wehub-resource-sync
90 lines
2.4 KiB
Python
Executable File
90 lines
2.4 KiB
Python
Executable File
import numpy as np
|
|
|
|
|
|
class Tensor(object):
|
|
"""
|
|
a base class for tensor object
|
|
"""
|
|
__array_ufunc__ = None
|
|
|
|
def __init__(self, value, function=None):
|
|
"""
|
|
construct Tensor object
|
|
Parameters
|
|
----------
|
|
value : array_like
|
|
value of this tensor
|
|
function : Function
|
|
function output this tensor
|
|
"""
|
|
if not isinstance(value, (int, float, np.number, np.ndarray)):
|
|
raise TypeError(
|
|
"Unsupported class for Tensor: {}".format(type(value))
|
|
)
|
|
self.value = value
|
|
self.function = function
|
|
|
|
def __format__(self, *args, **kwargs):
|
|
return self.__repr__()
|
|
|
|
def __repr__(self):
|
|
if isinstance(self.value, np.ndarray):
|
|
return (
|
|
"{0}(shape={1.shape}, dtype={1.dtype})"
|
|
.format(self.__class__.__name__, self.value)
|
|
)
|
|
else:
|
|
return (
|
|
"{0}(value={1})".format(self.__class__.__name__, self.value)
|
|
)
|
|
|
|
@property
|
|
def ndim(self):
|
|
if hasattr(self.value, "ndim"):
|
|
return self.value.ndim
|
|
else:
|
|
return 0
|
|
|
|
@property
|
|
def shape(self):
|
|
if hasattr(self.value, "shape"):
|
|
return self.value.shape
|
|
else:
|
|
return ()
|
|
|
|
@property
|
|
def size(self):
|
|
if hasattr(self.value, "size"):
|
|
return self.value.size
|
|
else:
|
|
return 1
|
|
|
|
def backward(self, delta=1, **kwargs):
|
|
"""
|
|
back-propagate error
|
|
Parameters
|
|
----------
|
|
delta : array_like
|
|
derivative with respect to this array
|
|
"""
|
|
if isinstance(delta, np.ndarray):
|
|
if delta.shape != self.shape:
|
|
raise ValueError(
|
|
"shapes {} and {} not aligned"
|
|
.format(delta.shape, self.shape)
|
|
)
|
|
elif isinstance(delta, (int, float, np.number)):
|
|
if self.shape != ():
|
|
raise ValueError(
|
|
"delta must be np.ndarray"
|
|
)
|
|
else:
|
|
raise TypeError(
|
|
"unsupported class for delta"
|
|
)
|
|
self._backward(delta, **kwargs)
|
|
|
|
def _backward(self, delta, **kwargs):
|
|
if hasattr(self.function, "backward"):
|
|
self.function.backward(delta, **kwargs)
|