interpolate_nearest_2d_op

→TensorImplPtr

int interpolate_nearest_2d_op(const int & input, int H_out, int W_out)

Public 2-D nearest-neighbour resize (no autograd).

Each output pixel receives the value of the nearest input pixel:

y[i, j] = x\!\left[\lfloor (i + 0.5) \cdot H_{\text{in}} / H_{\text{out}} \rfloor,\; \lfloor (j + 0.5) \cdot W_{\text{in}} / W_{\text{out}} \rfloor\right]

No backward node is attached — the operation is inference-only.

inputTensorImplPtr

4-D input tensor (N, C, H_in, W_in).

H_outint

Desired output height.

W_outint

Desired output width.

TensorImplPtr

Resampled tensor of shape (N, C, H_out, W_out).

Loading…