fold_op

→TensorImplPtr

int fold_op(const int & x, const int & output_size, const int & kernel_size, const int & stride, const int & padding, const int & dilation)

source edit

Python wrappers

pylucid.nn.functional.foldfunction

Fold (col2im) — the inverse of Unfold for 2-D inputs.

Scatters each column back into its source location in the spatial grid; values at overlapping positions are summed (this is the formal adjoint of Unfold, not a true inverse when stride permits overlap).

Parameters

xTensorImplPtr

Column matrix of shape (N, C * kH * kW, L) produced by unfold_op or any equivalent column layout.

output_sizevector<int>

Target spatial size (outH, outW).

kernel_sizevector<int>

Kernel size per spatial axis used during the matching unfold.

stridevector<int>

Stride per spatial axis.

paddingvector<int>

Zero-padding per spatial axis.

dilationvector<int>

Dilation per spatial axis.

Returns

TensorImplPtr

Output of shape (N, C, outH, outW) with overlapping contributions summed.