Package ai.djl.nn.convolutional

Class Conv2d

java.lang.Object

ai.djl.nn.AbstractBaseBlock

ai.djl.nn.AbstractBlock

ai.djl.nn.convolutional.Convolution

ai.djl.nn.convolutional.Conv2d

All Implemented Interfaces:

Block

public class Conv2d
extends Convolution

Being the pioneer of convolution layers, Conv2d layer works on two dimensions of input, LayoutType.WIDTH and LayoutType.HEIGHT as usually a Conv2d layer is used to process data with two spatial dimensions, namely image. The concept itself works just as how Convolution does, and each filter slides through an input data by two directions, first traversing the LayoutType.WIDTH then traverses each row of the data.

First proposed by LeCun et al.'s paper, 2-dimensional convolution layer gained its rising interest with the publication of paper about AlexNet for image classification task. It is still commonly used in image-related tasks and adapted in other tasks, including but not limited to 1-dimensional data which may be transformed to 2-dimensional data, though Conv1d is now available for use.

The input to a Conv2d is an NDList with a single 4-D NDArray. The layout of the NDArray must be "NCHW". The shapes are

• data: (batch_size, channel, height, width)
• weight: (num_filter, channel, kernel[0], kernel[1])
• bias: (num_filter,)
• out: (batch_size, num_filter, out_height, out_width)
out_height = f(height, kernel[0], pad[0], stride[0], dilate[0])
out_width = f(width, kernel[1], pad[1], stride[1], dilate[1])
where f(x, k, p, s, d) = floor((x + 2 * p - d * (k - 1) - 1)/s) + 1

Both weight and bias are learn-able parameters.

See Also:

Convolution

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static class	Conv2d.Builder	The Builder to construct a Conv2d type of Block.

Nested classes/interfaces inherited from class ai.djl.nn.convolutional.Convolution

Convolution.ConvolutionBuilder<T extends Convolution.ConvolutionBuilder>

Field Summary

Fields inherited from class ai.djl.nn.convolutional.Convolution

bias, dilation, filters, groups, includeBias, kernelShape, padding, stride, weight

Fields inherited from class ai.djl.nn.AbstractBlock

children, parameters

Fields inherited from class ai.djl.nn.AbstractBaseBlock

inputNames, inputShapes, outputDataTypes, version

Constructor Summary

Constructors

Modifier	Constructor	Description
protected	Conv2d(Conv2d.Builder builder)

Method Summary

Modifier and Type	Method	Description
static Conv2d.Builder	builder()	Creates a builder to build a Conv2d.
static NDList	conv2d(NDArray input, NDArray weight)	Applies 2D convolution over an input signal composed of several input planes.
static NDList	conv2d(NDArray input, NDArray weight, NDArray bias)	Applies 2D convolution over an input signal composed of several input planes.
static NDList	conv2d(NDArray input, NDArray weight, NDArray bias, Shape stride)	Applies 2D convolution over an input signal composed of several input planes.
static NDList	conv2d(NDArray input, NDArray weight, NDArray bias, Shape stride, Shape padding)	Applies 2D convolution over an input signal composed of several input planes.
static NDList	conv2d(NDArray input, NDArray weight, NDArray bias, Shape stride, Shape padding, Shape dilation)	Applies 2D convolution over an input signal composed of several input planes.
static NDList	conv2d(NDArray input, NDArray weight, NDArray bias, Shape stride, Shape padding, Shape dilation, int groups)	Applies 2D convolution over an input signal composed of several input planes.
protected LayoutType[]	getExpectedLayout()	Returns the expected layout of the input.
protected java.lang.String	getStringLayout()	Returns the string representing the layout of the input.
protected int	numDimensions()	Returns the number of dimensions of the input.

Methods inherited from class ai.djl.nn.convolutional.Convolution

beforeInitialize, forwardInternal, getDilation, getFilters, getGroups, getKernelShape, getOutputShapes, getPadding, getStride, isIncludeBias, loadMetadata, prepare

Methods inherited from class ai.djl.nn.AbstractBlock

addChildBlock, addChildBlock, addChildBlockSingleton, addParameter, getChildren, getDirectParameters

Methods inherited from class ai.djl.nn.AbstractBaseBlock

cast, clear, describeInput, forward, forward, forwardInternal, getInputShapes, getOutputDataTypes, getParameters, initialize, initializeChildBlocks, isInitialized, loadParameters, readInputShapes, saveInputShapes, saveMetadata, saveParameters, setInitializer, setInitializer, setInitializer, toString

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, wait, wait, wait

Methods inherited from interface ai.djl.nn.Block

forward, freezeParameters, getOutputShapes

Constructor Detail

Conv2d

protected Conv2d(Conv2d.Builder builder)

Method Detail

getExpectedLayout

protected LayoutType[] getExpectedLayout()

Returns the expected layout of the input.

Specified by:

getExpectedLayout in class Convolution

Returns:

the expected layout of the input

getStringLayout

protected java.lang.String getStringLayout()

Returns the string representing the layout of the input.

Specified by:

getStringLayout in class Convolution

Returns:

the string representing the layout of the input

numDimensions

protected int numDimensions()

Returns the number of dimensions of the input.

Specified by:

numDimensions in class Convolution

Returns:

the number of dimensions of the input

conv2d

public static NDList conv2d(NDArray input,
                            NDArray weight)

Applies 2D convolution over an input signal composed of several input planes.

Parameters:

input - the input NDArray of shape (batchSize, inputChannel, height, width)

weight - filters NDArray of shape (outChannel, inputChannel/groups, height, width)

Returns:

the output of the conv2d operation

conv2d

public static NDList conv2d(NDArray input,
                            NDArray weight,
                            NDArray bias)

Applies 2D convolution over an input signal composed of several input planes.

Parameters:

input - the input NDArray of shape (batchSize, inputChannel, height, width)

weight - filters NDArray of shape (outChannel, inputChannel/groups, height, width)

bias - bias NDArray of shape (outChannel)

Returns:

the output of the conv2d operation

conv2d

public static NDList conv2d(NDArray input,
                            NDArray weight,
                            NDArray bias,
                            Shape stride)

Applies 2D convolution over an input signal composed of several input planes.

Parameters:

input - the input NDArray of shape (batchSize, inputChannel, height, width)

weight - filters NDArray of shape (outChannel, inputChannel/groups, height, width)

bias - bias NDArray of shape (outChannel)

stride - the stride of the convolving kernel: Shape(height, width)

Returns:

the output of the conv2d operation

conv2d

public static NDList conv2d(NDArray input,
                            NDArray weight,
                            NDArray bias,
                            Shape stride,
                            Shape padding)

Applies 2D convolution over an input signal composed of several input planes.

Parameters:

input - the input NDArray of shape (batchSize, inputChannel, height, width)

weight - filters NDArray of shape (outChannel, inputChannel/groups, height, width)

bias - bias NDArray of shape (outChannel)

stride - the stride of the convolving kernel: Shape(height, width)

padding - implicit paddings on both sides of the input: Shape(height, width)

Returns:

the output of the conv2d operation

conv2d

public static NDList conv2d(NDArray input,
                            NDArray weight,
                            NDArray bias,
                            Shape stride,
                            Shape padding,
                            Shape dilation)

Applies 2D convolution over an input signal composed of several input planes.

Parameters:

input - the input NDArray of shape (batchSize, inputChannel, height, width)

weight - filters NDArray of shape (outChannel, inputChannel/groups, height, width)

bias - bias NDArray of shape (outChannel)

stride - the stride of the convolving kernel: Shape(height, width)

padding - implicit paddings on both sides of the input: Shape(height, width)

dilation - the spacing between kernel elements: Shape(height, width)

Returns:

the output of the conv2d operation

conv2d

public static NDList conv2d(NDArray input,
                            NDArray weight,
                            NDArray bias,
                            Shape stride,
                            Shape padding,
                            Shape dilation,
                            int groups)

Applies 2D convolution over an input signal composed of several input planes.

Parameters:

input - the input NDArray of shape (batchSize, inputChannel, height, width)

weight - filters NDArray of shape (outChannel, inputChannel/groups, height, width)

bias - bias NDArray of shape (outChannel)

stride - the stride of the convolving kernel: Shape(height, width)

padding - implicit paddings on both sides of the input: Shape(height, width)

dilation - the spacing between kernel elements: Shape(height, width)

groups - split input into groups: input channel(input.size(1)) should be divisible by the number of groups

Returns:

the output of the conv2d operation

builder

public static Conv2d.Builder builder()

Creates a builder to build a Conv2d.

Returns:

a new builder