Skip to content

gradient

GradientOp

Bases: TensorOp

Return the gradients of finals w.r.t. inputs.

Parameters:

Name Type Description Default
inputs Union[str, List[str]]

The tensor(s) to compute gradients with respect to.

 required
finals Union[str, List[str]]

The tensor(s) to compute gradients from.

 required
outputs Union[str, List[str]]

The key(s) under which to save the gradients.

 required
mode Union[None, str, Iterable[str]]

What mode(s) to execute this Op in. For example, "train", "eval", "test", or "infer". To execute regardless of mode, pass None. To execute in all modes except for a particular one, you can pass an argument like "!infer" or "!train".

 'eval'
Source code in fastestimator\fastestimator\op\tensorop\gradient\gradient.py 
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
class GradientOp(TensorOp):
    """Return the gradients of finals w.r.t. inputs.

    Args:
        inputs: The tensor(s) to compute gradients with respect to.
        finals: The tensor(s) to compute gradients from.
        outputs: The key(s) under which to save the gradients.
        mode: What mode(s) to execute this Op in. For example, "train", "eval", "test", or "infer". To execute
            regardless of mode, pass None. To execute in all modes except for a particular one, you can pass an argument
            like "!infer" or "!train".
    """
    def __init__(self,
                 inputs: Union[str, List[str]],
                 finals: Union[str, List[str]],
                 outputs: Union[str, List[str]],
                 mode: Union[None, str, Iterable[str]] = "eval"):
        inputs = to_list(inputs)
        finals = to_list(finals)
        outputs = to_list(outputs)
        assert len(inputs) == len(finals) == len(outputs), \
            "GradientOp requires the same number of inputs, finals, and outputs"
        inputs.extend(finals)
        super().__init__(inputs=inputs, outputs=outputs, mode=mode)
        self.retain_graph = True

    def forward(self, data: List[Tensor], state: Dict[str, Any]) -> List[Tensor]:
        initials = data[:len(data) // 2]
        finals = data[len(data) // 2:]
        results = []
        for initial, final in zip(initials, finals):
            results.append(get_gradient(final, initial, tape=state['tape'], retain_graph=self.retain_graph))
        return results