def _parse_inputs(inputs: Iterable[Any]) -> List[Tensor]:
''' Parse all inputs that are not Nodes to Tensors
'''
return [
x if isinstance(x, _Node) else Tensor(x, name=str(x)) for x in inputs
]
def forward(self, input: Tensor, target: Tensor) -> Tensor:
# Avoid division by zero
input.value = clip(input.value, 1e-16, 1 - 1e-16)
return -self.reduction_fn(sum(target * log(input), dim=1,
keepdim=True),
dim=self.dim,
keepdim=self.keepdim)
def randint(*shape: int,
low=0,
high=100,
diff=False,
name='Tensor[randint]') -> Tensor:
''' Return random integers from low (inclusive) to high (exclusive).
'''
return Tensor(np_randint(low, high=high, size=shape), diff=diff, name=name)
def __init__(self, *children: Union[Tensor, ndarray, List[Number],
Number]):
super(Function, self).__init__(*_parse_inputs(children),
name=self.__class__.__name__)
# This output placeholder is reused when possible
self._output_placeholder = Tensor(
None,
diff=any(x.diff for x in self.children) and modes.DIFF_ENABLED,
creator=self if modes.DIFF_ENABLED else None,
name=self._generate_tensor_name())
if modes.DIFF_ENABLED: # If graph building is enabled.
# Allocate space for parent's output (output placeholder)
for child in self.children:
child.parents_outputs.append(self._output_placeholder)
def randn(*shape: int, diff=False, name='Tensor[randn]') -> Tensor:
''' Return a sample (or samples) from the "standard normal" distribution.
'''
return Tensor(np_randn(*shape), diff=diff, name=name)
def rand(*shape: int, diff=False, name='Tensor[rand]') -> Tensor:
''' Random values in a given shape.
'''
return Tensor(np_rand(*shape), diff=diff, name=name)
def __init__(self, name):
super(Add1, self).__init__(name=name)
self.one = Tensor(1)
def __init__(self, name):
super(Mul2, self).__init__(name=name)
self.two = Tensor(2)
def __init__(self, name):
super(CustomFlow, self).__init__(name=name)
self.two = Tensor(2)
self.fourty_two = Tensor(42)