ttnn.fold

ttnn.fold(input: ttnn.Tensor, stride_h: int, stride_w: int, collapse_output: bool | None) → None

NHWC space-to-depth. Packs every stride_h * stride_w neighbourhood at position (h * stride_h, w * stride_w) into the channel dim at position (h, w). Shapes below use the padded H/W/C (any HW / channel padding is applied before folding).

Output shape: (N, Hp/stride_h, Wp/stride_w, Cp * stride_h * stride_w) where Hp, Wp, Cp are the padded input dims.