Convolutions

Forward and backward convolution passes. More...

Functions
AFAPI array	convolve2GradientNN (const array &incoming_gradient, const array &original_signal, const array &original_filter, const array &convolved_output, const dim4 stride, const dim4 padding, const dim4 dilation, convGradientType grad_type)
	C++ interface for calculating backward pass gradient of 2D convolution This function calculates the gradient with respect to the output of the convolve2NN function that uses the machine learning formulation for the dimensions of the signals and filters. More...
AFAPI af_err	af_convolve2_gradient_nn (af_array *out, const af_array incoming_gradient, const af_array original_signal, const af_array original_filter, const af_array convolved_output, const unsigned stride_dims, const dim_t *strides, const unsigned padding_dims, const dim_t *paddings, const unsigned dilation_dims, const dim_t *dilations, af_conv_gradient_type grad_type)
	C interface for calculating backward pass gradient of 2D convolution This function calculates the gradient with respect to the output of the af::convolve2NN() function that uses the machine learning formulation for the dimensions of the signals and filters. More...

Detailed Description

Forward and backward convolution passes.

Function Documentation

af_convolve2_gradient_nn()

AFAPI af_err af_convolve2_gradient_nn	(	af_array *	out,
		const af_array	incoming_gradient,
		const af_array	original_signal,
		const af_array	original_filter,
		const af_array	convolved_output,
		const unsigned	stride_dims,
		const dim_t *	strides,
		const unsigned	padding_dims,
		const dim_t *	paddings,
		const unsigned	dilation_dims,
		const dim_t *	dilations,
		af_conv_gradient_type	grad_type
	)

C interface for calculating backward pass gradient of 2D convolution This function calculates the gradient with respect to the output of the af::convolve2NN() function that uses the machine learning formulation for the dimensions of the signals and filters.

Parameters

[out]	out	gradient wrt/gradType
[in]	incoming_gradient	gradients to be distributed in backwards pass
[in]	original_signal	input signal to forward pass of convolution assumed structure of input is ( d0 x d1 x d2 x N )
[in]	original_filter	input filter to forward pass of convolution assumed structure of input is ( d0 x d1 x d2 x N )
[in]	convolved_output	output from forward pass of convolution
[in]	stride_dims	specifies number of stride dimensions
[in]	strides	array of stride values
[in]	padding_dims	number of padding dimensions
[in]	paddings	array of padding values
[in]	dilation_dims	number of dilation dimensions
[in]	dilations	array of dilation values
[in]	grad_type	specifies which gradient to return

Returns

AF_SUCCESS if the execution completes properly

convolve2GradientNN()

AFAPI array convolve2GradientNN	(	const array &	incoming_gradient,
		const array &	original_signal,
		const array &	original_filter,
		const array &	convolved_output,
		const dim4	stride,
		const dim4	padding,
		const dim4	dilation,
		convGradientType	grad_type
	)

C++ interface for calculating backward pass gradient of 2D convolution This function calculates the gradient with respect to the output of the convolve2NN function that uses the machine learning formulation for the dimensions of the signals and filters.

Parameters

[in]	incoming_gradient	gradients to be distributed in backwards pass
[in]	original_signal	input signal to forward pass of convolution assumed structure of input is ( d0 x d1 x d2 x N )
[in]	original_filter	input filter to forward pass of convolution assumed structure of input is ( d0 x d1 x d2 x N )
[in]	convolved_output	output from forward pass of convolution
[in]	stride	specifies strides along each dimension for original convolution
[in]	padding	specifies padding width along each dimension for original convolution
[in]	dilation	specifies filter dilation along each dimension for original convolution
[in]	grad_type	specifies which gradient to return

Returns

gradient wrt/grad_type

Note

Make sure you pass in both dim0, and dim1 in your dim4 arguments. The third and fourth dimensions are currently ignored.