einops_reduce_op

→TensorImplPtr

int einops_reduce_op(const int & a, const int & pattern, int reduction, const int & axes_lengths)

source edit

Reduce a tensor over one or more axes specified by an einops pattern.

Axes present on the lhs but absent from the rhs are collapsed using the given reduction code. axes_lengths is forwarded to the internal rearrange step that aligns axes before the reduction.

Math

For reduction = sum:

y_{j_1, \dots, j_m} = \sum_{k_1, \dots, k_r} x_{i_1, \dots, i_n}

where $\{k_\bullet\}$ are the axes dropped between lhs and rhs. The mean / max / min / prod variants substitute the corresponding monoid.

See Also

einops_rearrange_op, einops_repeat_op, einsum_op

Parameters

aTensorImplPtr

Input tensor.

patternstd::string

Einops pattern "lhs -> rhs" with at least one axis dropped on the rhs.

reductionint

Reduction code: 1=mean, 2=sum, 3=max, 4=min, 5=prod. Must match the Python-side enum.

axes_lengthsstd::map<std::string, std::int64_t>

Named sizes for axes whose size cannot be inferred from a.shape.

Returns

TensorImplPtr

Reduced tensor of the shape implied by rhs.

Raises

EinopsPatternError

If the pattern is malformed or no axis is dropped (use rearrange instead).

:invalid_argument

If reduction is outside the accepted code range.