onehot

Onehot

Bases: NumpyOp

Transform an integer label to one-hot-encoding.

This can be desirable for increasing robustness against incorrect labels: https://towardsdatascience.com/label-smoothing-making-model-robust-to-incorrect-labels-2fae037ffbd0

Parameters:

Name	Type	Description	Default
inputs	Union[str, Iterable[str], Callable]	Input key(s) of labels to be onehot encoded.	required
outputs	Union[str, Iterable[str]]	Output key(s) of labels.	required
num_classes	int	Total number of classes.	required
label_smoothing	float	Smoothing factor, after smoothing class output is: 1 - label_smoothing + label_smoothing / num_classes, the other class output is: label_smoothing / num_classes.	0.0
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".	None

Source code in fastestimator\fastestimator\op\numpyop\univariate\onehot.py

class Onehot(NumpyOp):
    """Transform an integer label to one-hot-encoding.

    This can be desirable for increasing robustness against incorrect labels:
    https://towardsdatascience.com/label-smoothing-making-model-robust-to-incorrect-labels-2fae037ffbd0

    Args:
        inputs: Input key(s) of labels to be onehot encoded.
        outputs: Output key(s) of labels.
        num_classes: Total number of classes.
        label_smoothing: Smoothing factor, after smoothing class output is: 1 - label_smoothing + label_smoothing
            / num_classes, the other class output is: label_smoothing / num_classes.
        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, Iterable[str], Callable],
                 outputs: Union[str, Iterable[str]],
                 num_classes: int,
                 label_smoothing: float = 0.0,
                 mode: Union[None, str, Iterable[str]] = None):
        super().__init__(inputs=inputs, outputs=outputs, mode=mode)
        self.num_classes = num_classes
        self.label_smoothing = label_smoothing
        self.in_list, self.out_list = True, True

    def forward(self, data: List[Union[int, np.ndarray]], state: Dict[str, Any]) -> List[np.ndarray]:
        return [self._apply_onehot(elem) for elem in data]

    def _apply_onehot(self, data: Union[int, np.ndarray]) -> np.ndarray:
        class_index = np.array(data)
        assert "int" in str(class_index.dtype)
        assert class_index.size == 1, "data must have only one item"
        class_index = class_index.item()
        assert class_index < self.num_classes, "label value should be smaller than num_classes"
        output = np.full((self.num_classes), fill_value=self.label_smoothing / self.num_classes)
        output[class_index] = 1.0 - self.label_smoothing + self.label_smoothing / self.num_classes
        return output