3D Transposed Convolution

Applies a 3D transposed convolution (also called deconvolution) over an input signal. Useful for 3D volumetric data upsampling, such as in medical imaging or video generation.

Usage

mlx_conv_transpose3d(
  input,
  weight,
  stride = c(1L, 1L, 1L),
  padding = c(0L, 0L, 0L),
  dilation = c(1L, 1L, 1L),
  output_padding = c(0L, 0L, 0L),
  groups = 1L
)

Arguments

input

Input mlx array. Shape depends on dimensionality (see individual functions).

weight

Weight array. Shape depends on dimensionality (see individual functions).

stride

Stride of the convolution. Can be a scalar or vector (length depends on dimensionality). Default: 1 for 1D, c(1,1) for 2D, c(1,1,1) for 3D.

padding

Amount of zero padding. Can be a scalar or vector (length depends on dimensionality). Default: 0 for 1D, c(0,0) for 2D, c(0,0,0) for 3D.

dilation

Spacing between kernel elements. Can be a scalar or vector (length depends on dimensionality). Default: 1 for 1D, c(1,1) for 2D, c(1,1,1) for 3D.

output_padding

Additional size added to output shape. Can be a scalar or length-3 vector. Default: c(0, 0, 0)

groups

Number of blocked connections from input to output channels. Default: 1.

Value

An mlx array with the transposed convolution result

Details

Input has shape (batch, depth, height, width, in_channels) for 'NDHWC' layout. Weight has shape (out_channels, kernel_d, kernel_h, kernel_w, in_channels).

See also

mlx_conv3d(), mlx_conv_transpose1d(), mlx_conv_transpose2d()

mlx.nn