sometimes

Sometimes

Bases: NumpyOp

Perform a NumpyOp with a given probability.

Note that Sometimes should not be used to wrap an op whose output key(s) do not already exist in the data dictionary. This would result in a problem when future ops / traces attempt to reference the output key, but Sometimes declined to generate it. If you want to create a default value for a new key, simply use a LambdaOp before invoking the Sometimes.

Parameters:

Name	Type	Description	Default
numpy_op	NumpyOp	The operator to be performed.	required
prob	float	The probability of execution, which should be in the range: [0-1).	0.5

Source code in fastestimator\fastestimator\op\numpyop\meta\sometimes.py

@traceable()
class Sometimes(NumpyOp):
    """Perform a NumpyOp with a given probability.

    Note that Sometimes should not be used to wrap an op whose output key(s) do not already exist in the data
    dictionary. This would result in a problem when future ops / traces attempt to reference the output key, but
    Sometimes declined to generate it. If you want to create a default value for a new key, simply use a LambdaOp before
    invoking the Sometimes.

    Args:
        numpy_op: The operator to be performed.
        prob: The probability of execution, which should be in the range: [0-1).
    """
    def __init__(self, numpy_op: NumpyOp, prob: float = 0.5) -> None:
        # We're going to try to collect any missing output keys from the data dictionary so that they don't get
        # overridden when Sometimes chooses not to execute.
        inps = set(numpy_op.inputs)
        outs = set(numpy_op.outputs)
        self.extra_inputs = list(outs - inps)  # Used by traceability
        self.inp_idx = len(numpy_op.inputs)
        super().__init__(inputs=numpy_op.inputs + self.extra_inputs, outputs=numpy_op.outputs, mode=numpy_op.mode)
        # Note that in_list and out_list will always be true
        self.op = numpy_op
        self.prob = prob

    def __getstate__(self) -> Dict[str, Dict[Any, Any]]:
        return {'op': self.op.__getstate__() if hasattr(self.op, '__getstate__') else {}}

    def forward(self, data: List[np.ndarray], state: Dict[str, Any]) -> List[np.ndarray]:
        """Execute the wrapped operator a certain fraction of the time.

        Args:
            data: The information to be passed to the wrapped operator.
            state: Information about the current execution context, for example {"mode": "train"}.

        Returns:
            The original `data`, or the `data` after running it through the wrapped operator.
        """
        if self.prob > np.random.uniform():
            data = data[:self.inp_idx]  # Cut off the unnecessary inputs
            if not self.op.in_list:
                data = data[0]
            data = self.op.forward(data, state)
            if not self.op.out_list:
                data = [data]
        else:
            data = [data[self.inputs.index(out)] for out in self.outputs]
        return data

forward

Execute the wrapped operator a certain fraction of the time.

Parameters:

Name	Type	Description	Default
data	List[np.ndarray]	The information to be passed to the wrapped operator.	required
state	Dict[str, Any]	Information about the current execution context, for example {"mode": "train"}.	required

Returns:

Type	Description
List[np.ndarray]	The original data, or the data after running it through the wrapped operator.

Source code in fastestimator\fastestimator\op\numpyop\meta\sometimes.py

def forward(self, data: List[np.ndarray], state: Dict[str, Any]) -> List[np.ndarray]:
    """Execute the wrapped operator a certain fraction of the time.

    Args:
        data: The information to be passed to the wrapped operator.
        state: Information about the current execution context, for example {"mode": "train"}.

    Returns:
        The original `data`, or the `data` after running it through the wrapped operator.
    """
    if self.prob > np.random.uniform():
        data = data[:self.inp_idx]  # Cut off the unnecessary inputs
        if not self.op.in_list:
            data = data[0]
        data = self.op.forward(data, state)
        if not self.op.out_list:
            data = [data]
    else:
        data = [data[self.inputs.index(out)] for out in self.outputs]
    return data