col2im_1d_f32

void col2im_1d_f32(const float * cols, float * dx, int C, int L, int KL, int OL, int stride_l, int pad_l, int dilation_l)

source edit

Single-precision 1-D col2im fold (gradient transform).

Inverse of im2col_1d_f32: scatter-adds each column of cols back into the corresponding positions of $dx$ . Overlapping receptive fields (stride < kernel) accumulate by sum — this is the mathematically correct adjoint that the conv1d backward needs.

Math

$dx[c, l] = \sum_{o, k : o \cdot s + k \cdot d - p = l} \text{cols}[c \cdot KL + k, o]$

Parameters

colsconst float*

Column matrix, shape

(C \cdot KL, OL)

dxfloat*

Output gradient tensor, shape

(C, L)

; must be zeroed by the caller before this call.

C, L, KL, OLint

Channel count, input length, kernel length, output length.

stride_l, pad_l, dilation_lint

Match the convolution forward.