from __future__ import annotations
from typing import Any
import numpy as np
from .grad import dag
from .utils import check
from ._tensor import _Tensor
[docs]
def from_numpy(array: np.ndarray, requires_grad: bool = False) -> _Tensor:
"""
Create a tensor that shares storage with an existing NumPy array.
Parameters
----------
array : numpy.ndarray
Source array. The tensor reuses the same underlying storage.
requires_grad : bool, default=False
Whether to track operations on the returned tensor for automatic
differentiation.
Returns
-------
_Tensor
Tensor view over ``array``.
Raises
------
TypeError
If ``array`` is not a NumPy array.
RuntimeError
If ``array`` is read-only.
Notes
-----
Mutations are shared between the NumPy array and the returned tensor.
See Also
--------
`torch.from_numpy <https://pytorch.org/docs/stable/generated/torch.from_numpy.html>`_
"""
# fast path to create tensor from numpy array without copy
# /!\ Warning ! Shared storage, changes to the original array will be reflected in the tensor and vice versa
if not isinstance(array, np.ndarray):
raise TypeError(f"Input should be a np.ndarray but got {type(array)}.")
if array.flags.writeable==False:
raise RuntimeError("Writing to a tensor created from a read-only NumPy array is not supported.")
new_tensor = _Tensor(array=array, requires_grad=requires_grad)
if requires_grad:
new_tensor.node_id = dag.create_node(parents_id=[None], operation=None, result=new_tensor, is_leaf=True)
return new_tensor
[docs]
def tensor(array: Any, requires_grad: bool = False, dtype: Any = None) -> _Tensor:
"""
Create a tensor from a Python scalar, list, or NumPy array.
Parameters
----------
array : scalar, list, or numpy.ndarray
Input data used to build the tensor.
requires_grad : bool, default=False
Whether to track operations on the returned tensor for automatic
differentiation.
dtype : numpy.dtype or type, optional
Requested tensor dtype. If omitted, the dtype is inferred from the
input.
Returns
-------
_Tensor
Newly created tensor.
Notes
-----
Unlike :func:`from_numpy`, this constructor copies NumPy inputs.
See Also
--------
`torch.tensor <https://pytorch.org/docs/stable/generated/torch.tensor.html>`_
"""
# Instantiating tensor from numpy array copies the underlying buffer. For no-copy use from_numpy.
new_tensor = _Tensor(array=array, requires_grad=requires_grad, dtype=dtype)
if requires_grad:
new_tensor.node_id = dag.create_node(parents_id=[None], operation=None, result=new_tensor, is_leaf=True)
return new_tensor
[docs]
def randn(*shape: int, requires_grad: bool = False, dtype: Any = np.float32) -> _Tensor:
"""
Return a tensor filled with samples from the standard normal distribution.
Parameters
----------
*shape : int
Output shape.
requires_grad : bool, default=False
Whether to track gradients for the returned tensor.
dtype : numpy.dtype or type, default=numpy.float32
Output dtype.
Returns
-------
_Tensor
Randomly initialized tensor.
See Also
--------
`torch.randn <https://pytorch.org/docs/stable/generated/torch.randn.html>`_
"""
if not isinstance(shape, tuple):
raise TypeError(f"Expected a shape of type tuple, got {type(shape)} instead")
if 0 in shape:
raise ValueError("At least one of the dimension is empty")
return tensor(np.random.randn(*shape).astype(dtype), requires_grad=requires_grad)
[docs]
def ones(*shape: int, requires_grad: bool = False, dtype: Any = np.float32) -> _Tensor:
"""
Return a tensor filled with ones.
Parameters
----------
*shape : int
Output shape.
requires_grad : bool, default=False
Whether to track gradients for the returned tensor.
dtype : numpy.dtype or type, default=numpy.float32
Output dtype.
Returns
-------
_Tensor
Tensor filled with ones.
See Also
--------
`torch.ones <https://pytorch.org/docs/stable/generated/torch.ones.html>`_
"""
if not isinstance(shape, tuple):
raise TypeError(f"Expected a shape of type tuple, got {type(shape)} instead")
if 0 in shape:
raise ValueError("At least one of the dimension is empty")
return tensor(np.ones(shape).astype(dtype), requires_grad=requires_grad)
[docs]
def zeros(*shape: int, requires_grad: bool = False, dtype: Any = np.float32) -> _Tensor:
"""
Return a tensor filled with zeros.
Parameters
----------
*shape : int
Output shape.
requires_grad : bool, default=False
Whether to track gradients for the returned tensor.
dtype : numpy.dtype or type, default=numpy.float32
Output dtype.
Returns
-------
_Tensor
Tensor filled with zeros.
See Also
--------
`torch.zeros <https://pytorch.org/docs/stable/generated/torch.zeros.html>`_
"""
if not isinstance(shape, tuple):
raise TypeError(f"Expected a shape of type tuple, got {type(shape)} instead")
if 0 in shape:
raise ValueError("At least one of the dimension is empty")
return tensor(np.zeros(shape).astype(dtype), requires_grad=requires_grad)