Ctrl+K

enp.interp

Contents

enp.interp#

etils.enp.interp(x: etils.enp.array_types.typing.Array, from_: Tuple[int | float | etils.enp.array_types.typing.Array | Tuple[Any, ...] | List[Any], int | float | etils.enp.array_types.typing.Array | Tuple[Any, ...] | List[Any]], to: Tuple[int | float | etils.enp.array_types.typing.Array | Tuple[Any, ...] | List[Any], int | float | etils.enp.array_types.typing.Array | Tuple[Any, ...] | List[Any]], *, axis: int = -1, xnp: Any = Ellipsis) → etils.enp.array_types.typing.FloatArray[source]#

Linearly scale the given value by the given range.

Somehow similar to np.interp or scipy.interpolate.inter1d with some differences like support scaling an axis by a different factors and extrapolate values outside the boundaries.

from_ and to are expected to be (min, max) tuples and the function interpolate between the two ranges.

Example: Normalizing a uint8 image to (-1, 1).

img = jnp.array([
    [0, 0],
    [127, 255],
])
img = enp.interp(img, (0, 255), (0, 1))
img == jnp.array([
    [-1, -1],
    [0.498..., 1],
])

min and max can be either float values or array like structure, in which case the numpy broadcasting rules applies (x should be a Array[… d] and min/max values should be Array[d].

Example: Converting normalized 3d coordinates to world coordinates.

coords = enp.interp(coords, from_=(-1, 1), to=(0, (h, w, d)))

  • coords[:, 0] is interpolated from (-1, 1) to (0, h)

  • coords[:, 1] is interpolated from (-1, 1) to (0, w)

  • coords[:, 2] is interpolated from (-1, 1) to (0, d)

Parameters:

  • x – Array to scale

  • from – Range of x.

  • to – Range to which normalize x.

  • axis – Axis on which normalizing. Only relevant if from_ or to items contains range value.

  • xnp – Numpy module to use

Returns:

Float tensor with same shape as x, but with normalized coordinates.

Contents