conv_transpose2d_op

→TensorImplPtr

int conv_transpose2d_op(const int & x, const int & W, const int & b, int stride_h, int stride_w, int pad_h, int pad_w, int opad_h, int opad_w)

source edit

Python wrappers

pylucid.nn.functional.conv_transpose2dfunction

Two-dimensional transposed convolution (fractionally-strided conv).

The 2-D upsampling primitive used by VAE / GAN decoders, U-Net expansion paths, and super-resolution heads. Equivalent to the gradient of a 2-D convolution w.r.t. its input.

Parameters

xTensorImplPtr

Input of shape (B, C_in, H, W).

WTensorImplPtr

Filters of shape (C_in, C_out, KH, KW).

bTensorImplPtr

Bias of shape (C_out,) or empty.

stride_h, stride_wint= None

Per-axis upsampling factor. Default: 1.

pad_h, pad_wint= None

Per-axis padding subtracted from the output extent. Default: 0.

opad_h, opad_wint= None

Per-axis one-sided output padding. Default: 0.

Returns

TensorImplPtr

Output of shape (B, C_out, H_out, W_out) with $H_\text{out} = (H - 1) s_h - 2 p_h + KH + p^\text{out}_h$ and analogously for $W_\text{out}$ .

Notes

When stride > 1 and kernel_size is not divisible by stride the output can exhibit characteristic checkerboard artefacts. A common mitigation is to choose kernel_size = stride * n or to substitute bilinear upsampling followed by a regular convolution.