unique_inverse(x: array, /) Tuple[array, array]

Returns the unique elements of an input array x and the indices from the set of unique elements that reconstruct x.

Data-dependent output shape

The shapes of two of the output arrays for this function depend on the data values in the input array; hence, array libraries which build computation graphs (e.g., JAX, Dask, etc.) may find this function difficult to implement without knowing array values. Accordingly, such libraries may choose to omit this function. See Data-dependent output shapes section for more details.


Uniqueness should be determined based on value equality (see equal()). For input arrays having floating-point data types, value-based equality implies the following behavior.

  • As nan values compare as False, nan values should be considered distinct.

  • As complex floating-point values having at least one nan component compare as False, complex floating-point values having nan components should be considered distinct.

  • As -0 and +0 compare as True, signed zeros should not be considered distinct, and the corresponding unique element will be implementation-dependent (e.g., an implementation could choose to return -0 if -0 occurs before +0).

As signed zeros are not distinct, using inverse_indices to reconstruct the input array is not guaranteed to return an array having the exact same values.


x (array) – input array. If x has more than one dimension, the function must flatten x and return the unique elements of the flattened array.


out (Tuple[array, array]) – a namedtuple (values, inverse_indices) whose

  • first element must have the field name values and must be a one-dimensional array containing the unique elements of x. The array must have the same data type as x.

  • second element must have the field name inverse_indices and must be an array containing the indices of values that reconstruct x. The array must have the same shape as x and have the default array index data type.


The order of unique elements is not specified and may vary between implementations.


Changed in version 2022.12: Added complex data type support.

Changed in version 2023.12: Clarified flattening behavior.