Tile a tensor by repeating it along multiple axes.

reps[d] specifies how many times to repeat along dimension d. If reps is longer than a.rank(), the input is conceptually left-padded with size-1 dimensions before tiling, so the output rank equals reps.size().

Math

$y_{[i_0, \ldots, i_{r-1}]} = x_{[i_0 \bmod D_0, \ldots, i_{r-1} \bmod D_{r-1}]}$ where $D_d$ is the input's effective size along dimension $d$.

Shape

Let r = reps.size() and n = a.rank(). After left-padding a with max(r - n, 0) size-1 dims: out.shape[d] = a_padded.shape[d] * reps[d] for d in [0, max(r, n)).

See Also

repeat_op — single-axis, interleaved variant.