upsample_node.h File Reference

Bilinear 2D upsampling node compatible with aten::upsample_bilinear2d.vec. More...

#include "sdfg/data_flow/library_nodes/math/tensor/tensor_node.h"
#include "sdfg/serializer/json_serializer.h"

Go to the source code of this file.

Classes
class	sdfg::math::tensor::UpsampleBilinear2DNode
	2D bilinear upsampling operation More...
class	sdfg::math::tensor::UpsampleBilinear2DNodeSerializer
	Serializer for UpsampleBilinear2DNode. More...

Namespaces
namespace	sdfg
namespace	sdfg::math
namespace	sdfg::math::tensor

Functions
data_flow::LibraryNodeCode	sdfg::math::tensor::LibraryNodeType_UpsampleBilinear2D ("ml::UpsampleBilinear2D")

Detailed Description

Bilinear 2D upsampling node compatible with aten::upsample_bilinear2d.vec.

This file defines the UpsampleBilinear2DNode class which implements a 2D bilinear resize following the PyTorch aten::upsample_bilinear2d.vec operator semantics:

upsample_bilinear2d.vec(Tensor input, SymInt[]? output_size, bool align_corners, float[]? scale_factors) -> Tensor

Input/Output Requirements

• Input connector "X": Input tensor [N, C, H_in, W_in]
• Input connector "Y": Output tensor [N, C, H_out, W_out] (written in place)

Source-coordinate mapping

For each output index o along a spatial dimension the corresponding (fractional) source coordinate is computed exactly as PyTorch does:

• align_corners: src = o * (In - 1) / (Out - 1) (0 if Out == 1)
• otherwise: src = max(0, (o + 0.5) * rscale - 0.5) where rscale is 1 / scale_factor when explicit scale factors were provided, and In / Out otherwise.

The integer neighbours are i0 = floor(src) and i1 = min(i0 + 1, In - 1) with interpolation weight lambda = src - i0.