import torch
import torch._six
from typing import Optional, List, DefaultDict
import warnings
from collections import defaultdict
import sys
import traceback



def _type(self, dtype=None, non_blocking=False, **kwargs):
    """Returns the type if `dtype` is not provided, else casts this object to
    the specified type.

    If this is already of the correct type, no copy is performed and the
    original object is returned.

    Args:
        dtype (type or string): The desired type
        non_blocking (bool): If ``True``, and the source is in pinned memory
            and destination is on the GPU or vice versa, the copy is performed
            asynchronously with respect to the host. Otherwise, the argument
            has no effect.
        **kwargs: For compatibility, may contain the key ``async`` in place of
            the ``non_blocking`` argument. The ``async`` arg is deprecated.
    """
    non_blocking = _get_async_or_non_blocking('type', non_blocking, kwargs)
    if dtype is None:
        return self.__module__ + '.' + self.__class__.__name__

    if isinstance(dtype, str):
        dtype = _import_dotted_name(dtype)
    if dtype == type(self):
        return self
    if self.is_sparse:
        if not dtype.is_sparse:
            raise RuntimeError("Cannot cast sparse tensor to dense tensor")
        new_module_name = dtype.__module__.replace('.sparse', '')
        new_values_type_name = new_module_name + '.' + dtype.__name__
        new_values = torch.Tensor._values(self).type(new_values_type_name, non_blocking)
        new_indices_type_name = new_module_name + '.LongTensor'
        new_indices = torch.Tensor._indices(self).type(new_indices_type_name, non_blocking)
        return dtype(new_indices, new_values, self.size())
    if dtype.is_sparse:
        raise RuntimeError("Cannot cast dense tensor to sparse tensor")
    return dtype(self.size()).copy_(self, non_blocking)


def _cuda(self, device=None, non_blocking=False, **kwargs):
    """Returns a copy of this object in CUDA memory.

    If this object is already in CUDA memory and on the correct device, then
    no copy is performed and the original object is returned.

    Args:
        device (int): The destination GPU id. Defaults to the current device.
        non_blocking (bool): If ``True`` and the source is in pinned memory,
            the copy will be asynchronous with respect to the host. Otherwise,
            the argument has no effect.
        **kwargs: For compatibility, may contain the key ``async`` in place of
            the ``non_blocking`` argument.
    """
    non_blocking = _get_async_or_non_blocking('cuda', non_blocking, kwargs)
    if self.is_cuda:
        if device is None:
            device = torch.cuda.current_device()
        if self.get_device() == device:
            return self
    else:
        if device is None:
            device = -1
    with torch.cuda.device(device):
        if self.is_sparse:
            new_type = getattr(torch.cuda.sparse, self.__class__.__name__)
            indices = torch.Tensor._indices(self).cuda(device, non_blocking)
            values = torch.Tensor._values(self).cuda(device, non_blocking)
            return new_type(indices, values, self.size())
        else:
            new_type = getattr(torch.cuda, self.__class__.__name__)
            return new_type(self.size()).copy_(self, non_blocking)


def _get_async_or_non_blocking(function_name, non_blocking, kwargs):
    if not kwargs:
        return non_blocking
    if len(kwargs) != 1 or 'async' not in kwargs:
        message = "{}() got an unexpected keyword argument '{}'"
        argument = list(kwargs.keys()).pop()
        raise TypeError(message.format(function_name, argument))
    warnings.warn("'async' is deprecated; use 'non_blocking'")
    return kwargs['async']


# Note [Don't serialize hooks]
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Since time immemorial, we have serialized the backward hooks associated with
# variables.  This kind of half-worked--Python can pickle global functions
# (but not closures!)--but there were problems.
#
#   - It's fragile.  If you serialize a backward hook into a saved
#     model, and then you rename the function associated with the hook,
#     now your saved model is broken and you can't load it anymore.
#
#   - It's not actually used.  The standard recommendation is to
#     serialize the *state_dict* of a model, not the model itself
#     (since this is more stable to code changes affecting the model
#     serialization), and the state dict saves "data" only, thus
#     stripping the the backward hooks.  In some cases, hooks are
#     essential to the well-functioning of a model (e.g., DDP),
#     but DDP already manages readding the hooks!
#
#   - We didn't serialize them in many cases.  Prior to #10220, we
#     were dropping backward hooks in ForkingPickler.  We "fixed" this
#     to be convenient with other serialization sites, but lack of
#     serializing backward hooks wasn't actually the root cause of
#     the bug.
#
# With these cases in mind, we have decided that a better strategy
# is to just NOT serialize hooks at all.
#
# Since this is a BC-breaking change, we should warn when we previously
# serialized a hook, but no longer do so. This will be done by adding a special
# sentinel property to hooks will be used to suppress this warning. If a hook
# has the property _torch_serialize_ignore, we will not emit a warning if we
# attempt to serialize a Tensor with this hook attached to it.
#
# By the way, when _backward_hooks is skipped, we must give an EMPTY
# OrderedDict(), if you pass a None you'll run afoul #12219.


def _rebuild_tensor(storage, storage_offset, size, stride):
    # first construct a tensor with the correct dtype/device
    t = torch.tensor([], dtype=storage.dtype, device=storage.device)
    return t.set_(storage, storage_offset, size, stride)


def _rebuild_tensor_v2(storage, storage_offset, size, stride, requires_grad, backward_hooks):
    tensor = _rebuild_tensor(storage, storage_offset, size, stride)
    tensor.requires_grad = requires_grad
    # NB: This line exists only for backwards compatibility; the
    # general expectation is that backward_hooks is an empty
    # OrderedDict.  See Note [Don't serialize hooks]
    tensor._backward_hooks = backward_hooks
    return tensor


_sparse_tensors_to_validate: List["torch.Tensor"] = []

# In _legacy_load() in serialization.py we unpickle storages after the sparse
# tensors have been already unpickled. Those storages contain data necessary for
# validating sparse tensors: indices and values. That's why sparse tensors are
# first unpickled without any validation, and then this function is called just
# before _legacy_load() returns, so that all the sparse tensors can be validated
# in bulk.
#
# The same procedure must be followed by _load() in serialization.py because due
# to Pickler semantics, we have to use the same (non-validating) function for
# unpickling sparse tensors, regardless of the caller.
def _validate_loaded_sparse_tensors():
    try:
        for t in _sparse_tensors_to_validate:
            torch._validate_sparse_coo_tensor_args(t._indices(), t._values(),
                                                   t.size())
    finally:
        _sparse_tensors_to_validate.clear()

def _rebuild_sparse_tensor(layout, data):
    if layout == torch.sparse_coo:
        indices, values, size = data
        result = torch._sparse_coo_tensor_unsafe(indices, values, size)
        _sparse_tensors_to_validate.append(result)
        return result

    raise NotImplementedError("rebuilding sparse tensor for layout %s" % (layout))


def _rebuild_xla_tensor(data, dtype, device, requires_grad):
    tensor = torch.from_numpy(data).to(dtype=dtype, device=device)
    tensor.requires_grad = requires_grad
    return tensor


def _rebuild_qtensor(storage, storage_offset, size, stride, quantizer_params, requires_grad, backward_hooks):
    qscheme = quantizer_params[0]
    if qscheme == torch.per_tensor_affine:
        _, scale, zero_point = quantizer_params
        tensor = torch._empty_affine_quantized(size, scale=scale, zero_point=zero_point, dtype=storage.dtype)
    elif qscheme in (torch.per_channel_affine, torch.per_channel_affine_float_qparams):
        _, scales, zero_points, axis = quantizer_params
        if type(scales) is list and type(zero_points) is list:
            if qscheme == torch.per_channel_affine:
                scales = torch.tensor(scales, dtype=torch.double)
                zero_points = torch.tensor(zero_points, dtype=torch.long)
            else:
                scales = torch.tensor(scales, dtype=torch.float)
                zero_points = torch.tensor(zero_points, dtype=torch.float)
        tensor = torch._empty_per_channel_affine_quantized(
            size, scales=scales, zero_points=zero_points, axis=axis, dtype=storage.dtype)
    else:
        raise RuntimeError("Can't deserialize quantized tensor with qscheme {}".format(qscheme))
    tensor.set_(storage, storage_offset, size, stride)
    tensor.requires_grad = requires_grad
    # NB: This line exists only for backwards compatibility; the
    # general expectation is that backward_hooks is an empty
    # OrderedDict.  See Note [Don't serialize hooks]
    tensor._backward_hooks = backward_hooks
    return tensor

def _rebuild_parameter(data, requires_grad, backward_hooks):
    param = torch.nn.Parameter(data, requires_grad)
    # NB: This line exists only for backwards compatibility; the
    # general expectation is that backward_hooks is an empty
    # OrderedDict.  See Note [Don't serialize hooks]
    param._backward_hooks = backward_hooks

    return param


def _import_dotted_name(name):
    components = name.split('.')
    obj = __import__(components[0])
    for component in components[1:]:
        obj = getattr(obj, component)
    return obj


# Taken from python 3.5 docs
def _accumulate(iterable, fn=lambda x, y: x + y):
    'Return running totals'
    # _accumulate([1,2,3,4,5]) --> 1 3 6 10 15
    # _accumulate([1,2,3,4,5], operator.mul) --> 1 2 6 24 120
    it = iter(iterable)
    try:
        total = next(it)
    except StopIteration:
        return
    yield total
    for element in it:
        total = fn(total, element)
        yield total


def _flatten_dense_tensors(tensors):
    """Flatten dense tensors into a contiguous 1D buffer. Assume tensors are of
    same dense type.

    Since inputs are dense, the resulting tensor will be a concatenated 1D
    buffer. Element-wise operation on this buffer will be equivalent to
    operating individually.

    Args:
        tensors (Iterable[Tensor]): dense tensors to flatten.

    Returns:
        A contiguous 1D buffer containing input tensors.
    """
    if len(tensors) == 1:
        return tensors[0].contiguous().view(-1)
    flat = torch.cat([t.contiguous().view(-1) for t in tensors], dim=0)
    return flat


def _flatten_sparse_tensors(tensors):
    """Flatten sparse tensors into two contiguous 1D buffers, one of indices and
    one of values. Assume tensors are of same sparse type.

    Args:
        tensors (Iterable[Tensor]): sparse tensors to flatten.

    Returns:
        A tuple of two contiguous 1D buffers, one containing input tensors'
        indices and the other containing the values.
    """
    flat_indices = _flatten_dense_tensors([torch.Tensor._indices(t) for t in tensors])
    flat_values = _flatten_dense_tensors([torch.Tensor._values(t) for t in tensors])
    return flat_indices, flat_values


def _unflatten_dense_tensors(flat, tensors):
    """View a flat buffer using the sizes of tensors. Assume that tensors are of
    same dense type, and that flat is given by _flatten_dense_tensors.

    Args:
        flat (Tensor): flattened dense tensors to unflatten.
        tensors (Iterable[Tensor]): dense tensors whose sizes will be used to
          unflatten flat.

    Returns:
        Unflattened dense tensors with sizes same as tensors and values from
        flat.
    """
    outputs = []
    offset = 0
    for tensor in tensors:
        numel = tensor.numel()
        outputs.append(flat.narrow(0, offset, numel).view_as(tensor))
        offset += numel
    return tuple(outputs)


def _unflatten_sparse_tensors(flat, tensors):
    """View flat buffer (containing indices and values) using the sizes of
    tensors. Assume that tensors are of same sparse type, and that flat is given
    by _flatten_sparse_tensors.

    Args:
        flat (tuple(Tensor, Tensor)): flattened indices and values of sparse
          tensors to unflatten.
        tensors (Iterable[Tensor]): sparse tensors whose sizes will be used to
          unflatten flat.

    Returns:
        Unflattened sparse tensors with sizes same as tensors and values from
        flat.
    """
    flat_indices, flat_values = flat
    indices = _unflatten_dense_tensors(flat_indices, [torch.Tensor._indices(t) for t in tensors])
    values = _unflatten_dense_tensors(flat_values, [torch.Tensor._values(t) for t in tensors])
    outputs = []
    for t, i, v in zip(tensors, indices, values):
        outputs.append(t.new(i, v, t.size()))
    return tuple(outputs)


def _reorder_tensors_as(tensors, ordered_tensors):
    """Assume that tensors are of same order as ordered_tensors within their
    types, e.g., from _take_tensors. Reorder them to be of same order as
    ordered_tensors.

    Args:
        tensors (Iterable[Tensor]): tensors to be reordered. They should be of
          the same order as ordered_tensors within their own types.
        ordered_tensors (Iterable[Tensor]): tensors whose order will be the
          reference.

    Returns:
        Ordered tuple of tensors with contents from tensors and order of
        ordered_tensors.
    """
    type_dict = defaultdict(list)
    for tensor in tensors:
        type_dict[tensor.type()].append(tensor)
    type_dict_ = {t: iter(coll) for t, coll in type_dict.items()}
    return tuple(next(type_dict_[tensor.type()]) for tensor in ordered_tensors)