Trait coaster_nn::Convolution

pub trait Convolution<F>: NN<F> {
    fn new_convolution_config(
        &self, 
        src: &SharedTensor<F>, 
        dest: &SharedTensor<F>, 
        filter: &SharedTensor<F>, 
        algo_fwd: ConvForwardAlgo, 
        algo_bwd_filter: ConvBackwardFilterAlgo, 
        algo_bwd_data: ConvBackwardDataAlgo, 
        stride: &[i32], 
        zero_padding: &[i32]
    ) -> Result<Self::CC, Error>;
    fn convolution(
        &self, 
        filter: &SharedTensor<F>, 
        x: &SharedTensor<F>, 
        result: &mut SharedTensor<F>, 
        workspace: &mut SharedTensor<u8>, 
        config: &Self::CC
    ) -> Result<(), Error>;
    fn convolution_grad_filter(
        &self, 
        src_data: &SharedTensor<F>, 
        dest_diff: &SharedTensor<F>, 
        filter_diff: &mut SharedTensor<F>, 
        workspace: &mut SharedTensor<u8>, 
        config: &Self::CC
    ) -> Result<(), Error>;
    fn convolution_grad_data(
        &self, 
        filter: &SharedTensor<F>, 
        x_diff: &SharedTensor<F>, 
        result_diff: &mut SharedTensor<F>, 
        workspace: &mut SharedTensor<u8>, 
        config: &Self::CC
    ) -> Result<(), Error>;
}

Provides the functionality for a Backend to support Convolution operations.

Required Methods

fn new_convolution_config(
    &self,
    src: &SharedTensor<F>,
    dest: &SharedTensor<F>,
    filter: &SharedTensor<F>,
    algo_fwd: ConvForwardAlgo,
    algo_bwd_filter: ConvBackwardFilterAlgo,
    algo_bwd_data: ConvBackwardDataAlgo,
    stride: &[i32],
    zero_padding: &[i32]
) -> Result<Self::CC, Error>

Creates a new ConvolutionConfig, which needs to be passed to further convolution Operations.

fn convolution(
    &self,
    filter: &SharedTensor<F>,
    x: &SharedTensor<F>,
    result: &mut SharedTensor<F>,
    workspace: &mut SharedTensor<u8>,
    config: &Self::CC
) -> Result<(), Error>

Computes a CNN convolution over the input Tensor x.

Saves the result to result.

fn convolution_grad_filter(
    &self,
    src_data: &SharedTensor<F>,
    dest_diff: &SharedTensor<F>,
    filter_diff: &mut SharedTensor<F>,
    workspace: &mut SharedTensor<u8>,
    config: &Self::CC
) -> Result<(), Error>

Computes the gradient of a CNN convolution with respect to the filter.

Saves the result to filter_diff.

fn convolution_grad_data(
    &self,
    filter: &SharedTensor<F>,
    x_diff: &SharedTensor<F>,
    result_diff: &mut SharedTensor<F>,
    workspace: &mut SharedTensor<u8>,
    config: &Self::CC
) -> Result<(), Error>

Computes the gradient of a CNN convolution over the input Tensor x with respect to the data.

Saves the result to result_diff.

Implementors

• impl<T> Convolution<T> for Backend<Native> where
    T: Add<T, Output = T> + Mul<T, Output = T> + Default + Copy, 
• impl<T> Convolution<T> for Backend<Cuda> where
    T: Float + DataTypeInfo,