Fork me on GitHub

src/arraymancer/nn_primitives/fallback/conv

  Source Edit

Procs

proc col2im[T](input: Tensor[T]; channels, height, width: int;
               kernel_size: Size2D; padding: Size2D = (0, 0);
               stride: Size2D = (1, 1)): Tensor[T]
Convert blocks of an image from columns back to an image, collapsed pixels are summed   Source Edit
proc im2col[T](input: Tensor[T]; kernel_size: Size2D; padding: Size2D = (0, 0);
               stride: Size2D = (1, 1); result: var Tensor[T])
Convert blocks of an image into columns, useful for preprocessing an image before convolutions   Source Edit
proc im2colgemm_conv2d[T](input, kernel, bias: Tensor[T];
                          padding: Size2D = (0, 0); stride: Size2D = (1, 1)): Tensor[
    T]
Compute cross-correlate for image with the given kernel weights   Source Edit
proc im2colgemm_conv2d_gradient[T](input, kernel: Tensor[T];
                                   padding: Size2D = (0, 0);
                                   stride: Size2D = (1, 1);
                                   grad_output: Tensor[T];
                                   grad_input, grad_weight: var Tensor[T])
Computes gradients w.r.t input and weights for a 2D convolution   Source Edit
Arraymancer Technical reference Tutorial Spellbook (How-To's) Under the hood