conv_node.h File Reference

Convolution operation node compatible with ONNX Conv operator. More...

#include "sdfg/data_flow/library_nodes/math/blas/blas_node.h"
#include "sdfg/data_flow/library_nodes/math/tensor/spatial_tensor_node.h"
#include "sdfg/data_flow/library_nodes/math/tensor/tensor_node.h"

Go to the source code of this file.

Classes
class	sdfg::math::tensor::ConvNode
	Convolution operation following ONNX Conv operator specification. More...
struct	sdfg::math::tensor::ConvNode::ConvExpandPrerequisits
class	sdfg::math::tensor::ConvNodeSerializer
	Serializer for ConvNode. More...

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

Functions
data_flow::LibraryNodeCode	sdfg::math::tensor::LibraryNodeType_Conv ("ml::Conv")

Detailed Description

Convolution operation node compatible with ONNX Conv operator.

This file defines the ConvNode class which implements a tensor convolution operation following the ONNX Conv operator specification. The node is expanded using the im2col transformation into a GEMM operation.

ONNX Conv Operator Compatibility

The ConvNode implements the ONNX Conv operator with the following parameters:

• Input tensor X: [N, C_in, D1, D2, ..., Dn] for n-dimensional convolution
• Weight tensor W: [C_out, C_in/group, k1, k2, ..., kn]
• Optional bias tensor B: [C_out]
• Output tensor Y: [N, C_out, D1_out, D2_out, ..., Dn_out]

Supported attributes:

• kernel_shape: Shape of the convolution kernel
• strides: Stride along each spatial axis
• pads: Padding for the beginning and ending along each spatial axis
• dilations: Dilation along each spatial axis
• group: Number of groups for grouped convolutions