Struct ConvolutionLayer 

pub struct ConvolutionLayer<F>
where
    F: Float,
{ /* private fields */ }

Rust implementation of a convolutional layer. The weight matrix shall have dimension (in that order) (input channels, output channels, kernel width, kernel height), to comply with the order in which pytorch weights are saved.

Implementations

impl<F> ConvolutionLayer<F>

where F: 'static + Float + AddAssign,

pub fn new( weights: ArrayBase<OwnedRepr<F>, Dim<[usize; 4]>>, bias_array: Option<ArrayBase<OwnedRepr<F>, Dim<[usize; 1]>>>, stride: usize, padding: Padding, ) -> ConvolutionLayer<F>

Creates new convolution layer. The weights are given in Pytorch layout. (out channels, in channels, kernel height, kernel width) Bias: (output height * output width, 1)

pub fn new_tf( weights: ArrayBase<OwnedRepr<F>, Dim<[usize; 4]>>, bias_array: Option<ArrayBase<OwnedRepr<F>, Dim<[usize; 1]>>>, stride: usize, padding: Padding, ) -> ConvolutionLayer<F>

Creates new convolution layer. The weights are given in Tensorflow layout. (kernel height, kernel width, in channels, out channels)

pub fn convolve( &self, image: &ArrayBase<OwnedRepr<F>, Dim<[usize; 3]>>, ) -> ArrayBase<OwnedRepr<F>, Dim<[usize; 3]>>

Analog to conv2d.

Trait Implementations

impl<F: FloatLikePrimitive> Layer<ArrayBase<OwnedRepr<F>, Dim<[usize; 3]>>, ArrayBase<OwnedRepr<F>, Dim<[usize; 3]>>> for ConvolutionLayer<F>

fn forward_pass(&self, input: &Array3<F>) -> Array3<F>