col2im_f32

void col2im_f32(const float * cols, float * dx, int C, int H, int W, int KH, int KW, int OH, int OW, int stride_h, int stride_w, int pad_h, int pad_w, int dilation_h, int dilation_w)

source edit

Single-precision 2-D col2im fold (gradient transform).

Adjoint of im2col_f32: scatters each receptive-field column of cols back into the corresponding input positions of $dx$ , accumulating by addition where windows overlap. Used in the conv2d backward pass.

Parameters

colsconst float*

Column matrix, shape

(C \cdot K_H \cdot K_W, O_H \cdot O_W)

dxfloat*

Output gradient tensor, shape

(C, H, W)

; must be zeroed by the caller before this call.

C, H, W, KH, KW, OH, OWint

Shape parameters.

stride_h, stride_w, pad_h, pad_w, dilation_h, dilation_wint

Match the forward convolution.

Notes

Overlapping receptive fields accumulate by sum — this is the mathematically correct adjoint of im2col under any stride / dilation.