires_grad:
                if inp.layout == torch._mkldnn:  # type: ignore[attr-defined]
                    raise ValueError("MKLDNN inputs are not support for forward AD gradcheck.")

                inp = fwAD.make_dual(inp, torch.zeros_like(inp))
                # If inp is a differentiable view, the dual might not be the tangent given to
                # make_dual, so read it explicitly from the dual tensor
                fw_grads.append(fwAD.unpack_dual(inp)[1])
            dual_inputs.append(inp)

        if all_u:
            # Do the full reduction in one pass
            # To be consistent with numerical evaluation, we actually compute one reduction per input
            for i, (fw_grad, u) in enumerate(zip(fw_grads, all_u)):
                fw_grad.copy_(u.view_as(fw_grad))
                raw_outputs = _as_tuple(fn(*dual_inputs))
                dual_outputs = filter(_is_float_or_complex_tensor, raw_outputs)
                for index_o, d_o in enumerate(dual_outputs):
                    val, res = fwAD.unpack_dual(d_o)
                    if check_grad_dtypes and val.is_complex() != res.is_complex():
                        raise GradcheckError('Forward AD gradient has dtype mismatch.')

                    # Remove extra dimension of size 1 corresponding to the reduced input
                    jacobians[i][index_o].squeeze_(0)
                    if res is None:
                        jacobians[i][index_o].zero_()
                    else:
                        jacobians[i][index_o].copy_(res.reshape(-1))
                fw_grad.zero_()
        else:
            # Reconstruct the full Jacobian column by column
            for i, fw_grad in enumerate(fw_grads):
                for lin_idx, grad_idx in enumerate(product(*[range(m) for m in fw_grad.size()])):
                    fw_grad[grad_idx] = 1.
                    raw_outputs = _as_tuple(fn(*dual_inputs))
                    dual_outputs = filter(_is_float_or_complex_tensor, raw_outputs)
                    for index_o, d_o in enumerate(dual_outputs):
                        val, res = fwAD.unpack_dual(d_o)
                        if check_grad_dtypes and val.is_complex() != res.is_complex():
                            raise GradcheckError('Forward AD gradient has dtype mismatch.')

                        if res is None:
                            jacobians[i][index_o][lin_idx].zero_()
                        else:
                            jacobians[i][index_o][lin_idx].copy_(res.reshape(-1))
                    fw_grad[grad_idx] = 0.

    return jacobians

def _get_input_to_perturb(input):
    # Prepare the input so that it can be modified in-place and do certain
    # operations that require the tensor to have strides. If fast_mode=False,
    # _iter_tensor would handle the below cases:
    if input.layout == torch._mkldnn:  # type: ignore[attr-defined] # no attr _mkldnn
        # Convert to dense so we can perform operations that require strided tensors
        input_to_perturb = input.to_dense()
    elif input.layout == torch.sparse_coo:
        # Clone because input may require grad, and copy_ calls resize_,
        # which is not allowed for .data
        input_to_perturb = input.clone()
    else:
        input_to_perturb = input.data
    return input_to_perturb


def _with_prepare_inputs(fn, inputs, input_idx, input_to_perturb, fast_mode=False):
    # Wraps `fn` so that its inputs are already supplied
    def wrapped_fn():
        inp = tuple(_prepare_input(a, input_to_perturb if i == input_idx else None, fast_mode)
                    if is_tensor_like(a) else a for i, a in enumerate(_as_tuple(inputs)))
        return tuple(a.clone() for a in _as_tuple(fn(*inp)))
    return wrapped_fn


def _get_numerical_jvp_fn(wrapped_fn, input_to_perturb, eps, nbhd_checks_fn):
    # Wraps jvp_fn so that certain arguments are already supplied
    def jvp_fn(delta):
        return _compute_numerical_gradient(wrapped_fn, input_to_perturb, delta, eps, nbhd_checks_fn)
    return jvp_fn


def _reshape_tensor_or_tuple(u, shape):
    # We don't need to reshape when input corresponding to u is sparse
    if isinstance(u, tuple):
        if u[0].layout != torch.sparse_coo:
            return (u[0].reshape(shape), u[1].reshape(shape))
    else:
        if u.layout != torch.sparse_coo:
            return u.reshape(shape)
    return u


def _mul_tensor_or_tuple(u, k):
    if isinstance(u, tuple):
        return (k * u[0], k * u[1])
    else:
        return k * u


def _get_numerical_jvp_wrt_specific_input(fn, input_idx, inputs, u, eps, is_forward_ad=False) -> List[torch.Tensor]:
    input = inputs[input_idx]
    input_to_perturb = _get_input_to_perturb(input)
    wrapped_fn = _with_prepare_inputs(fn, inputs, input_idx, input_to_perturb, True)
    nbhd_checks_fn = functools.partial(check_outputs_same_dtype_and_shape, eps=eps)
    jvp_fn = _get_numerical_jvp_fn(wrapped_fn, input_to_perturb, eps, nbhd_checks_fn)
    u = _reshape_tensor_or_tuple(u, input_to_perturb.shape)
    u = _mul_tensor_or_tuple(u, eps)
    return _compute_numerical_jvps_wrt_specific_input(jvp_fn, u, input.is_complex(), is_forward_ad)


def _get_numerical_vJu(fn, inputs, inp_indices, func_out, all_u, all_v, eps, is_forward_ad):
    # Note that all_v can also be None, in that case, this function only computes Ju.
    reduced_jacobians: List[List[torch.Tensor]] = []
    for i, (inp_idx, u) in enumerate(zip(inp_indices, all_u)):
        all_Ju = _get_numerical_jvp_wrt_specific_input(fn, inp_idx, inputs, u, eps, is_forward_ad)
        # Filter out the Ju for non floating point outputs
        filtered_Ju = []
        func_out = _as_tuple(func_out)
        assert len(all_Ju) == len(func_out)
        for Ju, output in zip(all_Ju, func_out):
            if _is_float_or_complex_tensor(output):
                filtered_Ju.append(Ju)
            else:
                # TODO: handle the other Ju
                pass
        if all_v is not None:
            jacobian_scalars: List[torch.Tensor] = []
            for v, Ju in zip(all_v, filtered_Ju):
                jacobian_scalars.append(_dot_with_type_promotion(v, Ju))
            reduced_jacobians.append(jacobian_scalars)
        else:
            reduced_jacobians.append(filtered_Ju)
    return reduced_jacobians


def _check_jacobians_equal(j1, j2, atol):
    # Check whether the max difference between two Jacobian tensors are within some
    # tolerance `atol`.
    for j1_x, j2_x in zip(j1, j2):
        if j1_x.numel() != 0 and (j1_x - j2_x).abs().max() > atol:
            return False
    return True


def _stack_and_check_tensors(list_of_list_of_tensors, inputs,
                             numel_outputs) -> Tuple[Tuple[torch.Tensor, ...], bool, bool]:
    # For the ith tensor in the inner list checks whether it has the same size and
    # dtype as the ith differentiable input.
    out_jacobians = _allocate_jacobians_with_inputs(inputs, numel_outputs)
    diff_input_list = list(_iter_tensors(inputs, True))
    correct_grad_sizes = True
    correct_grad_types = True
    for i, tensor_list in enumerate(list_of_list_of_tensors):
        inp = diff_input_list[i]
        out_jacobian = out_jacobians[i]
        for j, tensor in enumerate(tensor_list):
            if tensor is not None and tensor.size() != inp.size():
                correct_grad_sizes = False
            elif tensor is not None and tensor.dtype != inp.dtype:
                correct_grad_types = False
            if tensor is None:
                out_jacobian[:, j].zero_()
            else:
                dense = tensor.to_dense() if not tensor.layout == torch.strided else tensor
                assert out_jacobian[:, j].numel() == dense.numel()
                out_jacobian[:, j] = dense.reshape(-1)
    return out_jacobians, correct_grad_sizes, correct_grad_types


FAILED_NONDET_MSG = """\n
NOTE: If your op relies on non-deterministic operations i.e., it is listed here:
https://pytorch.org/docs/stable/generated/torch.use_deterministic_algorithms.html
this failure might be expected.

If you are adding a new operator, please file an issue and then use one of the
workarounds. The workaround depends on how your test invokes gradcheck/gradgradcheck.
If the test
- manually invokes gradcheck/gradgradcheck, then call gradcheck/gradgradcheck
  with `nondet_tol=<tol>` as a keyword argument.
- is OpInfo-based (e.g., in test_ops.py), then modify the OpInfo for the test
  to have `gradcheck_nondet_tol=<tol>`.
- is a Module test (e.g., in common_nn.py), then modify the corresponding
  module_test entry to have `gradcheck_nondet_tol=<tol>`
"""


def _check_analytical_jacobian_attributes(inputs, output, nondet_tol, check_grad_dtypes,
                                          fast_mode=False, v=None) -> Tuple[torch.Tensor, ...]:
    # This is used by both fast and slow mode:
    #  - For slow mode, vjps[i][j] is the jth row the Jacobian wrt the ith
    #    input.
    #  - For fast mode, vjps[i][0] is a linear combination of the rows
    #    of the Jacobian wrt the ith input
    diff_input_list = list(_iter_tensors(inputs, True))

    def vjp_fn(grad_output):
        return torch.autograd.grad(output, diff_input_list, grad_output,
                                   retain_graph=True, allow_unused=True)
    # Compute everything twice to check for nondeterminism (which we call reentrancy)
    if fast_mode:
        vjps1 = _get_analytical_vjps_wrt_specific_output(vjp_fn, output.clone(), v)
        vjps2 = _get_analytical_vjps_wrt_specific_output(vjp_fn, output.clone(), v)
    else:
        vjps1 = _compute_analytical_jacobian_rows(vjp_fn, output.clone())
        vjps2 = _compute_analytical_jacobian_rows(vjp_fn, output.clone())

    output_numel = output.numel() if not fast_mode else 1
    jacobians1, types_ok, sizes_ok = _stack_and_check_tensors(vjps1, inputs, output_numel)
    jacobians2, _, _ = _stack_and_check_tensors(vjps2, inputs, output_numel)
    reentrant = _check_jacobians_equal(jacobians1, jacobians2, nondet_tol)

    if not types_ok and check_grad_dtypes:
        raise GradcheckError('Gradient has dtype mismatch')
    if not sizes_ok:
        raise GradcheckError('Analytical gradient has incorrect size')
    if not reentrant:
        raise GradcheckError('Backward is not reentrant, i.e., running backward with '
                             'same input and grad_output multiple times gives different values, '
                             'although analytical gradient matches numerical gradient.'
                             f'The tolerance for nondeterminism was {nondet_tol}.' +
                             FAILED_NONDET_MSG)
    return jacobians1


def _get_analytical_vJu_backward_mode(inputs, outputs, nondet_tol, check_grad_dtypes, all_v, all_u):
    reduced_jacobians: List[List[torch.Tensor]] = []
    for output, v in zip(outputs, all_v):
        all_vJ = _check_analytical_jacobian_attributes(inputs, output, nondet_tol, check_grad_dtypes,
                                                       fast_mode=True, v=v)
        jacobian_scalars: List[torch.Tensor] = []
        for vJ, u in zip(all_vJ, all_u):
            # Why do we need squeeze here? vJ is a 2-d tensor so that we can reuse
            # the error checking logic from slow mode
            vJ = vJ.T.squeeze(0)
            if vJ.is_complex():  # C -> R
                tv = torch.view_as_real(vJ.resolve_conj())
                tr = tv.select(-1, 0)
                ti = tv.select(-1, 1)
                jacobian_scalars.append(tr.dot(u[0]) + 1j * ti.dot(u[1]))
            else:  # R -> R
                jacobian_scalars.append(vJ.dot(u))
        reduced_jacobians.append(jacobian_scalars)
    return reduced_jacobians

def get_analytical_jacobian(inputs, output, nondet_tol=0.0, grad_out=1.0):
    # Replicates the behavior of the old get_analytical_jacobian before the refactor
    # This shares much of its code with _check_analytical_jacobian_attributes
    warnings.warn("get_analytical_jacobian was part of PyTorch's private API and not "
                  "meant to be exposed. We are deprecating it and it will be removed "
                  "in a future version of PyTorch. If you have a specific use for "
                  "this or feature request for this to be a stable API, please file "
                  "us an issue at https://github.com/pytorch/pytorch/issues/new")
    if grad_out != 1.0:  # grad_out param is only kept for backward compatibility reasons
        raise ValueError("Expected grad_out to be 1.0. get_analytical_jacobian no longer "
                         "supports values of grad_out != 1.0.")
    if output.is_complex():
        raise ValueError("Expected output to be non-complex. get_analytical_jacobian no "
                         "longer supports functions that return complex outputs.")
    diff_input_list = list(_iter_tensors(inputs, True))

    def vjp_fn(grad_output):
        return torch.autograd.grad(output, diff_input_list, grad_output,
                                   retain_graph=True, allow_unused=True)
    # Compute everything twice to check for nondeterminism (which we call reentrancy)
    vjps1 = _compute_analytical_jacobian_rows(vjp_fn, output.clone())
    vjps2 = _compute_analytical_jacobian_rows(vjp_fn, output.clone())

    output_numel = output.numel()
    jacobians1, types_ok, sizes_ok = _stack_and_check_tensors(vjps1, inputs, output_numel)
    jacobians2, _, _ = _stack_and_check_tensors(vjps2, inputs, output_numel)
    reentrant = _check_jacobians_equal(jacobians1, jacobians2, nondet_tol)

    return jacobians1, reentrant, sizes_ok, types_ok


def _get_analytical_jacobian(inputs, outputs, input_idx, output_idx):
    # Computes the analytical Jacobian in slow mode for a single input-output pair.
    # Forgoes performing checks on dtype, shape, and reentrancy.
    jacobians = _check_analytical_jacobian_attributes(inputs, outputs[output_idx],
                                                      nondet_tol=float('inf'), check_grad_dtypes=False)
    return jacobians[input_idx]


def _compute_analytical_jacobian_rows(vjp_fn, sample_output) -> List[List[Optional[torch.Tensor]]]:
    # Computes Jacobian row-by-row using backward function `vjp_fn` = v^T J
    # NB: this function does not assume vjp_fn(v) to return tensors with the same
    # number of elements for different v. This is checked when we later combine the
    # rows into a single tensor.
    grad_out_base = torch.zeros_like(sample_output, memory_format=torch.legacy_contiguous_format)
    flat_grad_out = grad_out_base.view(-1)
    # jacobians_rows[i][j] represents the jth row of the ith input
    jacobians_rows: List[List[Optional[torch.Tensor]]] = []
    for j in range(flat_grad_out.numel()):
        flat_grad_out.zero_()
        flat_grad_out[j] = 1.0
        grad_inputs = vjp_fn(grad_out_base)
        for i, d_x in enumerate(grad_inputs):
            if j == 0:
                jacobians_rows.append([])
            jacobians_rows[i] += [d_x.clone() if isinstance(d_x, torch.Tensor) else None]
    return jacobians_rows


def _get_analytical_vjps_wrt_specific_output(vjp_fn, sample_output, v) -> List[List[Optional[torch.Tensor]]]:
    vjps: List[List[Optional[torch.Tensor]]] = []
    grad_inputs = vjp_fn(v.reshape(sample_output.shape))
    for vjp in grad_inputs:
        vjps.append([vjp.clone() if isinstance(vjp, torch.Tensor) else None])
    return vjps


def _check_inputs(tupled_inputs, check_sparse_nnz) -> bool:
    if not check_sparse_nnz and any(t.is_sparse for t in tupled_inputs if isinstance(t, torch.Tensor)):
        raise GradcheckError('gradcheck expects all tensor inputs are dense when check_sparse_nnz is set to False.')
    # Make sure that gradients are saved for at least one input
    any_input_requiring_grad = False
    for idx, inp in enumerate(tupled_inputs):
        if is_tensor_like(inp) and inp.requires_grad:
            if not (inp.dtype == torch.float64 or inp.dtype == torch.complex128):
                warnings.warn(
                    f'Input #{idx} requires gradient and '
                    'is not a double precision floating point or complex. '
                    'This check will likely fail if all the inputs are '
                    'not of double precision floating point or complex. ')
            content = inp._values() if inp.is_sparse else inp
            # TODO: To cover more problematic cases, replace stride = 0 check with
            # "any overlap in memory" once we have a proper function to check it.
            if content.layout is not torch._mkldnn:  # type: ignore[attr-defined]
                if not all(st > 0 or sz <= 1 for st, sz in zip(content.stride(), content.size())):
                    raise RuntimeError(
                        f'The {idx}th input has a dimension with stride 0. gradcheck only '
                        'supports inputs that are non-overlapping to be able to '
                        'compute the numerical gradients correctly. You should call '
                        '.contiguous on the input before passing it to gradcheck.')
            any_input_requiring_grad = True
            inp.retain_grad()
    if not any_input_requiring_grad:
        raise ValueError(
            'gradcheck expects at least one input tensor to require gradient, '
            'but none of the them have requires_grad=True.')
    return True


def _check_outputs(outputs) -> None:
    if any(t.layout == torch.sparse_coo for t in outputs if isinstance(t, torch.Tensor)):
        # it is easier to call to_dense() on the sparse output than
        # to modify analytical jacobian
        raise ValueError('Sparse output is not supported at gradcheck yet. '
                         'Please call to_dense() on the output of fn for gradcheck.')
    if any(t.layout == torch._mkldnn for t in outputs if isinstance(t, torch.Tensor)):  # type: ignore[attr-defined]
        raise ValueError('MKLDNN output is not supported at gradcheck yet. '
                         'Please call to_dense() on the output of fn for gradcheck.')


def _check_no_differentiable_outputs(func, inputs, func_out, eps) -> bool:
    # When there are no differentiable outputs, numerical gradient for a function is
    # expected to be zero.
    jacobians_all_inputs_outputs = _get_numerical_jacobian(func, inputs, func_out, eps=eps)
    for jacobians_all_outputs_and_fixed_input in jacobians_all_inputs_outputs:
        for jacobian in jacobians_all_outputs_and_fixed_input:
            if torch.ne(jacobian, 0).sum() > 0:
                raise GradcheckError('Numerical gradient for function expected to be zero')
    return True


def _check_no_differentiable_outputs_fast(func, func_out, all_inputs, inputs_indices,
                                          all_u, eps, nondet_tol):
    for inp_idx, u in zip(inputs_indices, all_u):
        jvps = _get_numerical_jvp_wrt_specific_input(func, inp_idx, all_inputs, u, eps)
        for jvp in jvps:
            if jvp.numel() == 0:
                continue
            if (jvp - torch.zeros_like(jvp)).abs().max() > nondet_tol:
                raise GradcheckError('Numerical gradient for function expected to be zero')
    return True


FAILED_BATCHED_GRAD_MSG = """
gradcheck or gradgradcheck failed while testing batched gradient computation.
This could have been invoked in a number of ways (via a test that calls
gradcheck/gradgradcheck directly or via an autogenerated test).

If you are adding a new operator, please file an issue and then use one of the
workarounds. The workaround depends on how your test invokes gradcheck/gradgradcheck.
If the test
- manually invokes gradcheck/gradgradcheck, then call gradcheck/gradgradcheck
  with `check_batched_grad=False` as a keyword argument.
- is OpInfo-based (e.g., in test_ops.py), then modify the OpInfo for the test
  to have `check_batched_grad=False` and/or `check_batched_gradgrad=False`.
- is common_method_invocations-based, then add your test to the denylist
  EXCLUDE_BATCHED_GRAD_TESTS in test_autograd.py

If you're modifying an existing operator that supports batched grad computation,
or wish to make a new operator work with batched grad computation, please read
the following.

To compute batched grads (e.g., jacobians, hessians), we vmap over the backward
computation. The most common failure case is if there is a 'vmap-incompatible
operation' in the backward pass. Please see
NOTE: [How to write vmap-compatible backward formulas]
in the codebase for an explanation of how to fix this.
""".strip()

def _get_failed_batched_grad_test_msg(output_idx, input_idx, res, exp):
    return f"""
For output {output_idx} and input {input_idx}:

{FAILED_BATCHED_GRAD_MSG}

Got:
{res}

Expected:
{exp}
""".strip()


def _test_batched_grad(input, output, output_idx) -> bool:
    # NB: _test_batched_grad compares two autograd.grad invocations with a single
    # vmap(autograd.grad) invocation. It's not exactly a "gradcheck" in the
    # sense that we're not comparing an analytical jacobian with a numeric one,
    # but it is morally similar (we could have computed a full analytic jac
    # via vmap, but that is potentially slow)
    diff_input_list = list(_iter_tensors(input, True))
    grad = functools.partial(torch.autograd.grad, output, diff_input_list, retain_graph=True, allow_unused=True)

    def vjp(v):
        results = grad(v)
        results = tuple(grad if grad is not None else
                        torch.zeros([], dtype=inp.dtype, device=inp.device).expand(inp.shape)
                        for grad, inp in zip(results, diff_input_list))
        return results

    grad_outputs = [torch.randn_like(output) for _ in range(2)]

    expected = [vjp(gO) for gO in grad_outputs]
    expected = [torch.stack(shards) for shards in zip(*expected)]

    # Squash warnings since these are expected to happen in most cases
    # NB: this doesn't work for CUDA tests: https://github.com/pytorch/pytorch/issues/50209
    with warnings.catch_warnings():
        warnings.filterwarnings("ignore", message="Batching rule not implemented")
        warnings.filterwarnings("ignore", message="torch.vmap is an experimental prototype")
        try:
            result = vmap(vjp)(torch.stack(grad_outputs))
        except RuntimeError as ex:
            # It's OK that we're not raising the error at the correct callsite.
            # That's because the callsite is always going to inside the Python
            # autograd.grad instead of the C++ traceback of what line in the
            # backward formula
            raise GradcheckError(
                f'While computing batched gradients, got: {ex}\n\n{FAILED_BATCHED_GRAD_MSG}')

    for input_idx, (res, exp) in enumerate(zip(result, expected)):
        if torch.allclose(res, exp):
            continue
        raise GradcheckError(_get_failed_batched_grad_test_msg(output_idx, input_idx, res, exp))
    return True


def _test_backward_mul_by_grad_output(outputs, inputs, check_sparse_nnz) -> bool:
    # Tests that backward is multiplied by grad_output
    diff_input_list: List[torch.Tensor] = list(_iter_tensors(inputs, True))
    if not diff_input_list:
        raise GradcheckError("no Tensors requiring grad found in input")
    grads_input = torch.autograd.grad(outputs, diff_input_list,
                                      [torch.zeros_like(o, memory_format=torch.legacy_contiguous_format) for o in outputs],
                                      allow_unused=True)
    for gi, di in zip(grads_input, diff_input_list):
        if gi is None:
            continue
        if isinstance(gi, torch.Tensor) and gi.layout != torch.strided:
            if gi.layout != di.layout:
                raise GradcheckError('grad is incorrect layout (' + str(gi.layout) + ' is not ' + str(di.layout) + ')')
            if gi.layout == torch.sparse_coo:
                if gi.sparse_dim() != di.sparse_dim():
                    raise GradcheckError('grad is sparse tensor, but has incorrect sparse_dim')
                if gi.dense_dim() != di.dense_dim():
                    raise GradcheckError('grad is sparse tensor, but has incorrect dense_dim')
            gi = gi.to_dense()
            di = di.to_dense()

        if check_sparse_nnz:
            if not torch.allclose(gi, torch.zeros_like(gi)):
                raise GradcheckError('backward not multiplied by grad_output')
        elif not gi.eq(0).all():
            raise GradcheckError('backward not multiplied by grad_output')
        if gi.dtype != di.dtype or gi.device != di.device or gi.is_sparse != di.is_sparse:
            raise GradcheckError("grad is incorrect type")
        if gi.size() != di.size():
            raise GradcheckError('grad is incorrect size')
    return True


def _test_undefined_grad(func, outputs, inputs) -> bool:
    diff_input_list: List[torch.Tensor] = list(_iter_tensors(inputs, True))
    if not diff_input_list:
        raise GradcheckError("no Tensors requiring grad found in input")

    def warn_bc_breaking():
        warnings.warn((
            'Backwards compatibility: New undefined gradient support checking '
            'feature is enabled by default, but it may break existing callers '
            'of this function. If this is true for you, you can call this '
            'function with "check_undefined_grad=False" to disable the feature'))

    def check_undefined_grad_support(output_to_check):
        grads_output = [torch.zeros_like(o, memory_format=torch.legacy_contiguous_format) for o in output_to_check]
        try:
            grads_input = torch.autograd.grad(output_to_check, diff_input_list,
                                              grads_output, allow_unused=True)
        except RuntimeError:
            warn_bc_breaking()
            raise GradcheckError((
                'Expected backward function to handle undefined output grads. '
                'Please look at "Notes about undefined output gradients" in '
                '"tools/autograd/derivatives.yaml"'))

        for gi, i in zip(grads_input, diff_input_list):
            if (gi is not None) and (not gi.eq(0).all()):
                warn_bc_breaking()
                raise GradcheckError((
                    'Expected all input grads to be undefined or zero when all output grads are undefined '
                    'or zero. Please look at "Notes about undefined output gradients" in '
                    '"tools/autograd/derivatives.yaml"'))
        return True

    # All backward functions must work properly if all output grads are undefined
    outputs_to_check = [[
        torch._C._functions.UndefinedGrad()(o) for o in _differentiable_outputs(func(*inputs))
        # This check filters out Tensor-likes that aren't instances of Tensor.
        if isinstance(o, torch.Tensor)
    ]]

    # If there are multiple output grads, we should be able to undef one at a time without error
    if len(outputs_to_check[0]) > 1:
        for undef_grad_idx in range(len(outputs)):
            output_to_check = _differentiable_outputs(func(*inputs))
            outputs_to_check.append([
                torch._C._functions.UndefinedGrad()(o) if idx == undef_grad_idx else o
                for idx, o in enumerate(output_to_check)])

    return all(check_undefined_grad_support(output) for output in outputs_to_check)


def _as_tuple(x):
    if isinstance(x, tuple):
        return x
    elif isinstance(x, list):
        return tuple(x)
    else:
        return x,


def _differentiable_outputs(x):
    return tuple(o for o in _as_tuple(x) if o.requires_grad)


def _get_notallclose_msg(analytical, numerical, output_idx, input_idx, complex_indices,
                         test_imag=False, is_forward_ad=False) -> str:
    out_is_complex = (not is_forward_ad) and complex_indices and output_idx in complex_indices
    inp_is_complex = is_forward_ad and complex_indices and input_idx in complex_indices
    part = "imaginary" if test_imag else "real"
    element = "inputs" if is_forward_ad else "outputs"
    prefix = "" if not (out_is_complex or inp_is_complex) else \
        f"While considering the {part} part of complex {element} only, "
    mode = "computed with forward mode " if is_forward_ad else ""
    return prefix + 'Jacobian %smismatch for output %d with respect to input %d,\n' \
        'numerical:%s\nanalytical:%s\n' % (mode, output_idx, input_idx, numerical, analytical)


def _transpose(matrix_of_tensors):
    # returns list of tuples
    return list(zip(*matrix_of_tensors))


def _real_and_imag_output(fn):
    # returns new functions real(fn), and imag(fn) where real(fn) and imag(fn) behave the same as
    # the original fn, except torch.real or torch.imag are applied to the complex outputs
    def apply_to_c_outs(fn, fn_to_apply):
        def wrapped_fn(*inputs):
            outs = _as_tuple(fn(*inputs))
            return tuple(fn_to_apply(o) if o.is_complex() else o for o in outs)
        return wrapped_fn

    return apply_to_c_outs(fn, torch.real), apply_to_c_outs(fn, torch.imag)

def _real_and_imag_input(fn, complex_inp_indices):
    # returns new functions that take real inputs instead of complex inputs and compute fn(x + 0 * 1j)
    # and f(x * 1j).
    def apply_to_c_inps(fn, fn_to_apply):
        def wrapped_fn(*inputs):
            new_inputs = list(inputs)
            for should_be_complex in complex_inp_indices:
                new_inputs[should_be_complex] = fn_to_apply(new_inputs[should_be_complex])
            return _as_tuple(fn(*new_inputs))
        return wrapped_fn
    return apply_to_c_inps(fn, lambda x: x + 0 * 1j), apply_to_c_inps(fn, lambda x: x * 1j)


def _gradcheck_real_imag(gradcheck_fn, func, func_out, tupled_inputs, outputs, eps, rtol,
                         atol, check_grad_dtypes, check_forward_ad, nondet_tol):
    complex_out_indices = [i for i, o in enumerate(outputs) if o.is_complex()]
    has_any_complex_output = any(o.is_complex() for o in _as_tuple(func_out))
    if has_any_complex_output:
        real_fn, imag_fn = _real_and_imag_output(func)

        imag_func_out = imag_fn(*tupled_inputs)
        imag_outputs = _differentiable_outputs(imag_func_out)
        gradcheck_fn(imag_fn, imag_func_out, tupled_inputs, imag_outputs, eps,
                     rtol, atol, check_grad_dtypes, nondet_tol,
                     complex_indices=complex_out_indices, test_imag=True)

        real_func_out = real_fn(*tupled_inputs)
        real_outputs = _differentiable_outputs(real_func_out)
        gradcheck_fn(real_fn, real_func_out, tupled_inputs, real_outputs, eps,
                     rtol, atol, check_grad_dtypes, nondet_tol, complex_indices=complex_out_indices)
    else:
        gradcheck_fn(func, func_out, tupled_inputs, outputs, eps,
                     rtol, atol, check_grad_dtypes, nondet_tol)

    if check_forward_ad:
        complex_inp_indices = [i for i, inp in enumerate(tupled_inputs) if is_tensor_like(inp) and inp.is_complex()]
        if complex_inp_indices:
            real_fn, imag_fn = _real_and_imag_input(func, complex_inp_indices)

            imag_inputs = [inp.imag if is_tensor_like(inp) and inp.is_complex() else inp for inp in tupled_inputs]
            imag_func_out = imag_fn(*imag_inputs)
            diff_imag_func_out = _differentiable_outputs(imag_func_out)
            gradcheck_fn(imag_fn, imag_func_out, imag_inputs, diff_imag_func_out, eps,
                         rtol, atol, check_grad_dtypes, nondet_tol,
                         complex_indices=complex_inp_indices, test_imag=True, use_forward_ad=True)

            real_inputs = [inp.real if is_tensor_like(inp) and inp.is_complex() else inp for inp in tupled_inputs]
            real_func_out = real_fn(*real_inputs)
            diff_real_func_out = _differentiable_outputs(real_func_out)
            gradcheck_fn(real_fn, real_func_out, real_inputs, diff_real_func_out, eps,
                         rtol, atol, check_grad_dtypes, nondet_tol, complex_indices=complex_inp_indices,
                         use_forward_ad=True)
        else:
            gradcheck_fn(func, func_out, tupled_inputs, outputs, eps,
                         rtol, atol, check_grad_dtypes, nondet_tol, use_forward_ad=True)


def _slow_gradcheck(func, func_out, tupled_inputs, outputs, eps, rtol, atol, check_grad_dtypes,
                    nondet_tol, *, use_forward_ad=False, complex_indices=None, test_imag=False):
    func_out = _as_tuple(func_out)
    if not outputs:
        return _check_no_differentiable_outputs(func, tupled_inputs, func_out, eps)

    numerical = _transpose(_get_numerical_jacobian(func, tupled_inputs, outputs, eps=eps, is_forward_ad=use_forward_ad))

    if use_forward_ad:
        analytical_forward = _get_analytical_jacobian_forward_ad(func, tupled_inputs, func_out, check_grad_dtypes=check_grad_dtypes)

        for i, n_per_out in enumerate(numerical):
            for j, n in enumerate(n_per_out):
                a = analytical_forward[j][i]
                if not _allclose_with_type_promotion(a, n.to(a.device), rtol, atol):
                    raise GradcheckError(_get_notallclose_msg(a, n, i, j, complex_indices, test_imag,
                                                              is_forward_ad=True))
    else:
        for i, o in enumerate(outputs):
            analytical = _check_analytical_jacobian_attributes(tupled_inputs, o, nondet_tol, check_grad_dtypes)

            for j, (a, n) in enumerate(zip(analytical, numerical[i])):
                if not _allclose_with_type_promotion(a, n.to(a.device), rtol, atol):
                    raise GradcheckError(_get_notallclose_msg(a, n, i, j, complex_indices, test_imag))

    return True


def _dot_with_type_promotion(u, v):
    assert u.dim() == 1 and v.dim() == 1
    return (u * v).sum()


def _allclose_with_type_promotion(a, b, rtol, atol):
    promoted_type = torch.promote_types(a.dtype, b.dtype)
    a = a.to(dtype=promoted_type)
    b = b.to(dtype=promoted_type)
    return torch.allclose(a, b, rtol, atol)


def _to_real_dtype(dtype):
    if dtype == torch.complex128:
        return torch.float64
    elif dtype == torch.complex64:
        return torch.float32
    else:
        return dtype

def _vec_from_tensor(x, generator, downcast_complex=False):
    # Create a random vector with the same number of elements as x and the same
    # dtype/device. If x is complex and downcast_complex is False, we create a
    # complex tensor with only real component.
    if x.layout == torch.sparse_coo:
        # For sparse, create a random sparse vec with random values in the same
        # indices. Make sure size is set so that it isn't inferred to be smaller.
        x_values = x._values()
        dtype = _to_real_dtype(x.dtype) if downcast_complex else x.dtype
        values = torch.rand(x_values.numel(), generator=generator) \
            .to(dtype=dtype, device=x.device) \
            .reshape(x_values.shape)
        values /= values.norm()
        vec = torch.sparse_coo_tensor(x._indices(), values, x.size())
    else:
        dtype = _to_real_dtype(x.dtype) if downcast_complex else x.dtype
        vec = torch.rand(x.numel(), generator=generator).to(dtype=dtype, device=x.device)
        vec /= vec.norm()
    return vec


def _get_inp_tensors(tupled_inputs):
    inp_idx_tup = [(i, t) for i, t in enumerate(tupled_inputs) if is_tensor_like(t) and t.requires_grad]
    return [tup[0] for tup in inp_idx_tup], [tup[1] for tup in inp_idx_tup]


def _adjusted_atol(atol, u, v):
    # In slow gradcheck, we compare A and B element-wise, i.e., for some a, b we
    # allow: |a - b| < atol + rtol * b. But since we now compare q1 = v^T A u and
    # q2 = v^T B u, we must allow |q1 - q2| < v^T E u + rtol * v^T B u, where E is
    # the correctly sized matrix in which each entry is atol.
    #
    # We see that atol needs to be scaled by v^T M u (where M is an all-ones M x N
    # matrix): v^T M u = \sum_{i} \sum_{j} u_i * v_j = (\sum_{i} u_i)(\sum_{i} v_i)
    # TODO: properly handle case when u is tuple instead of only taking first element
    u = u[0] if isinstance(u, tuple) else u
    sum_u = torch.sparse.sum(u) if u.layout == torch.sparse_coo else u.sum()
    sum_v = 1. if v is None else torch.sparse.sum(v) if v.layout == torch.sparse_coo else v.sum()
    return atol * float(sum_u) * float(sum_v)


FAST_FAIL_SLOW_OK_MSG = """
Fast gradcheck failed but element-wise differences are small. This means that the
test might've passed in slow_mode!

If you are adding a new operator, please file an issue and then use one of the
workarounds. The workaround depends on how your test invokes gradcheck/gradgradcheck:

If the test
- manually invokes gradcheck/gradgradcheck, then call gradcheck/gradgradcheck
  with `fast_mode=False` as a keyword argument.
- is OpInfo-based (e.g., in test_ops.py), then modify the OpInfo for the test
  to have `gradcheck_fast_mode=False`
- is a Module test (e.g., in common_nn.py), then modify the corresponding
  module_test entry to have `gradcheck_fast_mode=False`
""".strip()


def _run_slow_mode_and_get_error(func, tupled_inputs, outputs, input_idx, output_idx, rtol, atol, is_forward_ad):
    # Compute jacobians in slow mode for better error message
    slow_numerical = _get_numerical_jacobian(func, tupled_inputs, outputs, is_forward_ad=is_forward_ad)[input_idx][output_idx]
    if is_forward_ad:
        def new_fn(inp):
            new_inputs = list(tupled_inputs)
            new_inputs[input_idx] = inp
            return _as_tuple(func(*new_inputs))[output_idx]
        slow_analytical = _get_analytical_jacobian_forward_ad(new_fn, (tupled_inputs[input_idx],), (outputs[output_idx],))[0][0]
    else:
        slow_analytical = _get_analytical_jacobian(tupled_inputs, outputs, input_idx, output_idx)


    # Assume jacobians are non-empty and have the same shape
    slow_max_diff = (slow_numerical - slow_analytical).abs().max()

    slow_allclose = torch.allclose(slow_analytical, slow_numerical, rtol, atol)
    msg = ("\nThe above quantities relating the numerical and analytical jacobians are computed \n"
           "in fast mode. See: https://github.com/pytorch/pytorch/issues/53876 for more background \n"
           "about fast mode. Below, we recompute numerical and analytical jacobians in slow mode:\n\n"
           f"Numerical:\n {slow_numerical}\n"
           f"Analytical:\n{slow_analytical}\n\n"
           f"The max per-element difference (slow mode) is: {slow_max_diff}.\n")
    if slow_allclose:
        # Slow gradcheck would've passed!
        msg += FAST_FAIL_SLOW_OK_MSG
    return msg


def _to_flat_dense_if_sparse(tensor):
    if tensor.layout == torch.sparse_coo:
        return tensor.to_dense().reshape(-1)
    else:
        return tensor


def _make_vectors(inp_tensors, outputs, *, use_forward_ad):
    # Use our own generator to avoid messing with the user's RNG state
    g_cpu = torch.Generator()
    all_u = []
    all_u_dense = []
    for inp in inp_tensors:
        ur = _vec_from_tensor(inp, g_cpu, True)
        ur_dense = _to_flat_dense_if_sparse(ur)
        if inp.is_complex():
            ui = _vec_from_tensor(inp, g_cpu, True)
            all_u.append((ur, ui))
            ui_dense = _to_flat_dense_if_sparse(ui)
            all_u_dense.append((ur_dense, ui_dense))
        else:
            all_u.append(ur)
            all_u_dense.append(ur_dense)
    all_v = None if use_forward_ad else [_vec_from_tensor(out, g_cpu) for out in outputs]
    return all_v, all_u, all_u_dense


def _check_analytical_numerical_equal(all_analytical, all_numerical, complex_indices, tupled_inputs, outputs,
                                      func, all_v, all_u, rtol, atol, test_imag, *, is_forward_ad=False):
    for i, all_numerical_for_input_i in enumerate(all_numerical):
        for j, n in enumerate(all_numerical_for_input_i):
            # Forward AD generates the transpose of what this function expects
            if is_forward_ad:
                a = all_analytical[i][j]
            else:
                a = all_analytical[j][i]
            n = n.to(device=a.device)
            updated_atol = _adjusted_atol(atol, all_u[i], all_v[j] if all_v else None)
            if not _allclose_with_type_promotion(a, n.to(a.device), rtol, updated_atol):
                jacobians_str = _run_slow_mode_and_get_error(func, tupled_inputs, outputs, i, j, rtol, atol, is_forward_ad)
                raise GradcheckError(_get_notallclose_msg(a, n, j, i, complex_indices, test_imag, is_forward_ad) + jacobians_str)


def _fast_gradcheck(func, func_out, inputs, outputs, eps, rtol,
                    atol, check_grad_dtypes, nondet_tol, *, use_forward_ad=False, complex_indices=None, test_imag=False):
    # See https://github.com/pytorch/pytorch/issues/53876 for details
    inp_tensors_idx, inp_tensors = _get_inp_tensors(inputs)
    all_v, all_u, all_u_dense = _make_vectors(inp_tensors, outputs, use_forward_ad=use_forward_ad)

    numerical_vJu = _get_numerical_vJu(func, inputs, inp_tensors_idx, func_out, all_u, all_v, eps, is_forward_ad=use_forward_ad)
    if use_forward_ad:
        assert all_v is None
        analytical_vJu = _get_analytical_jacobian_forward_ad(func, inputs, _as_tuple(func_out),
                                                             all_u=all_u, check_grad_dtypes=check_grad_dtypes)
    else:
        if not outputs:
            _check_no_differentiable_outputs_fast(func, func_out, inputs, inp_tensors_idx, all_u, eps, nondet_tol)

        analytical_vJu = _get_analytical_vJu_backward_mode(inputs, outputs, nondet_tol, check_grad_dtypes, all_v, all_u_dense)

    _check_analytical_numerical_equal(analytical_vJu, numerical_vJu, complex_indices,
                                      inputs, outputs, func, all_v, all_u, rtol, atol, test_imag, is_forward_ad=use_forward_ad)

    return True


# Note [VarArg of Tensors]
# ~~~~~~~~~~~~~~~~~~~~~~~~
# 'func' accepts a vararg of tensors, which isn't expressable in the type system at the moment.
# If https://mypy.readthedocs.io/en/latest/additional_features.html?highlight=callable#extended-callable-types is accepted,
# the '...' first argument of Callable can be replaced with VarArg(Tensor).
# For now, we permit any input.
# the '...' first argument of Callable can be replaced with VarArg(Tensor).
# For now, we permit any input.
[docs]def gradcheck(
    func: Callable[..., Union[_TensorOrTensors]],  # See Note [VarArg of Tensors]
    inputs: _TensorOrTensors,
    *,
    eps: float = 1e-6,
    atol: float = 1e-5,
    rtol: float = 1e-3,
    raise_exception: bool = True,
    check_sparse_nnz: bool = False,
    nondet_tol: float = 0.0,
    check_undefined_grad: bool = True,
    check_grad_dtypes: bool = False,
    check_batched_grad: bool = False,
    check_forward_ad: bool = False,
    fast_mode: bool = False,
) -> bool:
    r"""Check gradients computed via small finite differences against analytical
    gradients w.r.t. tensors in :attr:`inputs` that are of floating point or complex type
    and with ``requires_grad=True``.

    The check between numerical and analytical gradients uses :func:`~torch.allclose`.

    For most of the complex functions we consider for optimization purposes, no notion of
    Jacobian exists. Instead, gradcheck verifies if the numerical and analytical values of
    the Wirtinger and Conjugate Wirtinger derivatives are consistent. Because the gradient
    computation is done under the assumption that the overall function has a real-valued
    output, we treat functions with complex output in a special way. For these functions,
    gradcheck is applied to two real-valued functions corresponding to taking the real
    components of the complex outputs for the first, and taking the imaginary components
    of the complex outputs for the second. For more details, check out
    :ref:`complex_autograd-doc`.

    .. note::
        The default values are designed for :attr:`input` of double precision.
        This check will likely fail if :attr:`input` is of less precision, e.g.,
        ``FloatTensor``.

    .. warning::
       If any checked tensor in :attr:`input` has overlapping memory, i.e.,
       different indices pointing to the same memory address (e.g., from
       :func:`torch.expand`), this check will likely fail because the numerical
       gradients computed by point perturbation at such indices will change
       values at all other indices that share the same memory address.

    Args:
        func (function): a Python function that takes Tensor inputs and returns
            a Tensor or a tuple of Tensors
        inputs (tuple of Tensor or Tensor): inputs to the function
        eps (float, optional): perturbation for finite differences
        atol (float, optional): absolute tolerance
        rtol (float, optional): relative tolerance
        raise_exception (bool, optional): indicating whether to raise an exception if
            the check fails. The exception gives more information about the
            exact nature of the failure. This is helpful when debugging gradchecks.
        check_sparse_nnz (bool, optional): if True, gradcheck allows for SparseTensor input,
            and for any SparseTensor at input, gradcheck will perform check at nnz positions only.
        nondet_tol (float, optional): tolerance for non-determinism. When running
            identical inputs through the differentiation, the results must either match
            exactly (default, 0.0) or be within this tolerance.
        check_undefined_grad (bool, optional): if True, check if undefined output grads
            are supported and treated as zeros, for ``Tensor`` outputs.
        check_batched_grad (bool, optional): if True, check if we can compute
            batched gradients using prototype vmap support. Defaults to False.
        check_forward_ad (bool, optional): if True, check that the gradients computed with forward
            mode AD match the numerical ones. Defaults to False.
        fast_mode (bool, optional): Fast mode for gradcheck and gradgradcheck is currently only
            implemented for R to R functions. If none of the inputs and outputs are complex
            a faster implementation of gradcheck that no longer computes the entire jacobian
            is run; otherwise, we fall back to the slow implementation.

    Returns:
        True if all differences satisfy allclose condition
    """
    # This is just a wrapper that handles the raise_exception logic
    args = locals().copy()
    args.pop("raise_exception")
    if not raise_exception:
        try:
            return _gradcheck_helper(**args)
        except GradcheckError as e:
            return False
    else:
        return _gradcheck_helper(**args)


def _gradcheck_helper(func, inputs, eps, atol, rtol, check_sparse_nnz, nondet_tol, check_undefined_grad,
                      check_grad_dtypes, check_batched_grad, check_forward_ad, fast_mode):
    tupled_inputs = _as_tuple(inputs)
    _check_inputs(tupled_inputs, check_sparse_nnz)

    func_out = func(*tupled_inputs)
    outputs = _differentiable_outputs(func_out)
    _check_outputs(outputs)

    gradcheck_fn = _fast_gradcheck if fast_mode else _slow_gradcheck
    _gradcheck_real_imag(gradcheck_fn, func, func_out, tupled_inputs, outputs, eps,
                         rtol, atol, check_grad_dtypes, check_forward_ad=check_forward_ad, nondet_tol=nondet_tol)

    for i, o in enumerate(outputs):
        if check_batched_grad:
            _test_batched_grad(tupled_inputs, o, i)

    _test_backward_mul_by_grad_output(outputs, tupled_inputs, check_sparse_nnz)

    if check_undefined_grad:
        _test_undefined_grad(func, outputs, tupled_inputs)
    return True


[docs]def gradgradcheck(
    func: Callable[..., _TensorOrTensors],  # See Note [VarArg of Tensors]
    inputs: _TensorOrTensors,
    grad_outputs: Optional[_TensorOrTensors] = None,
    *,
    eps: float = 1e-6,
    atol: float = 1e-5,
    rtol: float = 1e-3,
    gen_non_contig_grad_outputs: bool = False,
    raise_exception: bool = True,
    nondet_tol: float = 0.0,
    check_undefined_grad: bool = True,
    check_grad_dtypes: bool = False,
    check_batched_grad: bool = False,
    fast_mode: bool = False,
) -> bool:
    r"""Check gradients of gradients computed via small finite differences
    against analytical gradients w.r.t. tensors in :attr:`inputs` and
    :attr:`grad_outputs` that are of floating point or complex type and with
    ``requires_grad=True``.

    This function checks that backpropagating through the gradients computed
    to the given :attr:`grad_outputs` are correct.

    The check between numerical and analytical gradients uses :func:`~torch.allclose`.

    .. note::
        The default values are designed for :attr:`input` and
        :attr:`grad_outputs` of double precision. This check will likely fail if
        they are of less precision, e.g., ``FloatTensor``.

    .. warning::
       If any checked tensor in :attr:`input` and :attr:`grad_outputs` has
       overlapping memory, i.e., different indices pointing to the same memory
       address (e.g., from :func:`torch.expand`), this check will likely fail
       because the numerical gradients computed by point perturbation at such
       indices will change values at all other indices that share the same
       memory address.

    Args:
        func (function): a Python function that takes Tensor inputs and returns
            a Tensor or a tuple of Tensors
        inputs (tuple of Tensor or Tensor): inputs to the function
        grad_outputs (tuple of Tensor or Tensor, optional): The gradients with
            respect to the function's outputs.
        eps (float, optional): perturbation for finite differences
        atol (float, optional): absolute tolerance
        rtol (float, optional): relative tolerance
        gen_non_contig_gr
Shortcuts

r"""Definition of the DataLoader and associated iterators that subclass _BaseDataLoaderIter

To support these two classes, in `./_utils` we define many utility methods and
functions to be run in multiprocessing. E.g., the data loading worker loop is
in `./_utils/worker.py`.
"""

import os
import threading
import itertools
import warnings
import queue
from typing import Any, Callable, TypeVar, Generic, Sequence, List, Optional

import multiprocessing as python_multiprocessing
import torch
import torch.multiprocessing as multiprocessing
from torch._utils import ExceptionWrapper
from torch._six import string_classes

from . import IterableDataset, Sampler, SequentialSampler, RandomSampler, BatchSampler, Dataset
from . import _utils

T_co = TypeVar('T_co', covariant=True)
T = TypeVar('T')
_worker_init_fn_t = Callable[[int], None]

# Ideally we would parameterize `DataLoader` by the return type of `collate_fn`, but there is currently no way to have that
# type parameter set to a default value if the user doesn't pass in a custom 'collate_fn'.
# See https://github.com/python/mypy/issues/3737.
_collate_fn_t = Callable[[List[T]], Any]


# This function used to be defined in this file. However, it was moved to
# _utils/collate.py. Although it is rather hard to access this from user land
# (one has to explicitly directly `import torch.utils.data.dataloader`), there
# probably is user code out there using it. This aliasing maintains BC in this
# aspect.
default_collate: _collate_fn_t = _utils.collate.default_collate

get_worker_info = _utils.worker.get_worker_info

class _DatasetKind(object):
    Map = 0
    Iterable = 1

    @staticmethod
    def create_fetcher(kind, dataset, auto_collation, collate_fn, drop_last):
        if kind == _DatasetKind.Map:
            return _utils.fetch._MapDatasetFetcher(dataset, auto_collation, collate_fn, drop_last)
        else:
            return _utils.fetch._IterableDatasetFetcher(dataset, auto_collation, collate_fn, drop_last)


class _InfiniteConstantSampler(Sampler):
    r"""Analogous to ``itertools.repeat(None, None)``.
    Used as sampler for :class:`~torch.utils.data.IterableDataset`.

    Args:
        data_source (Dataset): dataset to sample from
    """

    def __init__(self):
        super(_InfiniteConstantSampler, self).__init__(None)

    def __iter__(self):
        while True:
            yield None


[docs]class DataLoader(Generic[T_co]):
    r"""
    Data loader. Combines a dataset and a sampler, and provides an iterable over
    the given dataset.

    The :class:`~torch.utils.data.DataLoader` supports both map-style and
    iterable-style datasets with single- or multi-process loading, customizing
    loading order and optional automatic batching (collation) and memory pinning.

    See :py:mod:`torch.utils.data` documentation page for more details.

    Args:
        dataset (Dataset): dataset from which to load the data.
        batch_size (int, optional): how many samples per batch to load
            (default: ``1``).
        shuffle (bool, optional): set to ``True`` to have the data reshuffled
            at every epoch (default: ``False``).
        sampler (Sampler or Iterable, optional): defines the strategy to draw
            samples from the dataset. Can be any ``Iterable`` with ``__len__``
            implemented. If specified, :attr:`shuffle` must not be specified.
        batch_sampler (Sampler or Iterable, optional): like :attr:`sampler`, but
            returns a batch of indices at a time. Mutually exclusive with
            :attr:`batch_size`, :attr:`shuffle`, :attr:`sampler`,
            and :attr:`drop_last`.
        num_workers (int, optional): how many subprocesses to use for data
            loading. ``0`` means that the data will be loaded in the main process.
            (default: ``0``)
        collate_fn (callable, optional): merges a list of samples to form a
            mini-batch of Tensor(s).  Used when using batched loading from a
            map-style dataset.
        pin_memory (bool, optional): If ``True``, the data loader will copy Tensors
            into CUDA pinned memory before returning them.  If your data elements
            are a custom type, or your :attr:`collate_fn` returns a batch that is a custom type,
            see the example below.
        drop_last (bool, optional): set to ``True`` to drop the last incomplete batch,
            if the dataset size is not divisible by the batch size. If ``False`` and
            the size of dataset is not divisible by the batch size, then the last batch
            will be smaller. (default: ``False``)
        timeout (numeric, optional): if positive, the timeout value for collecting a batch
            from workers. Should always be non-negative. (default: ``0``)
        worker_init_fn (callable, optional): If not ``None``, this will be called on each
            worker subprocess with the worker id (an int in ``[0, num_workers - 1]``) as
            input, after seeding and before data loading. (default: ``None``)
        generator (torch.Generator, optional): If not ``None``, this RNG will be used
            by RandomSampler to generate random indexes and multiprocessing to generate
            `base_seed` for workers. (default: ``None``)
        prefetch_factor (int, optional, keyword-only arg): Number of samples loaded
            in advance by each worker. ``2`` means there will be a total of
            2 * num_workers samples prefetched across all workers. (default: ``2``)
        persistent_workers (bool, optional): If ``True``, the data loader will not shutdown
            the worker processes after a dataset has been consumed once. This allows to
            maintain the workers `Dataset` instances alive. (default: ``False``)


    .. warning:: If the ``spawn`` start method is used, :attr:`worker_init_fn`
                 cannot be an unpicklable object, e.g., a lambda function. See
                 :ref:`multiprocessing-best-practices` on more details related
                 to multiprocessing in PyTorch.

    .. warning:: ``len(dataloader)`` heuristic is based on the length of the sampler used.
                 When :attr:`dataset` is an :class:`~torch.utils.data.IterableDataset`,
                 it instead returns an estimate based on ``len(dataset) / batch_size``, with proper
                 rounding depending on :attr:`drop_last`, regardless of multi-process loading
                 configurations. This represents the best guess PyTorch can make because PyTorch
                 trusts user :attr:`dataset` code in correctly handling multi-process
                 loading to avoid duplicate data.

                 However, if sharding results in multiple workers having incomplete last batches,
                 this estimate can still be inaccurate, because (1) an otherwise complete batch can
                 be broken into multiple ones and (2) more than one batch worth of samples can be
                 dropped when :attr:`drop_last` is set. Unfortunately, PyTorch can not detect such
                 cases in general.

                 See `Dataset Types`_ for more details on these two types of datasets and how
                 :class:`~torch.utils.data.IterableDataset` interacts with
                 `Multi-process data loading`_.

    .. warning:: See :ref:`reproducibility`, and :ref:`dataloader-workers-random-seed`, and
                 :ref:`data-loading-randomness` notes for random seed related questions.
    """
    dataset: Dataset[T_co]
    batch_size: Optional[int]
    num_workers: int
    pin_memory: bool
    drop_last: bool
    timeout: float
    sampler: Sampler
    prefetch_factor: int
    _iterator : Optional['_BaseDataLoaderIter']
    __initialized = False

    def __init__(self, dataset: Dataset[T_co], batch_size: Optional[int] = 1,
                 shuffle: bool = False, sampler: Optional[Sampler] = None,
                 batch_sampler: Optional[Sampler[Sequence]] = None,
                 num_workers: int = 0, collate_fn: Optional[_collate_fn_t] = None,
                 pin_memory: bool = False, drop_last: bool = False,
                 timeout: float = 0, worker_init_fn: Optional[_worker_init_fn_t] = None,
                 multiprocessing_context=None, generator=None,
                 *, prefetch_factor: int = 2,
                 persistent_workers: bool = False):
        torch._C._log_api_usage_once("python.data_loader")

        if num_workers < 0:
            raise ValueError('num_workers option should be non-negative; '
                             'use num_workers=0 to disable multiprocessing.')

        if timeout < 0:
            raise ValueError('timeout option should be non-negative')

        if num_workers == 0 and prefetch_factor != 2:
            raise ValueError('prefetch_factor option could only be specified in multiprocessing.'
                             'let num_workers > 0 to enable multiprocessing.')
        assert prefetch_factor > 0

        if persistent_workers and num_workers == 0:
            raise ValueError('persistent_workers option needs num_workers > 0')

        self.dataset = dataset
        self.num_workers = num_workers
        self.prefetch_factor = prefetch_factor
        self.pin_memory = pin_memory
        self.timeout = timeout
        self.worker_init_fn = worker_init_fn
        self.multiprocessing_context = multiprocessing_context

        # Arg-check dataset related before checking samplers because we want to
        # tell users that iterable-style datasets are incompatible with custom
        # samplers first, so that they don't learn that this combo doesn't work
        # after spending time fixing the custom sampler errors.
        if isinstance(dataset, IterableDataset):
            self._dataset_kind = _DatasetKind.Iterable
            # NOTE [ Custom Samplers and IterableDataset ]
            #
            # `IterableDataset` does not support custom `batch_sampler` or
            # `sampler` since the key is irrelevant (unless we support
            # generator-style dataset one day...).
            #
            # For `sampler`, we always create a dummy sampler. This is an
            # infinite sampler even when the dataset may have an implemented
            # finite `__len__` because in multi-process data loading, naive
            # settings will return duplicated data (which may be desired), and
            # thus using a sampler with length matching that of dataset will
            # cause data lost (you may have duplicates of the first couple
            # batches, but never see anything afterwards). Therefore,
            # `Iterabledataset` always uses an infinite sampler, an instance of
            # `_InfiniteConstantSampler` defined above.
            #
            # A custom `batch_sampler` essentially only controls the batch size.
            # However, it is unclear how useful it would be since an iterable-style
            # dataset can handle that within itself. Moreover, it is pointless
            # in multi-process data loading as the assignment order of batches
            # to workers is an implementation detail so users can not control
            # how to batchify each worker's iterable. Thus, we disable this
            # option. If this turns out to be useful in future, we can re-enable
            # this, and support custom samplers that specify the assignments to
            # specific workers.
            if shuffle is not False:
                raise ValueError(
                    "DataLoader with IterableDataset: expected unspecified "
                    "shuffle option, but got shuffle={}".format(shuffle))
            elif sampler is not None:
                # See NOTE [ Custom Samplers and IterableDataset ]
                raise ValueError(
                    "DataLoader with IterableDataset: expected unspecified "
                    "sampler option, but got sampler={}".format(sampler))
            elif batch_sampler is not None:
                # See NOTE [ Custom Samplers and IterableDataset ]
                raise ValueError(
                    "DataLoader with IterableDataset: expected unspecified "
                    "batch_sampler option, but got batch_sampler={}".format(batch_sampler))
        else:
            self._dataset_kind = _DatasetKind.Map

        if sampler is not None and shuffle:
            raise ValueError('sampler option is mutually exclusive with '
                             'shuffle')

        if batch_sampler is not None:
            # auto_collation with custom batch_sampler
            if batch_size != 1 or shuffle or sampler is not None or drop_last:
                raise ValueError('batch_sampler option is mutually exclusive '
                                 'with batch_size, shuffle, sampler, and '
                                 'drop_last')
            batch_size = None
            drop_last = False
        elif batch_size is None:
            # no auto_collation
            if drop_last:
                raise ValueError('batch_size=None option disables auto-batching '
                                 'and is mutually exclusive with drop_last')

        if sampler is None:  # give default samplers
            if self._dataset_kind == _DatasetKind.Iterable:
                # See NOTE [ Custom Samplers and IterableDataset ]
                sampler = _InfiniteConstantSampler()
            else:  # map-style
                if shuffle:
                    sampler = RandomSampler(dataset, generator=generator)
                else:
                    sampler = SequentialSampler(dataset)

        if batch_size is not None and batch_sampler is None:
            # auto_collation without custom batch_sampler
            batch_sampler = BatchSampler(sampler, batch_size, drop_last)

        self.batch_size = batch_size
        self.drop_last = drop_last
        self.sampler = sampler
        self.batch_sampler = batch_sampler
        self.generator = generator

        if collate_fn is None:
            if self._auto_collation:
                collate_fn = _utils.collate.default_collate
            else:
                collate_fn = _utils.collate.default_convert

        self.collate_fn = collate_fn
        self.persistent_workers = persistent_workers

        self.__initialized = True
        self._IterableDataset_len_called = None  # See NOTE [ IterableDataset and __len__ ]

        self._iterator = None

        self.check_worker_number_rationality()

        torch.set_vital('Dataloader', 'enabled', 'True')  # type: ignore[attr-defined]

    def _get_iterator(self) -> '_BaseDataLoaderIter':
        if self.num_workers == 0:
            return _SingleProcessDataLoaderIter(self)
        else:
            self.check_worker_number_rationality()
            return _MultiProcessingDataLoaderIter(self)

    @property
    def multiprocessing_context(self):
        return self.__multiprocessing_context

    @multiprocessing_context.setter
    def multiprocessing_context(self, multiprocessing_context):
        if multiprocessing_context is not None:
            if self.num_workers > 0:
                if isinstance(multiprocessing_context, string_classes):
                    valid_start_methods = multiprocessing.get_all_start_methods()
                    if multiprocessing_context not in valid_start_methods:
                        raise ValueError(
                            ('multiprocessing_context option '
                             'should specify a valid start method in {!r}, but got '
                             'multiprocessing_context={!r}').format(valid_start_methods, multiprocessing_context))
                    # error: Argument 1 to "get_context" has incompatible type "Union[str, bytes]"; expected "str"  [arg-type]
                    multiprocessing_context = multiprocessing.get_context(multiprocessing_context)  # type: ignore[arg-type]

                if not isinstance(multiprocessing_context, python_multiprocessing.context.BaseContext):
                    raise TypeError(('multiprocessing_context option should be a valid context '
                                     'object or a string specifying the start method, but got '
                                     'multiprocessing_context={}').format(multiprocessing_context))
            else:
                raise ValueError(('multiprocessing_context can only be used with '
                                  'multi-process loading (num_workers > 0), but got '
                                  'num_workers={}').format(self.num_workers))

        self.__multiprocessing_context = multiprocessing_context

    def __setattr__(self, attr, val):
        if self.__initialized and attr in (
                'batch_size', 'batch_sampler', 'sampler', 'drop_last', 'dataset', 'persistent_workers'):
            raise ValueError('{} attribute should not be set after {} is '
                             'initialized'.format(attr, self.__class__.__name__))

        super(DataLoader, self).__setattr__(attr, val)

    # We quote '_BaseDataLoaderIter' since it isn't defined yet and the definition can't be moved up
    # since '_BaseDataLoaderIter' references 'DataLoader'.
    def __iter__(self) -> '_BaseDataLoaderIter':
        # When using a single worker the returned iterator should be
        # created everytime to avoid reseting its state
        # However, in the case of a multiple workers iterator
        # the iterator is only created once in the lifetime of the
        # DataLoader object so that workers can be reused
        if self.persistent_workers and self.num_workers > 0:
            if self._iterator is None:
                self._iterator = self._get_iterator()
            else:
                self._iterator._reset(self)
            return self._iterator
        else:
            return self._get_iterator()

    @property
    def _auto_collation(self):
        return self.batch_sampler is not None

    @property
    def _index_sampler(self):
        # The actual sampler used for generating indices for `_DatasetFetcher`
        # (see _utils/fetch.py) to read data at each time. This would be
        # `.batch_sampler` if in auto-collation mode, and `.sampler` otherwise.
        # We can't change `.sampler` and `.batch_sampler` attributes for BC
        # reasons.
        if self._auto_collation:
            return self.batch_sampler
        else:
            return self.sampler

    def __len__(self) -> int:
        if self._dataset_kind == _DatasetKind.Iterable:
            # NOTE [ IterableDataset and __len__ ]
            #
            # For `IterableDataset`, `__len__` could be inaccurate when one naively
            # does multi-processing data loading, since the samples will be duplicated.
            # However, no real use case should be actually using that behavior, so
            # it should count as a user error. We should generally trust user
            # code to do the proper thing (e.g., configure each replica differently
            # in `__iter__`), and give us the correct `__len__` if they choose to
            # implement it (this will still throw if the dataset does not implement
            # a `__len__`).
            #
            # To provide a further warning, we track if `__len__` was called on the
            # `DataLoader`, save the returned value in `self._len_called`, and warn
            # if the iterator ends up yielding more than this number of samples.

            # Cannot statically verify that dataset is Sized
            length = self._IterableDataset_len_called = len(self.dataset)  # type: ignore[assignment, arg-type]
            if self.batch_size is not None:  # IterableDataset doesn't allow custom sampler or batch_sampler
                from math import ceil
                if self.drop_last:
                    length = length // self.batch_size
                else:
                    length = ceil(length / self.batch_size)
            return length
        else:
            return len(self._index_sampler)

    def check_worker_number_rationality(self):
        # This function check whether the dataloader's worker number is rational based on
        # current system's resource. Current rule is that if the number of workers this
        # Dataloader will create is bigger than the number of logical cpus that is allowed to
        # use, than we will pop up a warning to let user pay attention.
        #
        # eg. If current system has 2 physical CPUs with 16 cores each. And each core support 2
        #     threads, then the total logical cpus here is 2 * 16 * 2 = 64. Let's say current
        #     DataLoader process can use half of them which is 32, then the rational max number of
        #     worker that initiated from this process is 32.
        #     Now, let's say the created DataLoader has num_works = 40, which is bigger than 32.
        #     So the warning message is triggered to notify the user to lower the worker number if
        #     necessary.
        #
        #
        # [Note] Please note that this function repects `cpuset` only when os.sched_getaffinity is
        #        available (available in most of Linux system, but not OSX and Windows).
        #        When os.sched_getaffinity is not available, os.cpu_count() is called instead, but
        #        it doesn't repect cpuset.
        #        We don't take threading into account since each worker process is single threaded
        #        at this time.
        #
        #        We don't set any threading flags (eg. OMP_NUM_THREADS, MKL_NUM_THREADS, etc)
        #        other than `torch.set_num_threads` to 1 in the worker process, if the passing
        #        in functions use 3rd party modules that rely on those threading flags to determine
        #        how many thread to create (eg. numpy, etc), then it is caller's responsibility to
        #        set those flags correctly.
        def _create_warning_msg(num_worker_suggest, num_worker_created, cpuset_checked):

            suggested_max_worker_msg = ((
                "Our suggested max number of worker in current system is {}{}, which is smaller "
                "than what this DataLoader is going to create.").format(
                    num_worker_suggest,
                    ("" if cpuset_checked else " (`cpuset` is not taken into account)"))
            ) if num_worker_suggest is not None else (
                "DataLoader is not able to compute a suggested max number of worker in current system.")

            warn_msg = (
                "This DataLoader will create {} worker processes in total. {} "
                "Please be aware that excessive worker creation might get DataLoader running slow or even freeze, "
                "lower the worker number to avoid potential slowness/freeze if necessary.").format(
                    num_worker_created,
                    suggested_max_worker_msg)
            return warn_msg

        if not self.num_workers or self.num_workers == 0:
            return

        # try to compute a suggested max number of worker based on system's resource
        max_num_worker_suggest = None
        cpuset_checked = False
        if hasattr(os, 'sched_getaffinity'):
            try:
                max_num_worker_suggest = len(os.sched_getaffinity(0))
                cpuset_checked = True
            except Exception:
                pass
        if max_num_worker_suggest is None:
            # os.cpu_count() could return Optional[int]
            # get cpu count first and check None in order to satify mypy check
            cpu_count = os.cpu_count()
            if cpu_count is not None:
                max_num_worker_suggest = cpu_count

        if max_num_worker_suggest is None:
            warnings.warn(_create_warning_msg(
                max_num_worker_suggest,
                self.num_workers,
                cpuset_checked))
            return

        if self.num_workers > max_num_worker_suggest:
            warnings.warn(_create_warning_msg(
                max_num_worker_suggest,
                self.num_workers,
                cpuset_checked))


class _BaseDataLoaderIter(object):
    def __init__(self, loader: DataLoader) -> None:
        self._dataset = loader.dataset
        self._dataset_kind = loader._dataset_kind
        self._IterableDataset_len_called = loader._IterableDataset_len_called
        self._auto_collation = loader._auto_collation
        self._drop_last = loader.drop_last
        self._index_sampler = loader._index_sampler
        self._num_workers = loader.num_workers
        self._prefetch_factor = loader.prefetch_factor
        self._pin_memory = loader.pin_memory and torch.cuda.is_available()
        self._timeout = loader.timeout
        self._collate_fn = loader.collate_fn
        self._sampler_iter = iter(self._index_sampler)
        self._base_seed = torch.empty((), dtype=torch.int64).random_(generator=loader.generator).item()
        self._persistent_workers = loader.persistent_workers
        self._num_yielded = 0
        self._profile_name = "enumerate(DataLoader)#{}.__next__".format(self.__class__.__name__)

    def __iter__(self) -> '_BaseDataLoaderIter':
        return self

    def _reset(self, loader, first_iter=False):
        self._sampler_iter = iter(self._index_sampler)
        self._num_yielded = 0
        self._IterableDataset_len_called = loader._IterableDataset_len_called

    def _next_index(self):
        return next(self._sampler_iter)  # may raise StopIteration

    def _next_data(self):
        raise NotImplementedError

    def __next__(self) -> Any:
        with torch.autograd.profiler.record_function(self._profile_name):
            if self._sampler_iter is None:
                self._reset()
            data = self._next_data()
            self._num_yielded += 1
            if self._dataset_kind == _DatasetKind.Iterable and \
                    self._IterableDataset_len_called is not None and \
                    self._num_yielded > self._IterableDataset_len_called:
                warn_msg = ("Length of IterableDataset {} was reported to be {} (when accessing len(dataloader)), but {} "
                            "samples have been fetched. ").format(self._dataset, self._IterableDataset_len_called,
                                                                  self._num_yielded)
                if self._num_workers > 0:
                    warn_msg += ("For multiprocessing data-loading, this could be caused by not properly configuring the "
                                 "IterableDataset replica at each worker. Please see "
                                 "https://pytorch.org/docs/stable/data.html#torch.utils.data.IterableDataset for examples.")
                warnings.warn(warn_msg)
            return data

    next = __next__  # Python 2 compatibility

    def __len__(self) -> int:
        return len(self._index_sampler)

    def __getstate__(self):
        # TODO: add limited pickling support for sharing an iterator
        # across multiple threads for HOGWILD.
        # Probably the best way to do this is by moving the sample pushing
        # to a separate thread and then just sharing the data queue
        # but signalling the end is tricky without a non-blocking API
        raise NotImplementedError("{} cannot be pickled", self.__class__.__name__)


class _SingleProcessDataLoaderIter(_BaseDataLoaderIter):
    def __init__(self, loader):
        super(_SingleProcessDataLoaderIter, self).__init__(loader)
        assert self._timeout == 0
        assert self._num_workers == 0

        self._dataset_fetcher = _DatasetKind.create_fetcher(
            self._dataset_kind, self._dataset, self._auto_collation, self._collate_fn, self._drop_last)

    def _next_data(self):
        index = self._next_index()  # may raise StopIteration
        data = self._dataset_fetcher.fetch(index)  # may raise StopIteration
        if self._pin_memory:
            data = _utils.pin_memory.pin_memory(data)
        return data


class _MultiProcessingDataLoaderIter(_BaseDataLoaderIter):
    r"""Iterates once over the DataLoader's dataset, as specified by the sampler"""

    # NOTE [ Data Loader Multiprocessing Shutdown Logic ]
    #
    # Preliminary:
    #
    # Our data model looks like this (queues are indicated with curly brackets):
    #
    #                main process                              ||
    #                     |                                    ||
    #               {index_queue}                              ||
    #                     |                                    ||
    #              worker processes                            ||     DATA
    #                     |                                    ||
    #            {worker_result_queue}                         ||     FLOW
    #                     |                                    ||
    #      pin_memory_thread of main process                   ||   DIRECTION
    #                     |                                    ||
    #               {data_queue}                               ||
    #                     |                                    ||
    #                data output                               \/
    #
    # P.S. `worker_result_queue` and `pin_memory_thread` part may be omitted if
    #      `pin_memory=False`.
    #
    #
    # Terminating multiprocessing logic requires very careful design. In
    # particular, we need to make sure that
    #
    #   1. The iterator gracefully exits the workers when its last reference is
    #      gone or it is depleted.
    #
    #      In this case, the workers should be gracefully exited because the
    #      main process may still need to continue to run, and we want cleaning
    #      up code in the workers to be executed (e.g., releasing GPU memory).
    #      Naturally, we implement the shutdown logic in `__del__` of
    #      DataLoaderIterator.
    #
    #      We delay the discussion on the logic in this case until later.
    #
    #   2. The iterator exits the workers when the loader process and/or worker
    #      processes exits normally or with error.
    #
    #      We set all workers and `pin_memory_thread` to have `daemon=True`.
    #
    #      You may ask, why can't we make the workers non-daemonic, and
    #      gracefully exit using the same logic as we have in `__del__` when the
    #      iterator gets deleted (see 1 above)?
    #
    #      First of all, `__del__` is **not** guaranteed to be called when
    #      interpreter exits. Even if it is called, by the time it executes,
    #      many Python core library resources may alreay be freed, and even
    #      simple things like acquiring an internal lock of a queue may hang.
    #      Therefore, in this case, we actually need to prevent `__del__` from
    #      being executed, and rely on the automatic termination of daemonic
    #      children.
    #
    #      Thus, we register an `atexit` hook that sets a global flag
    #      `_utils.python_exit_status`. Since `atexit` hooks are executed in the
    #      reverse order of registration, we are guaranteed that this flag is
    #      set before library resources we use are freed (which, at least in
    #      CPython, is done via an `atexit` handler defined in
    #      `multiprocessing/util.py`
    #      https://github.com/python/cpython/blob/c606624af8d4cb3b4a052fb263bb983b3f87585b/Lib/multiprocessing/util.py#L320-L362
    #      registered when an object requiring this mechanism is first
    #      created, e.g., `mp.Queue`
    #      https://github.com/python/cpython/blob/c606624af8d4cb3b4a052fb263bb983b3f87585b/Lib/multiprocessing/context.py#L100-L103
    #      https://github.com/python/cpython/blob/c606624af8d4cb3b4a052fb263bb983b3f87585b/Lib/multiprocessing/queues.py#L29
    #      )
    #
    #      So in `__del__`, we check if `_utils.python_exit_status` is set or
    #      `None` (freed), and perform no-op if so.
    #
    #      However, simply letting library clean-up codes run can also be bad,
    #      because such codes (i.e., `multiprocessing.util._exit_function()`)
    #      include join putting threads for `mp.Queue`, which can be blocking.
    #      Hence, the main process putting threads are called with
    #      `cancel_join_thread` at creation.  See later section
    #      [ 3b. A process won't hang when putting into a queue; ]
    #      for more details.
    #
    #      Here are two example cases where library clean-up codes can run
    #      before `__del__` is called:
    #
    #        1. If we hold onto a reference to the iterator, it more often
    #           than not tries to do `multiprocessing` library cleaning before
    #           clearing the alive referenced objects (https://github.com/pytorch/pytorch/issues/48666)
    #           and thus prevents our cleaning-up code to run first.
    #
    #        2. A similar issue araises when a `DataLoader` is used in a subprocess.
    #           When a process ends, it shuts the all its daemonic children
    #           down with a SIGTERM (instead of joining them without a timeout).
    #           Simiarly for threads, but by a different mechanism. This fact,
    #           together with a few implementation details of multiprocessing, forces
    #           us to make workers daemonic. All of our problems arise when a
    #           DataLoader is used in a subprocess, and are caused by multiprocessing
    #           code which looks more or less like this:
    #
    #               try:
    #                   your_function_using_a_dataloader()
    #               finally:
    #                   multiprocessing.util._exit_function()
    #
    #           The joining/termination mentioned above happens inside
    #           `_exit_function()`. Now, if `your_function_using_a_dataloader()`
    #           throws, the stack trace stored in the exception will prevent the
    #           frame which uses `DataLoaderIter` to be freed. If the frame has any
    #           reference to the `DataLoaderIter` (e.g., in a method of the iter),
    #           its  `__del__`, which starts the shutdown procedure, will not be
    #           called. That, in turn, means that workers aren't notified. Attempting
    #           to join in `_exit_function` will then result in a hang.
    #
    #           For context, `_exit_function` is also registered as an `atexit` call.
    #           So it is unclear to me (@ssnl) why this is needed in a finally block.
    #           The code dates back to 2008 and there is no comment on the original
    #           PEP 371 or patch https://bugs.python.org/issue3050 (containing both
    #           the finally block and the `atexit` registration) that explains this.
    #
    #
    #      Finally, another choice is to just shutdown workers with logic in 1
    #      above whenever we see an error in `next`. This isn't ideal because
    #        a. It prevents users from using try-catch to resume data loading.
    #        b. It doesn't prevent hanging if users have references to the
    #           iterator.
    #
    #   3. All processes exit if any of them die unexpectedly by fatal signals.
    #
    #      As shown above, the workers are set as daemonic children of the main
    #      process. However, automatic cleaning-up of such child processes only
    #      happens if the parent process exits gracefully (e.g., not via fatal
    #      signals like SIGKILL). So we must ensure that each process will exit
    #      even the process that should send/receive data to/from it were
    #      killed, i.e.,
    #
    #        a. A process won't hang when getting from a queue.
    #
    #           Even with carefully designed data dependencies (i.e., a `put()`
    #           always corresponding to a `get()`), hanging on `get()` can still
    #           happen when data in queue is corrupted (e.g., due to
    #           `cancel_join_thread` or unexpected exit).
    #
    #           For child exit, we set a timeout whenever we try to get data
    #           from `data_queue`, and check the workers' status on each timeout
    #           and error.
    #           See `_DataLoaderiter._get_batch()` and
    #           `_DataLoaderiter._try_get_data()` for details.
    #
    #           Additionally, for child exit on non-Windows platforms, we also
    #           register a SIGCHLD handler (which is supported on Windows) on
    #           the main process, which checks if any of the workers fail in the
    #           (Python) handler. This is more efficient and faster in detecting
    #           worker failures, compared to only using the above mechanism.
    #           See `DataLoader.cpp` and `_utils/signal_handling.py` for details.
    #
    #           For `.get()` calls where the sender(s) is not the workers, we
    #           guard them with timeouts, and check the status of the sender
    #           when timeout happens:
    #             + in the workers, the `_utils.worker.ManagerWatchdog` class
    #               checks the status of the main process.
    #             + if `pin_memory=True`, when getting from `pin_memory_thread`,
    #               check `pin_memory_thread` status periodically until `.get()`
    #               returns or see that `pin_memory_thread` died.
    #
    #        b. A process won't hang when putting into a queue;
    #
    #           We use `mp.Queue` which has a separate background thread to put
    #           objects from an unbounded buffer array. The background thread is
    #           daemonic and usually automatically joined when the process
    #           *exits*.
    #
    #           In case that the receiver has ended abruptly while
    #           reading from the pipe, the join will hang forever.  The usual
    #           solution for this in Python is calling  `q.cancel_join_thread`,
    #           which prevents automatically joining it when finalizing
    #           (exiting).
    #
    #           Nonetheless, `cancel_join_thread` must only be called when the
    #           queue is **not** going to be read from or write into by another
    #           process, because it may hold onto a lock or leave corrupted data
    #           in the queue, leading other readers/writers to hang.
    #
    #           Hence,
    #             + For worker processes, we only do so (for their output
    #               queues, i.e., `worker_result_queue`) before exiting.
    #             + For `pin_memory_thread`, its output queue `data_queue` is a
    #               `queue.Queue` that does blocking `put` if the queue is full.
    #               So there is no above problem, but as a result, in
    #               `_pin_memory_loop`, we do need to  wrap the `put` in a loop
    #               that breaks not only upon success, but also when the main
    #               process stops reading, i.e., is shutting down.
    #             + For loader process, we `cancel_join_thread()` for all
    #               `_index_queues` because the whole purpose of workers and
    #               `pin_memory_thread` is to serve the loader process.  If
    #               loader process is already exiting, we don't really care if
    #               the queues are corrupted.
    #
    #
    # Now let's get back to 1:
    #   how we gracefully exit the workers when the last reference to the
    #   iterator is gone.
    #
    # To achieve this, we implement the following logic along with the design
    # choices mentioned above:
    #
    # `workers_done_event`:
    #   A `multiprocessing.Event` shared among the main process and all worker
    #   processes. This is used to signal the workers that the iterator is
    #   shutting down. After it is set, they will not send processed data to
    #   queues anymore, and only wait for the final `None` before exiting.
    #   `done_event` isn't strictly needed. I.e., we can just check for `None`
    #   from the input queue, but it allows us to skip wasting resources
    #   processing data if we are already shutting down.
    #
    # `pin_memory_thread_done_event`:
    #   A `threading.Event` for a similar purpose to that of
    #   `workers_done_event`, but is for the `pin_memory_thread`. The reason
    #   that separate events are needed is that `pin_memory_thread` reads from
    #   the output queue of the workers. But the workers, upon seeing that
    #   `workers_done_event` is set, only wants to see the final `None`, and is
    #   not required to flush all data in the output queue (e.g., it may call
    #   `cancel_join_thread` on that queue if its `IterableDataset` iterator
    #   happens to exhaust coincidentally, which is out of the control of the
    #   main process). Thus, since we will exit `pin_memory_thread` before the
    #   workers (see below), two separete events are used.
    #
    # NOTE: In short, the protocol is that the main process will set these
    #       `done_event`s and then the corresponding processes/threads a `None`,
    #       and that they may exit at any time after receiving the `None`.
    #
    # NOTE: Using `None` as the final signal is valid, since normal data will
    #       always be a 2-tuple with the 1st element being the index of the data
    #       transferred (different from dataset index/key), and the 2nd being
    #       either the dataset key or the data sample (depending on which part
    #       of the data model the queue is at).
    #
    # [ worker processes ]
    #   While loader process is alive:
    #     Get from `index_queue`.
    #       If get anything else,
    #          Check `workers_done_event`.
    #            If set, continue to next iteration
    #                    i.e., keep getting until see the `None`, then exit.
    #            Otherwise, process data:
    #                If is fetching from an `IterableDataset` and the iterator
    #                    is exhausted, send an `_IterableDatasetStopIteration`
    #                    object to signal iteration end. The main process, upon
    #                    receiving such an object, will send `None` to this
    #                    worker and not use the corresponding `index_queue`
    #                    anymore.
    #       If timed out,
    #          No matter `workers_done_event` is set (still need to see `None`)
    #          or not, must continue to next iteration.
    #   (outside loop)
    #   If `workers_done_event` is set,  (this can be False with `IterableDataset`)
    #     `data_queue.cancel_join_thread()`.  (Everything is ending here:
    #                                          main process won't read from it;
    #                                          other workers will also call
    #                                          `cancel_join_thread`.)
    #
    # [ pin_memory_thread ]
    #   # No need to check main thread. If this thread is alive, the main loader
    #   # thread must be alive, because this thread is set as daemonic.
    #   While `pin_memory_thread_done_event` is not set:
    #     Get from `index_queue`.
    #       If timed out, continue to get in the next iteration.
    #       Otherwise, process data.
    #       While `pin_memory_thread_done_event` is not set:
    #         Put processed data to `data_queue` (a `queue.Queue` with blocking put)
    #         If timed out, continue to put in the next iteration.
    #         Otherwise, break, i.e., continuing to the out loop.
    #
    #   NOTE: we don't check the status of the main thread because
    #           1. if the process is killed by fatal signal, `pin_memory_thread`
    #              ends.
    #           2. in other cases, either the cleaning-up in __del__ or the
    #              automatic exit of daemonic thread will take care of it.
    #              This won't busy-wait either because `.get(timeout)` does not
    #              busy-wait.
    #
    # [ main process ]
    #   In the DataLoader Iter's `__del__`
    #     b. Exit `pin_memory_thread`
    #          i.   Set `pin_memory_thread_done_event`.
    #          ii   Put `None` in `worker_result_queue`.
    #          iii. Join the `pin_memory_thread`.
    #          iv.  `worker_result_queue.cancel_join_thread()`.
    #
    #     c. Exit the workers.
    #          i.   Set `workers_done_event`.
    #          ii.  Put `None` in each worker's `index_queue`.
    #          iii. Join the workers.
    #          iv.  Call `.cancel_join_thread()` on each worker's `index_queue`.
    #
    #        NOTE: (c) is better placed after (b) because it may leave corrupted
    #              data in `worker_result_queue`, which `pin_memory_thread`
    #              reads from, in which case the `pin_memory_thread` can only
    #              happen at timeing out, which is slow. Nonetheless, same thing
    #              happens if a worker is killed by signal at unfortunate times,
    #              but in other cases, we are better off having a non-corrupted
    #              `worker_result_queue` for `pin_memory_thread`.
    #
    #   NOTE: If `pin_memory=False`, there is no `pin_memory_thread` and (b)
    #         can be omitted
    #
    # NB: `done_event`s isn't strictly needed. E.g., we can just check for
    #     `None` from `index_queue`, but it allows us to skip wasting resources
    #     processing indices already in `index_queue` if we are already shutting
    #     down.

    def __init__(self, loader):
        super(_MultiProcessingDataLoaderIter, self).__init__(loader)

        assert self._num_workers > 0
        assert self._prefetch_factor > 0

        if loader.multiprocessing_context is None:
            multiprocessing_context = multiprocessing
        else:
            multiprocessing_context = loader.multiprocessing_context

        self._worker_init_fn = loader.worker_init_fn
        self._worker_queue_idx_cycle = itertools.cycle(range(self._num_workers))
        # No certainty which module multiprocessing_context is
        self._worker_result_queue = multiprocessing_context.Queue()  # type: ignore[var-annotated]
        self._worker_pids_set = False
        self._shutdown = False
        self._workers_done_event = multiprocessing_context.Event()

        self._index_queues = []
        self._workers = []
        for i in range(self._num_workers):
            # No certainty which module multiprocessing_context is
            index_queue = multiprocessing_context.Queue()  # type: ignore[var-annotated]
            # Need to `cancel_join_thread` here!
            # See sections (2) and (3b) above.
            index_queue.cancel_join_thread()
            w = multiprocessing_context.Process(
                target=_utils.worker._worker_loop,
                args=(self._dataset_kind, self._dataset, index_queue,
                      self._worker_result_queue, self._workers_done_event,
                      self._auto_collation, self._collate_fn, self._drop_last,
                      self._base_seed, self._worker_init_fn, i, self._num_workers,
                      self._persistent_workers))
            w.daemon = True
            # NB: Process.start() actually take some time as it needs to
            #     start a process and pass the arguments over via a pipe.
            #     Therefore, we only add a worker to self._workers list after
            #     it started, so that we do not call .join() if program dies
            #     before it starts, and __del__ tries to join but will get:
            #     AssertionError: can only join a started process.
            w.start()
            self._index_queues.append(index_queue)
            self._workers.append(w)

        if self._pin_memory:
            self._pin_memory_thread_done_event = threading.Event()

            # Queue is not type-annotated
            self._data_queue = queue.Queue()  # type: ignore[var-annotated]
            pin_memory_thread = threading.Thread(
                target=_utils.pin_memory._pin_memory_loop,
                args=(self._worker_result_queue, self._data_queue,
                      torch.cuda.current_device(),
                      self._pin_memory_thread_done_event))
            pin_memory_thread.daemon = True
            pin_memory_thread.start()
            # Similar to workers (see comment above), we only register
            # pin_memory_thread once it is started.
            self._pin_memory_thread = pin_memory_thread
        else:
            self._data_queue = self._worker_result_queue

        # .pid can be None only before process is spawned (not the case, so ignore)
        _utils.signal_handling._set_worker_pids(id(self), tuple(w.pid for w in self._workers))  # type: ignore[misc]
        _utils.signal_handling._set_SIGCHLD_handler()
        self._worker_pids_set = True
        self._reset(loader, first_iter=True)

    def _reset(self, loader, first_iter=False):
        super()._reset(loader, first_iter)
        self._send_idx = 0  # idx of the next task to be sent to workers
        self._rcvd_idx = 0  # idx of the next task to be returned in __next__
        # information about data not yet yielded, i.e., tasks w/ indices in range [rcvd_idx, send_idx).
        # map: task idx => - (worker_id,)        if data isn't fetched (outstanding)
        #                  \ (worker_id, data)   if data is already fetched (out-of-order)
        self._task_info = {}
        self._tasks_outstanding = 0  # always equal to count(v for v in task_info.values() if len(v) == 1)
        # A list of booleans representing whether each worker still has work to
        # do, i.e., not having exhausted its iterable dataset object. It always
        # contains all `True`s if not using an iterable-style dataset
        # (i.e., if kind != Iterable).
        # Not that this indicates that a worker still has work to do *for this epoch*.
        # It does not mean that a worker is dead. In case of `_persistent_workers`,
        # the worker will be reset to available in the next epoch.
        self._workers_status = [True for i in range(self._num_workers)]
        # We resume the prefetching in case it was enabled
        if not first_iter:
            for idx in range(self._num_workers):
                self._index_queues[idx].put(_utils.worker._ResumeIteration())
            resume_iteration_cnt = self._num_workers
            while resume_iteration_cnt > 0:
                return_idx, return_data = self._get_data()
                if isinstance(return_idx, _utils.worker._ResumeIteration):
                    assert return_data is None
                    resume_iteration_cnt -= 1
        # prime the prefetch loop
        for _ in range(self._prefetch_factor * self._num_workers):
            self._try_put_index()

    def _try_get_data(self, timeout=_utils.MP_STATUS_CHECK_INTERVAL):
        # Tries to fetch data from `self._data_queue` once for a given timeout.
        # This can also be used as inner loop of fetching without timeout, with
        # the sender status as the loop condition.
        #
        # This raises a `RuntimeError` if any worker died expectedly. This error
        # can come from either the SIGCHLD handler in `_utils/signal_handling.py`
        # (only for non-Windows platforms), or the manual check below on errors
        # and timeouts.
        #
        # Returns a 2-tuple:
        #   (bool: whether successfully get data, any: data if successful else None)
        try:
            data = self._data_queue.get(timeout=timeout)
            return (True, data)
        except Exception as e:
            # At timeout and error, we manually check whether any worker has
            # failed. Note that this is the only mechanism for Windows to detect
            # worker failures.
            failed_workers = []
            for worker_id, w in enumerate(self._workers):
                if self._workers_status[worker_id] and not w.is_alive():
                    failed_workers.append(w)
                    self._mark_worker_as_unavailable(worker_id)
            if len(failed_workers) > 0:
                pids_str = ', '.join(str(w.pid) for w in failed_workers)
                raise RuntimeError('DataLoader worker (pid(s) {}) exited unexpectedly'.format(pids_str)) from e
            if isinstance(e, queue.Empty):
                return (False, None)
            import tempfile
            import errno
            try:
                # Raise an exception if we are this close to the FDs limit.
                # Apparently, trying to open only one file is not a sufficient
                # test.
                # See NOTE [ DataLoader on Linux and open files limit ]
                fds_limit_margin = 10
                fs = [tempfile.NamedTemporaryFile() for i in range(fds_limit_margin)]
            except OSError as e:
                if e.errno == errno.EMFILE:
                    raise RuntimeError(
                        "Too many open files. Communication with the"
                        " workers is no longer possible. Please increase the"
                        " limit using `ulimit -n` in the shell or change the"
                        " sharing strategy by calling"
                        " `torch.multiprocessing.set_sharing_strategy('file_system')`"
                        " at the beginning of your code") from None
            raise

# NOTE [ DataLoader on Linux and open files limit ]
#
# On Linux when DataLoader is used with multiprocessing we pass the data between
# the root process and the workers through SHM files. We remove those files from
# the filesystem as soon as they are created and keep them alive by
# passing around their file descriptors through AF_UNIX sockets. (See
# docs/source/multiprocessing.rst and 'Multiprocessing Technical Notes` in
# the wiki (https://github.com/pytorch/pytorch/wiki).)
#
# This sometimes leads us to exceeding the open files limit. When that happens,
# and the offending file descriptor is coming over a socket, the `socket` Python
# package silently strips the file descriptor from the message, setting only the
# `MSG_CTRUNC` flag (which might be a bit misleading since the manpage says that
# it _indicates that some control data were discarded due to lack of space in
# the buffer for ancillary data_). This might reflect the C implementation of
# AF_UNIX sockets.
#
# This behaviour can be reproduced with the script and instructions at the
# bottom of this note.
#
# When that happens, the standard Python `multiprocessing` (and not
# `torch.multiprocessing`) raises a `RuntimeError: received 0 items of ancdata`
#
# Sometimes, instead of the FD being stripped, you may get an `OSError:
# Too many open files`, both in the script below and in DataLoader. However,
# this is rare and seems to be nondeterministic.
#
#
#   #!/usr/bin/env python3
#   import sys
#   import socket
#   import os
#   import array
#   import shutil
#   import socket
#
#
#   if len(sys.argv) != 4:
#       print("Usage: ", sys.argv[0], " tmp_dirname iteration (send|recv)")
#       sys.exit(1)
#
#   if __name__ == '__main__':
#       dirname = sys.argv[1]
#       sock_path = dirname + "/sock"
#       iterations = int(sys.argv[2])
#       def dummy_path(i):
#           return dirname + "/" + str(i) + ".dummy"
#
#
#       if sys.argv[3] == 'send':
#           while not os.path.exists(sock_path):
#               pass
#           client = socket.socket(socket.AF_UNIX, socket.SOCK_DGRAM)
#           client.connect(sock_path)
#           for i in range(iterations):
#               fd = os.open(dummy_path(i), os.O_WRONLY | os.O_CREAT)
#               ancdata = array.array('i', [fd])
#               msg = bytes([i % 256])
#               print("Sending fd ", fd, " (iteration #", i, ")")
#               client.sendmsg([msg], [(socket.SOL_SOCKET, socket.SCM_RIGHTS, ancdata)])
#
#
#       else:
#           assert sys.argv[3] == 'recv'
#
#           if os.path.exists(dirname):
#               raise Exception("Directory exists")
#
#           os.mkdir(dirname)
#
#           print("Opening socket...")
#           server = socket.socket(socket.AF_UNIX, socket.SOCK_DGRAM)
#           server.bind(sock_path)
#
#           print("Listening...")
#           for i in range(iterations):
#               a = array.array('i')
#               msg, ancdata, flags, addr = server.recvmsg(1, socket.CMSG_SPACE(a.itemsize))
#               assert(len(ancdata) == 1)
#               cmsg_level, cmsg_type, cmsg_data = ancdata[0]
#               a.frombytes(cmsg_data)
#               print("Received fd ", a[0], " (iteration #", i, ")")
#
#           shutil.rmtree(dirname)
#
# Steps to reproduce:
#
# 1. Run two shells and set lower file descriptor limit in the receiving one:
# (shell1) ulimit -n 1020
# (shell2) ulimit -n 1022
#
# 2. Run the script above with the `recv` option in the first shell
# (shell1) ./test_socket.py sock_tmp 1017 recv
#
# 3. Run the script with the `send` option in the second shell:
# (shell2) ./test_socket.py sock_tmp 1017 send

    def _get_data(self):
        # Fetches data from `self._data_queue`.
        #
        # We check workers' status every `MP_STATUS_CHECK_INTERVAL` seconds,
        # which we achieve by running `self._try_get_data(timeout=MP_STATUS_CHECK_INTERVAL)`
        # in a loop. This is the only mechanism to detect worker failures for
        # Windows. For other platforms, a SIGCHLD handler is also used for
        # worker failure detection.
        #
        # If `pin_memory=True`, we also need check if `pin_memory_thread` had
        # died at timeouts.
        if self._timeout > 0:
            success, data = self._try_get_data(self._timeout)
            if success:
                return data
            else:
                raise RuntimeError('DataLoader timed out after {} seconds'.format(self._timeout))
        elif self._pin_memory:
            while self._pin_memory_thread.is_alive():
                success, data = self._try_get_data()
                if success:
                    return data
            else:
                # while condition is false, i.e., pin_memory_thread died.
                raise RuntimeError('Pin memory thread exited unexpectedly')
            # In this case, `self._data_queue` is a `queue.Queue`,. But we don't
            # need to call `.task_done()` because we don't use `.join()`.
        else:
            while True:
                success, data = self._try_get_data()
                if success:
                    return data

    def _next_data(self):
        while True:
            # If the worker responsible for `self._rcvd_idx` has already ended
            # and was unable to fulfill this task (due to exhausting an `IterableDataset`),
            # we try to advance `self._rcvd_idx` to find the next valid index.
            #
            # This part needs to run in the loop because both the `self._get_data()`
            # call and `_IterableDatasetStopIteration` check below can mark
            # extra worker(s) as dead.
            while self._rcvd_idx < self._send_idx:
                info = self._task_info[self._rcvd_idx]
                worker_id = info[0]
                if len(info) == 2 or self._workers_status[worker_id]:  # has data or is still active
                    break
                del self._task_info[self._rcvd_idx]
                self._rcvd_idx += 1
            else:
                # no valid `self._rcvd_idx` is found (i.e., didn't break)
                if not self._persistent_workers:
                    self._shutdown_workers()
                raise StopIteration

            # Now `self._rcvd_idx` is the batch index we want to fetch

            # Check if the next sample has already been generated
            if len(self._task_info[self._rcvd_idx]) == 2:
                data = self._task_info.pop(self._rcvd_idx)[1]
                return self._process_data(data)

            assert not self._shutdown and self._tasks_outstanding > 0
            idx, data = self._get_data()
            self._tasks_outstanding -= 1
            if self._dataset_kind == _DatasetKind.Iterable:
                # Check for _IterableDatasetStopIteration
                if isinstance(data, _utils.worker._IterableDatasetStopIteration):
                    if self._persistent_workers:
                        self._workers_status[data.worker_id] = False
                    else:
                        self._mark_worker_as_unavailable(data.worker_id)
                    self._try_put_index()
                    continue

            if idx != self._rcvd_idx:
                # store out-of-order samples
                self._task_info[idx] += (data,)
            else:
                del self._task_info[idx]
                return self._process_data(data)

    def _try_put_index(self):
        assert self._tasks_outstanding < self._prefetch_factor * self._num_workers

        try:
            index = self._next_index()
        except StopIteration:
            return
        for _ in range(self._num_workers):  # find the next active worker, if any
            worker_queue_idx = next(self._worker_queue_idx_cycle)
            if self._workers_status[worker_queue_idx]:
                break
        else:
            # not found (i.e., didn't break)
            return

        self._index_queues[worker_queue_idx].put((self._send_idx, index))
        self._task_info[self._send_idx] = (worker_queue_idx,)
        self._tasks_outstanding += 1
        self._send_idx += 1

    def _process_data(self, data):
        self._rcvd_idx += 1
        self._try_put_index()
        if isinstance(data, ExceptionWrapper):
            data.reraise()
        return data

    def _mark_worker_as_unavailable(self, worker_id, shutdown=False):
        # Mark a worker as having finished its work e.g., due to
        # exhausting an `IterableDataset`. This should be used only when this
        # `_MultiProcessingDataLoaderIter` is going to continue running.

        assert self._workers_status[worker_id] or (self._persistent_workers and shutdown)

        # Signal termination to that specific worker.
        q = self._index_queues[worker_id]
        # Indicate that no more data will be put on this queue by the current
        # process.
        q.put(None)

        # Note that we don't actually join the worker here, nor do we remove the
        # worker's pid from C side struct because (1) joining may be slow, and
        # (2) since we don't join, the worker may still raise error, and we
        # prefer capturing those, rather than ignoring them, even though they
        # are raised after the worker has finished its job.
        # Joinning is deferred to `_shutdown_workers`, which it is called when
        # all workers finish their jobs (e.g., `IterableDataset` replicas) or
        # when this iterator is garbage collected.

        self._workers_status[worker_id] = False

        assert self._workers_done_event.is_set() == shutdown

    def _shutdown_workers(self):
        # Called when shutting down this `_MultiProcessingDataLoaderIter`.
        # See NOTE [ Data Loader Multiprocessing Shutdown Logic ] for details on
        # the logic of this function.
        python_exit_status = _utils.python_exit_status
        if python_exit_status is True or python_exit_status is None:
            # See (2) of the note. If Python is shutting down, do no-op.
            return
        # Normal exit when last reference is gone / iterator is depleted.
        # See (1) and the second half of the note.
        if not self._shutdown:
            self._shutdown = True
            try:
                # Normal exit when last reference is gone / iterator is depleted.
                # See (1) and the second half of the note.

                # Exit `pin_memory_thread` first because exiting workers may leave
                # corrupted data in `worker_result_queue` which `pin_memory_thread`
                # reads from.
                if hasattr(self, '_pin_memory_thread'):
                    # Use hasattr in case error happens before we set the attribute.
                    self._pin_memory_thread_done_event.set()
                    # Send something to pin_memory_thread in case it is waiting
                    # so that it can wake up and check `pin_memory_thread_done_event`
                    self._worker_result_queue.put((None, None))
                    self._pin_memory_thread.join()
                    self._worker_result_queue.cancel_join_thread()
                    self._worker_result_queue.close()

                # Exit workers now.
                self._workers_done_event.set()
                for worker_id in range(len(self._workers)):
                    # Get number of workers from `len(self._workers)` instead of
                    # `self._num_workers` in case we error before starting all
                    # workers.
                    # If we are using workers_status with persistent_workers
                    # we have to shut it down because the worker is paused
                    if self._persistent_workers or self._workers_status[worker_id]:
                        self._mark_worker_as_unavailable(worker_id, shutdown=True)
                for w in self._workers:
                    # We should be able to join here, but in case anything went
                    # wrong, we set a time

Recommend

How to Add Extra Information to Errors in Go Wrapping and Unwrapping Errors Unwrapping Errors with errors.Unwrap

How to Add Extra Information to Errors in Go Wrapping and Unwrapping Errors Wrapping Errors with fmt.Errorf

How to Add Extra Information to Errors in Go Handling Specific Errors Using Sentinel Errors

How to Add Extra Information to Errors in Go Returning and Handling Errors in Go

Como Instalar o Git no CentOS 7 Configurando o Git

Como Instalar o Git no CentOS 7 Instalar o Git Opção Dois — Instalar o Git a Partir do Código-fonte

Como Instalar o Git no CentOS 7 Instalar o Git Opção Um — Instalar o Git com Yum

Установка и использование Docker в Ubuntu 20.04 Шаг 8 — Загрузка образов Docker в репозиторий Docker

Установка и использование Docker в Ubuntu 20.04 Шаг 7 — Внесение изменений в контейнер для образа Docker

Установка и использование Docker в Ubuntu 20.04 Шаг 6 — Управление контейнерами Docker

Установка и использование Docker в Ubuntu 20.04 Шаг 5 — Запуск контейнеров Docker

Установка и использование Docker в Ubuntu 20.04 Шаг 4 — Работа с образами Docker

Установка и использование Docker в Ubuntu 20.04 Шаг 3 — Использование команды Docker

Установка и использование Docker в Ubuntu 20.04 Шаг 2 — Настройка команды Docker без sudo (необязательно)

Установка и использование Docker в Ubuntu 20.04 Шаг 1 — Установка Docker

Comment développer un site web Drupal 9 sur votre machine locale en utilisant Docker et DDEV Étape 6 — Gérer le conteneur DDEV

Comment développer un site web Drupal 9 sur votre machine locale en utilisant Docker et DDEV Étape 4 — Vérifier vos autorisations

Comment développer un site web Drupal 9 sur votre machine locale en utilisant Docker et DDEV Étape 3 — Configuration de votre projet Drupal 9 Option 2 — Utilisation de la ligne de commande

Comment développer un site web Drupal 9 sur votre machine locale en utilisant Docker et DDEV Étape 3 — Configuration de votre projet Drupal 9

Comment développer un site web Drupal 9 sur votre machine locale en utilisant Docker et DDEV Étape 2 — Déployer un nouveau site Drupal 9 en utilisant DDEV

Comment développer un site web Drupal 9 sur votre machine locale en utilisant Docker et DDEV Étape 1 — Installer DDEV Option 2 — Installer DDEV sur Linux