conv_data_backward

burn_cubecl::kernel::conv

Function conv_data_backward 

Source 
pub fn conv_data_backward<R: CubeRuntime, const N: usize>(
    out_grad: CubeTensor<R>,
    weights: CubeTensor<R>,
    in_shape: Shape,
    options: ConvOptions<N>,
    strategy: ConvStrategy,
) -> Result<CubeTensor<R>, ConvSetupError>
Expand description

Performs an N-dimensional convolution backwards data pass with the given strategy

  • input - The input feature map
  • weight - The weights (filter) applied to each kernel
  • in_shape - The shape of the input to the layer
  • options - The options to use for the convolution
  • strategy - The convolution algorithm to use. Autotune will pick the fastest available option.