2D Convolution

Torch-style 2D convolution in NCHW layout: x is [batch, in_channels, height, width], weight is [out_channels, in_channels / groups, kh, kw], output is [batch, out_channels, out_h, out_w]. Symmetric zero padding.

Usage

nv_conv2d(
  x,
  weight,
  stride = 1L,
  padding = 0L,
  dilation = 1L,
  groups = 1L,
  precision = "highest"
)

Arguments

x

(arrayish)
[N, C_in, H, W]. Can be any data type; x and weight are promoted to a common data type. An R value assumes the other operand's data type, and materializes at its default data type when that has none either.

weight

(arrayish)
[C_out, C_in / groups, kH, kW]. Promoted together with x – see x.

stride

(integer())
Length 1 or 2.

padding

(integer())
Symmetric padding, length 1 or 2.

dilation

(integer())
Kernel dilation, length 1 or 2.

groups

(integer(1))
Grouped/depthwise convolution.

precision

(character(1))
"highest", "high" or "default".

Value

(arrayish)
Has the operands' common data type, and shape [N, C_out, out_H, out_W].

See also

nv_conv1d(), nv_conv3d(), prim_convolution().

Examples

# one batch, one channel, 4x4, convolved with a 3x3 kernel
x <- nv_array(1:16, shape = c(1, 1, 4, 4), dtype = "f32")
weight <- nv_fill(1, shape = c(1, 1, 3, 3), dtype = "f32")
nv_conv2d(x, weight)
#> AnvlArray
#> (1,1,.,.) =
#>  54 90
#>  63 99
#> [ CPUf32{1,1,2,2} ] 

# two output channels give a result with two channels
weight2 <- nv_fill(1, shape = c(2, 1, 3, 3), dtype = "f32")
shape(nv_conv2d(x, weight2))
#> [1] 1 2 2 2