_clip_by_value

clip_by_value

Clip a tensor such that min_value <= tensor <= max_value.

Given an interval, values outside the interval are clipped. If min_value or max_value is not provided then clipping is not performed on lower or upper interval edge respectively.

This method can be used with Numpy data:

n = np.array([-5, 4, 2, 0, 9, -2])
b = fe.backend.clip_by_value(n, min_value=-2, max_value=3)  # [-2, 3, 2, 0, 3, -2]
b = fe.backend.clip_by_value(n, min_value=-2) # [-2, 4, 2, 0, 9, -2]

This method can be used with TensorFlow tensors:

t = tf.constant([-5, 4, 2, 0, 9, -2])
b = fe.backend.clip_by_value(t, min_value=-2, max_value=3)  # [-2, 3, 2, 0, 3, -2]
b = fe.backend.clip_by_value(t, min_value=-2) # [-2, 4, 2, 0, 9, -2]

This method can be used with PyTorch tensors:

p = torch.tensor([-5, 4, 2, 0, 9, -2])
b = fe.backend.clip_by_value(p, min_value=-2, max_value=3)  # [-2, 3, 2, 0, 3, -2]
b = fe.backend.clip_by_value(p, min_value=-2) # [-2, 4, 2, 0, 9, -2]

Parameters:

Name	Type	Description	Default
tensor	Tensor	The input value.	required
min_value	Union[int, float, Tensor, None]	The minimum value to clip to.	None
max_value	Union[int, float, Tensor, None]	The maximum value to clip to.	None

Returns:

Type	Description
Tensor	The tensor, with it's values clipped.

Raises:

Type	Description
ValueError	If tensor is an unacceptable data type.

Source code in fastestimator/fastestimator/backend/_clip_by_value.py

def clip_by_value(tensor: Tensor,
                  min_value: Union[int, float, Tensor, None] = None,
                  max_value: Union[int, float, Tensor, None] = None) -> Tensor:
    """Clip a tensor such that `min_value` <= tensor <= `max_value`.

    Given an interval, values outside the interval are clipped. If `min_value` or `max_value` is not provided then
    clipping is not performed on lower or upper interval edge respectively.

    This method can be used with Numpy data:
    ```python
    n = np.array([-5, 4, 2, 0, 9, -2])
    b = fe.backend.clip_by_value(n, min_value=-2, max_value=3)  # [-2, 3, 2, 0, 3, -2]
    b = fe.backend.clip_by_value(n, min_value=-2) # [-2, 4, 2, 0, 9, -2]
    ```

    This method can be used with TensorFlow tensors:
    ```python
    t = tf.constant([-5, 4, 2, 0, 9, -2])
    b = fe.backend.clip_by_value(t, min_value=-2, max_value=3)  # [-2, 3, 2, 0, 3, -2]
    b = fe.backend.clip_by_value(t, min_value=-2) # [-2, 4, 2, 0, 9, -2]
    ```

    This method can be used with PyTorch tensors:
    ```python
    p = torch.tensor([-5, 4, 2, 0, 9, -2])
    b = fe.backend.clip_by_value(p, min_value=-2, max_value=3)  # [-2, 3, 2, 0, 3, -2]
    b = fe.backend.clip_by_value(p, min_value=-2) # [-2, 4, 2, 0, 9, -2]
    ```

    Args:
        tensor: The input value.
        min_value: The minimum value to clip to.
        max_value: The maximum value to clip to.

    Returns:
        The `tensor`, with it's values clipped.

    Raises:
        ValueError: If `tensor` is an unacceptable data type.
    """
    assert min_value is not None or max_value is not None, "Both min_value and max_value must not be NoneType"
    if tf.is_tensor(tensor):
        if min_value is None:
            return tf.math.minimum(tensor, max_value)
        elif max_value is None:
            return tf.math.maximum(tensor, min_value)
        else:
            return tf.clip_by_value(tensor, clip_value_min=min_value, clip_value_max=max_value)
    elif isinstance(tensor, torch.Tensor):
        if isinstance(min_value, torch.Tensor):
            min_value = min_value.item()
        if isinstance(max_value, torch.Tensor):
            max_value = max_value.item()
        return tensor.clamp(min=min_value, max=max_value)
    elif isinstance(tensor, np.ndarray):
        return np.clip(tensor, a_min=min_value, a_max=max_value)
    else:
        raise ValueError("Unrecognized tensor type {}".format(type(tensor)))