jax.ops.segment_sum(data, segment_ids, num_segments=None, indices_are_sorted=False, unique_indices=False, bucket_size=None, mode=None)[source]#

Computes the sum within segments of an array.

Similar to TensorFlow’s segment_sum

  • data (jax.typing.ArrayLike) – an array with the values to be summed.

  • segment_ids (jax.typing.ArrayLike) – an array with integer dtype that indicates the segments of data (along its leading axis) to be summed. Values can be repeated and need not be sorted.

  • num_segments (int | None) – optional, an int with nonnegative value indicating the number of segments. The default is set to be the minimum number of segments that would support all indices in segment_ids, calculated as max(segment_ids) + 1. Since num_segments determines the size of the output, a static value must be provided to use segment_sum in a JIT-compiled function.

  • indices_are_sorted (bool) – whether segment_ids is known to be sorted.

  • unique_indices (bool) – whether segment_ids is known to be free of duplicates.

  • bucket_size (int | None) – size of bucket to group indices into. segment_sum is performed on each bucket separately to improve numerical stability of addition. Default None means no bucketing.

  • mode (GatherScatterMode | None) – a jax.lax.GatherScatterMode value describing how out-of-bounds indices should be handled. By default, values outside of the range [0, num_segments) are dropped and do not contribute to the sum.


An array with shape (num_segments,) + data.shape[1:] representing the segment sums.

Return type:



Simple 1D segment sum:

>>> data = jnp.arange(5)
>>> segment_ids = jnp.array([0, 0, 1, 1, 2])
>>> segment_sum(data, segment_ids)
Array([1, 5, 4], dtype=int32)

Using JIT requires static num_segments:

>>> from jax import jit
>>> jit(segment_sum, static_argnums=2)(data, segment_ids, 3)
Array([1, 5, 4], dtype=int32)