einops_repeat_op

→TensorImplPtr

int einops_repeat_op(const int & a, const int & pattern, const int & axes_lengths)

source edit

Tile / broadcast a tensor to introduce new axes specified by an einops pattern.

Axes present in rhs but absent from lhs are new and their sizes must be provided in axes_lengths. Existing axes may additionally be reordered and merged exactly as in einops_rearrange_op. Internally the op uses unsqueeze + broadcast_to so it avoids data copies whenever possible — repeat is materialised only when subsequent ops force a contiguous layout.

Math

For each new axis $k$ of size $n_k$ :

y_{\dots, k, \dots} = x_{\dots}, \quad k = 0, 1, \dots, n_k - 1

See Also

einops_rearrange_op, einops_reduce_op, einsum_op

Parameters

aTensorImplPtr

Input tensor.

patternstd::string

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

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

Sizes for the new axes introduced on the rhs. Must contain every axis name that does not appear on the lhs.

Returns

TensorImplPtr

Tensor with new axes broadcast/repeated to the requested sizes.

Raises

EinopsPatternError

If a new axis on the rhs is missing from axes_lengths.