Source code for mindspore.common.tensor

# Copyright 2020-2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""Tensor implementation."""
import numbers
import numpy as np

from mindspore import log as logger
from mindspore.communication.management import get_rank, get_group_size
from . import dtype as mstype
from ._register_for_tensor import tensor_operator_registry
from .._c_expression import Tensor as Tensor_
from .._checkparam import Validator as validator

__all__ = ['Tensor', 'RowTensor', 'SparseTensor']
np_types = (np.int8, np.int16, np.int32, np.int64,
            np.uint8, np.uint16, np.uint32, np.uint64, np.float16,
            np.float32, np.float64, np.bool_)


[docs]class Tensor(Tensor_): """ Tensor is used for data storage. Tensor inherits tensor object in C++. Some functions are implemented in C++ and some functions are implemented in Python. Args: input_data (Tensor, float, int, bool, tuple, list, numpy.ndarray): Input data of the tensor. dtype (:class:`mindspore.dtype`): Input data should be None, bool or numeric type defined in `mindspore.dtype`. The argument is used to define the data type of the output tensor. If it is None, the data type of the output tensor will be as same as the `input_data`. Default: None. shape (Union[tuple, list, int]): A list of integers, a tuple of integers or an integer as the shape of output. If `input_data` is available, `shape` doesn't need to be set. Default: None. init (Initializer): the information of init data. 'init' is used for delayed initialization in parallel mode. Usually, it is not recommended to use 'init' interface to initialize parameters in other conditions. If 'init' interface is used to initialize parameters, the `init_data` API need to be called to convert `Tensor` to the actual data. Outputs: Tensor, with the same shape as `input_data`. Examples: >>> import numpy as np >>> import mindspore as ms >>> from mindspore.common.tensor import Tensor >>> from mindspore.common.initializer import One >>> # initialize a tensor with input data >>> t1 = Tensor(np.zeros([1, 2, 3]), ms.float32) >>> assert isinstance(t1, Tensor) >>> assert t1.shape == (1, 2, 3) >>> assert t1.dtype == ms.float32 >>> >>> # initialize a tensor with a float scalar >>> t2 = Tensor(0.1) >>> assert isinstance(t2, Tensor) >>> assert t2.dtype == ms.float64 ... >>> # initialize a tensor with init >>> t3 = Tensor(shape = (1, 3), dtype=ms.float32, init=One()) >>> assert isinstance(t3, Tensor) >>> assert t3.shape == (1, 3) >>> assert t3.dtype == ms.float32 """ delta_seed = 0 def __init__(self, input_data=None, dtype=None, shape=None, init=None): self.init_finished = False # If input data is numpy number, convert it to np array if isinstance(input_data, np_types): input_data = np.array(input_data) if input_data is not None and shape is not None: raise ValueError("If input_data is available, shape doesn't need to be set") if init is not None and (shape is None or dtype is None): raise ValueError("init, dtype and shape must have values at the same time.") if ((input_data is not None and init is None) or (input_data is None and init is not None)) is False: raise TypeError("input_data and init can not be None at the same time.") if isinstance(shape, numbers.Number): shape = (shape,) if input_data is not None: if isinstance(input_data, np.ndarray) and input_data.ndim > 1 and input_data.size == 0: raise ValueError("input_data can not contain zero dimension.") if isinstance(input_data, (tuple, list)) and np.array(input_data).ndim > 1 \ and np.array(input_data).size == 0: raise ValueError("input_data can not contain zero dimension.") if shape is not None and not (hasattr(init, "__enable_zero_dim__") and init.__enable_zero_dim__): if 0 in shape: raise ValueError("Shape can not contain zero value.") # If input_data is tuple/list/numpy.ndarray, it's support in check_type method. if init is None: validator.check_value_type('input_data', input_data, (Tensor_, np.ndarray, list, tuple, float, int, bool), 'Tensor') valid_dtypes = (np.int8, np.int16, np.int32, np.int64, np.uint8, np.uint16, np.uint32, np.uint64, np.float16, np.float32, np.float64, np.bool_) if isinstance(input_data, np.ndarray) and input_data.dtype not in valid_dtypes: raise TypeError(f"For Tensor, the input_data is a numpy array, " f"but it's data type is not in supported list:\ {list(i.__name__ for i in valid_dtypes)}.") if isinstance(input_data, (tuple, list)): if np.array(input_data).dtype not in valid_dtypes: raise TypeError(f"For Tensor, the input_data is {input_data} that contain unsupported element.") if dtype is not None: validator.check_type_name('dtype', dtype, mstype.number_type + (mstype.bool_,), "Tensor") if isinstance(input_data, np.ndarray) and (not input_data.flags['FORC']): input_data = np.ascontiguousarray(input_data) if dtype is None: Tensor_.__init__(self, input_data) else: Tensor_.__init__(self, input_data, dtype) else: Tensor_.__init__(self, dtype, shape) self._virtual_flag = False self.init = init self.init_finished = True def __deepcopy__(self, memodict): new_obj = Tensor(self) new_obj.init = self.init new_obj._virtual_flag = self._virtual_flag # pylint:disable=w0212 return new_obj def __repr__(self): if self.init_finished: Tensor_.data_sync(self, False) return Tensor_.__repr__(self) return '' def __add__(self, other): out = tensor_operator_registry.get('__add__')(self, other) return out def __eq__(self, other): if not isinstance(other, (int, float, Tensor)): return False # bool type is not supported for `Equal` operator in backend. if self.dtype == mstype.bool_ or (isinstance(other, Tensor) and other.dtype == mstype.bool_): if isinstance(other, Tensor): return Tensor(np.array(self.asnumpy() == other.asnumpy())) return Tensor(np.array(self.asnumpy() == other)) return tensor_operator_registry.get('__eq__')(self, other) def __ne__(self, other): if not isinstance(other, (int, float, Tensor)): return True # bool type is not supported for `NotEqual` operator in backend. if self.dtype == mstype.bool_ or (isinstance(other, Tensor) and other.dtype == mstype.bool_): return Tensor(np.array(self.asnumpy() != other.asnumpy())) return tensor_operator_registry.get('__ne__')(self, other) def __hash__(self): return hash(id(self)) def __mul__(self, other): out = tensor_operator_registry.get('__mul__')(self, other) return out def __neg__(self): out = tensor_operator_registry.get('__neg__')(self) return out def __invert__(self): out = tensor_operator_registry.get('__logical_not__')(self) return out def __bool__(self): data = self.asnumpy() if data.shape == (): return bool(data) if data.shape == (1,): return bool(data[0]) raise ValueError("The truth value of an array with several elements is ambiguous.") def __index__(self): data = self.asnumpy() if not (data.dtype == "int8" or data.dtype == "int16" or data.dtype == "int32" or data.dtype == "int64" or data.dtype == "bool"): raise ValueError("Only integer tensors of a single element can be converted to an index.") if data.shape == (): return int(data) if data.shape == (1,): return int(data[0]) raise ValueError("Only integer tensors of a single element can be converted to an index.") def __pos__(self): return self def __iadd__(self, other): return self.__add__(other) def __radd__(self, other): out = tensor_operator_registry.get('__add__')(self, other) return out def __imul__(self, other): return self.__mul__(other) def __rmul__(self, other): out = tensor_operator_registry.get('__mul__')(self, other) return out def __truediv__(self, other): out = tensor_operator_registry.get('__truediv__')(self, other) return out def __rtruediv__(self, other): out = tensor_operator_registry.get('__truediv__')(other, self) return out def __sub__(self, other): out = tensor_operator_registry.get('__sub__')(self, other) return out def __isub__(self, other): return self.__sub__(other) def __rsub__(self, other): out = tensor_operator_registry.get('__sub__')(other, self) return out def __lt__(self, other): out = tensor_operator_registry.get('__lt__')(self, other) return out def __le__(self, other): out = tensor_operator_registry.get('__le__')(self, other) return out def __getitem__(self, index): if isinstance(index, int) and not isinstance(index, bool) and self.shape and index >= self.shape[0]: raise IndexError("index {} is out of bounds for axis 0 with size {}".format(index, self.shape[0])) out = tensor_operator_registry.get('__getitem__')(self, index) return out def __setitem__(self, index, value): out = tensor_operator_registry.get('__setitem__')(self, index, value) self.assign_value(out) return self def __gt__(self, other): out = tensor_operator_registry.get('__gt__')(self, other) return out def __ge__(self, other): out = tensor_operator_registry.get('__ge__')(self, other) return out def __len__(self): out = tensor_operator_registry.get('shape')(self) if out: return out[0] raise TypeError("Not support len of a 0-D tensor") def __mod__(self, other): return tensor_operator_registry.get('__mod__')(self, other) def __imod__(self, other): return self.__mod__(other) def __rmod__(self, other): return tensor_operator_registry.get('__mod__')(other, self) def __pow__(self, other): return tensor_operator_registry.get('__pow__')(self, other) def __floordiv__(self, other): return tensor_operator_registry.get('__floordiv__')(self, other) def __ifloordiv__(self, other): return self.__floordiv__(other) def __rfloordiv__(self, other): return tensor_operator_registry.get('__floordiv__')(other, self) def __str__(self): if self.dtype == mstype.type_none: return "Unknown Tensor type!" return str(self.asnumpy()) @property def shape(self): """Returns the shape of the tensor as a tuple.""" return self._shape @property def dtype(self): """Return the dtype of the tensor (:class:`mindspore.dtype`).""" return self._dtype @property def size(self): """Returns the total number of elements in tensor.""" return self._size @property def ndim(self): """Return the number of tensor dimensions.""" return len(self._shape) @property def has_init(self): """tensor is inited.""" return self.init is not None @property def itemsize(self): """Return the length of one tensor element in bytes.""" return self._itemsize @property def strides(self): """Return the tuple of bytes to step in each dimension when traversing a tensor.""" return self._strides @property def nbytes(self): """Return the total number of bytes taken by the tensor.""" return self._nbytes @property def T(self): """Return the transposed tensor.""" return self.transpose() @property def virtual_flag(self): """Mark tensor is virtual.""" return self._virtual_flag @virtual_flag.setter def virtual_flag(self, value): """The setter of virtual_flag.""" if not isinstance(value, bool): raise TypeError("virtual_flag must be bool.") self._virtual_flag = value
[docs] @staticmethod def from_numpy(array): """Convert numpy array to Tensor without copy data.""" return Tensor(Tensor_.from_numpy(array))
[docs] def asnumpy(self): """Convert tensor to numpy array.""" self.init_check() return Tensor_.asnumpy(self)
[docs] def flush_from_cache(self): """Flush cache data to host if tensor is cache enable.""" self.init_check() Tensor_._flush_from_cache(self)
[docs] def all(self, axis=(), keep_dims=False): """ Check all array elements along a given axis evaluate to True. Args: axis (Union[None, int, tuple(int)): Dimensions of reduction, when axis is None or empty tuple, reduce all dimensions. Default: (). keep_dims (bool): Whether to keep the reduced dimensions. Default: False. Returns: Tensor, has the same data type as x. """ self.init_check() if axis is None: axis = () return tensor_operator_registry.get('all')(keep_dims)(self, axis)
[docs] def any(self, axis=(), keep_dims=False): """ Check any array element along a given axis evaluate to True. Args: axis (Union[None, int, tuple(int)): Dimensions of reduction, when axis is None or empty tuple, reduce all dimensions. Default: (). keep_dims (bool): Whether to keep the reduced dimensions. Default: False. Returns: Tensor, has the same data type as x. """ self.init_check() if axis is None: axis = () return tensor_operator_registry.get('any')(keep_dims)(self, axis)
[docs] def view(self, *shape): r""" Reshape the tensor according to the input shape. Args: shape (Union[tuple(int), int]): Dimension of the output tensor. Returns: Tensor, has the same dimension as the input shape. """ self.init_check() if not shape: raise ValueError("The shape variable should not be empty") if isinstance(shape[0], tuple): if len(shape) != 1: raise ValueError(f"Only one tuple is needed, but got {shape}") shape = shape[0] return tensor_operator_registry.get('reshape')()(self, shape)
[docs] def expand_as(self, x): """ Expand the dimension of target tensor to the dimension of input tensor. Args: x (Tensor): The input tensor. The shape of input tensor must obey the broadcasting rule. Returns: Tensor, has the same dimension as input tensor. """ self.init_check() return tensor_operator_registry.get('broadcast_to')(x.shape)(self)
[docs] def abs(self): """ Return absolute value element-wisely. Returns: Tensor, has the same data type as x. """ self.init_check() return tensor_operator_registry.get('abs')()(self)
[docs] def mean(self, axis=(), keep_dims=False): """ Reduce a dimension of a tensor by averaging all elements in the dimension. Args: axis (Union[None, int, tuple(int), list(int)]): Dimensions of reduction, when axis is None or empty tuple, reduce all dimensions. Default: (). keep_dims (bool): Whether to keep the reduced dimensions. Default: False. Returns: Tensor, has the same data type as x. """ self.init_check() if axis is None: axis = () return tensor_operator_registry.get('mean')(keep_dims)(self, axis)
[docs] def transpose(self, *axes): r""" Return a view of the tensor with axes transposed. For a 1-D tensor this has no effect, as a transposed vector is simply the same vector. For a 2-D tensor, this is a standard matrix transpose. For a n-D tensor, if axes are given, their order indicates how the axes are permuted. If axes are not provided and tensor.shape = (i[0], i[1],...i[n-2], i[n-1]), then tensor.transpose().shape = (i[n-1], i[n-2], ... i[1], i[0]). Args: axes(Union[None, tuple(int), list(int), int], optional): If axes is None or blank, tensor.transpose() will reverse the order of the axes. If axes is tuple(int) or list(int), tensor.transpose() will transpose the tensor to the new axes order. If axes is int, this form is simply intended as a convenience alternative to the tuple/list form. Returns: Tensor, has the same dimension as input tensor, with axes suitably permuted. """ self.init_check() perm = validator.check_transpose_axis(axes, self.ndim) return tensor_operator_registry.get('transpose')()(self, perm)
[docs] def reshape(self, *shape): """ Give a new shape to a tensor without changing its data. Args: shape(Union[int, tuple(int), list(int)]): The new shape should be compatible with the original shape. If an integer, then the result will be a 1-D array of that length. One shape dimension can be -1. In this case, the value is inferred from the length of the array and remaining dimensions. Returns: Tensor, with new specified shape. """ self.init_check() new_shape = validator.check_reshape_shp(shape) return tensor_operator_registry.get('reshape')()(self, new_shape)
[docs] def ravel(self): """ Return a contiguous flattened tensor. Returns: Tensor, a 1-D tensor, containing the same elements of the input. """ self.init_check() reshape_op = tensor_operator_registry.get('reshape')() return reshape_op(self, (-1,))
[docs] def flatten(self, order='C'): r""" Return a copy of the tensor collapsed into one dimension. Args: order (str, optional): Can choose between 'C' and 'F'. 'C' means to flatten in row-major (C-style) order. 'F' means to flatten in column-major (Fortran-style) order. Only 'C' and 'F' are supported. Default: 'C'. Returns: Tensor, has the same data type as input. """ self.init_check() reshape_op = tensor_operator_registry.get('reshape')() trans_op = tensor_operator_registry.get('transpose')() order = validator.check_flatten_order(order) if order == 'C': return reshape_op(self, (-1,)) perm = tuple(range(self.ndim-1, -1, -1)) return reshape_op(trans_op(self, perm), (-1,))
[docs] def swapaxes(self, axis1, axis2): """ Interchange two axes of a tensor. Args: axis1 (int): First axis. axis2 (int): Second axis. Returns: Transposed tensor, has the same data type as the input. """ self.init_check() axis1, axis2 = validator.check_swapaxes_axis((axis1, axis2), self.ndim) if axis1 == axis2: return self if axis1 > axis2: axis1, axis2 = axis2, axis1 perm = tuple(range(0, self.ndim)) new_perm = None if axis2 + 1 < self.ndim: new_perm = perm[0:axis1] + perm[axis2:axis2+1] + \ perm[axis1+1:axis2] + perm[axis1:axis1+1] + perm[axis2+1:] else: new_perm = perm[0:axis1] + perm[axis2:axis2+1] + \ perm[axis1+1:axis2] + perm[axis1:axis1+1] return tensor_operator_registry.get('transpose')()(self, new_perm)
[docs] def squeeze(self, axis=None): """ Remove single-dimensional entries from the shape of a tensor. Args: axis (Union[None, int, list(int), tuple(int)], optional): Default is None. Returns: Tensor, with all or a subset of the dimensions of length 1 removed. """ self.init_check() if axis is None: return tensor_operator_registry.get('squeeze')(self) new_shape = validator.prepare_shape_for_squeeze(self.shape, axis) return tensor_operator_registry.get('reshape')()(self, new_shape)
[docs] def astype(self, dtype, copy=True): """ Return a copy of the tensor, casted to a specified type. Args: dtype (Union[:class:`mindspore.dtype`, str]): Designated tensor dtype, can be in format of :class:`mindspore.dtype.float32` or `float32`. Default: :class:`mindspore.dtype.float32`. copy (bool, optional): By default, astype always returns a newly allocated tensor. If this is set to false, the input tensor is returned instead of a copy if possible. Default: True. Returns: Tensor, with the designated dtype. """ self.init_check() dtype = validator.check_astype_dtype(dtype) if not copy and dtype == self.dtype: return self return tensor_operator_registry.get('cast')(self, dtype)
def init_check(self): if self.has_init: self.init_data() return self
[docs] def init_data(self, slice_index=None, shape=None, opt_shard_group=None): """ Get the tensor format data of this Tensor. The init_data function can be called once for the same tensor. Args: slice_index (int): Slice index of a parameter's slices. It is used when initialize a slice of a parameter, it guarantees that devices using the same slice can generate the same tensor. shape (list(int)): Shape of the slice, it is used when initialize a slice of the parameter. opt_shard_group(str): Optimizer shard group which is used in auto or semi auto parallel mode to get one shard of a parameter's slice. """ if self.init is None: raise TypeError("init_data must be set Tensor.init, init can't be None") if shape is None: shape = self.shape try: arr = np.ndarray(shape, dtype=mstype.dtype_to_nptype(self.dtype)) except ValueError: msg = "Error shape={}".format(shape) logger.error(msg) raise ValueError(msg) class seed_context: '''set and restore seed''' def __init__(self, init): self.init = init from .seed import get_seed global_seed = get_seed() self._np_seed = np.random.get_state()[1][0] self.need_set_seed = ((slice_index is not None) and (global_seed is None)) def __enter__(self): if self.need_set_seed: self.seed = self.init.seed np.random.seed(slice_index + Tensor.delta_seed) self.init.seed = slice_index + Tensor.delta_seed Tensor.delta_seed += 1 def __exit__(self, ptype, value, trace): if self.need_set_seed: np.random.seed(self._np_seed) self.init.seed, _ = self.seed with seed_context(self.init): self.init(arr) data = np.array(arr) if opt_shard_group: rank = get_rank(opt_shard_group) size = get_group_size(opt_shard_group) data = np.split(data, size)[rank] self.init = None self.assign_value(Tensor(data, dtype=self.dtype)) return self
[docs] def to_tensor(self, slice_index=None, shape=None, opt_shard_group=None): """Return init_data().""" logger.warning("WARN_DEPRECATED: The usage of to_tensor is deprecated." " Please use init_data") return self.init_data(slice_index, shape, opt_shard_group)
[docs]class RowTensor: """ A sparse representation of a set of tensor slices at given indices. An RowTensor is typically used to represent a subset of a larger tensor dense of shape [L0, D1, .. , DN] where L0 >> D0. The values in indices are the indices in the first dimension of the slices that have been extracted from the larger tensor. The dense tensor dense represented by an RowTensor slices has `dense[slices.indices[i], :, :, :, ...] = slices.values[i, :, :, :, ...]`. RowTensor can only be used in the `Cell`'s construct method. It is not supported in pynative mode at the moment. Args: indices (Tensor): A 1-D integer Tensor of shape [D0]. values (Tensor): A Tensor of any dtype of shape [D0, D1, ..., Dn]. dense_shape (tuple(int)): An integer tuple which contains the shape of the corresponding dense tensor. Returns: RowTensor, composed of `indices`, `values`, and `dense_shape`. Examples: >>> import mindspore as ms >>> import mindspore.nn as nn >>> class Net(nn.Cell): ... def __init__(self, dense_shape): ... super(Net, self).__init__() ... self.dense_shape = dense_shape ... def construct(self, indices, values): ... x = RowTensor(indices, values, self.dense_shape) ... return x.values, x.indices, x.dense_shape >>> >>> indices = Tensor([0]) >>> values = Tensor([[1, 2]], dtype=ms.float32) >>> out = Net((3, 2))(indices, values) >>> print(out[0]) [[1. 2.]] >>> print(out[1]) [0] >>> print(out[2]) (3, 2) """ def __init__(self, indices, values, dense_shape): "Init RowTensor" self.__indices = indices self.__values = values self.__dense_shape = dense_shape @property def indices(self): return self.__indices @property def values(self): return self.__values @property def dense_shape(self): return self.__dense_shape
[docs]class SparseTensor: """ A sparse representation of a set of nonzero elememts from a tensor at given indices. SparseTensor can only be used in the `Cell`'s construct method. Pynative mode not supported at the moment. For a tensor dense, its SparseTensor(indices, values, dense_shape) has `dense[indices[i]] = values[i]`. Args: indices (Tensor): A 2-D integer Tensor of shape `[N, ndims]`, where N and ndims are the number of `values` and number of dimensions in the SparseTensor, respectively. values (Tensor): A 1-D tensor of any type and shape `[N]`, which supplies the values for each element in `indices`. dense_shape (tuple(int)): A integer tuple of size `ndims`, which specifies the dense_shape of the sparse tensor. Returns: SparseTensor, composed of `indices`, `values`, and `dense_shape`. Examples: >>> import mindspore as ms >>> import mindspore.nn as nn >>> class Net(nn.Cell): ... def __init__(self, dense_shape): ... super(Net, self).__init__() ... self.dense_shape = dense_shape ... def construct(self, indices, values): ... x = SparseTensor(indices, values, self.dense_shape) ... return x.values, x.indices, x.dense_shape >>> >>> indices = Tensor([[0, 1], [1, 2]]) >>> values = Tensor([1, 2], dtype=ms.float32) >>> out = Net((3, 4))(indices, values) >>> print(out[0]) [1. 2.] >>> print(out[1]) [[0 1] [1 2]] >>> print(out[2]) (3, 4) """ def __init__(self, indices, values, dense_shape): "Init SparseTensor" self.__indices = indices self.__values = values self.__dense_shape = dense_shape @property def indices(self): return self.__indices @property def values(self): return self.__values @property def dense_shape(self): return self.__dense_shape
def _vm_compare(*args): """Implement `vm_compare` for tensor.""" obj_str = args[-1] if obj_str == "shape": fn = getattr(args[0].asnumpy(), obj_str) return fn if len(args) == 2: fn = getattr(args[0].asnumpy(), obj_str) return Tensor(fn()) if isinstance(args[0], Tensor): fn = getattr(args[0].asnumpy(), obj_str) y = args[1].asnumpy() if isinstance(args[1], Tensor) else args[1] else: obj_str = "__r" + obj_str[2:] fn = getattr(args[1].asnumpy(), obj_str) y = args[0] return Tensor(np.array(fn(y))) tensor_operator_registry.register('vm_compare', _vm_compare)