ai.djl.nn.convolutional.Conv2d Maven / Gradle / Ivy
/*
* Copyright 2019 Amazon.com, Inc. or its affiliates. All Rights Reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License"). You may not use this file except in compliance
* with the License. A copy of the License is located at
*
* http://aws.amazon.com/apache2.0/
*
* or in the "license" file accompanying this file. This file is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES
* OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions
* and limitations under the License.
*/
package ai.djl.nn.convolutional;
import ai.djl.ndarray.NDArray;
import ai.djl.ndarray.NDList;
import ai.djl.ndarray.types.LayoutType;
import ai.djl.ndarray.types.Shape;
import ai.djl.nn.Block;
import ai.djl.util.Preconditions;
/**
* Being the pioneer of convolution layers, {@code Conv2d} layer works on two dimensions of input,
* {@link LayoutType#WIDTH} and {@link LayoutType#HEIGHT} as usually a {@code Conv2d} layer is used
* to process data with two spatial dimensions, namely image. The concept itself works just as how
* {@link Convolution} does, and each filter slides through an input data by two directions, first
* traversing the {@link 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 {@link Conv1d} is now available for use.
*
*
The input to a {@code Conv2d} is an {@link ai.djl.ndarray.NDList} with a single 4-D {@link
* ai.djl.ndarray.NDArray}. The layout of the {@link ai.djl.ndarray.NDArray} must be "NCHW". The
* shapes are
*
*
* - {@code data: (batch_size, channel, height, width)}
*
- {@code weight: (num_filter, channel, kernel[0], kernel[1])}
*
- {@code bias: (num_filter,)}
*
- {@code out: (batch_size, num_filter, out_height, out_width)}
* {@code out_height = f(height, kernel[0], pad[0], stride[0], dilate[0])}
* {@code out_width = f(width, kernel[1], pad[1], stride[1], dilate[1])}
* {@code where f(x, k, p, s, d) = floor((x + 2 * p - d * (k - 1) - 1)/s) + 1}
*
*
* Both {@code weight} and {@code bias} are learn-able parameters.
*
* @see Convolution
*/
public class Conv2d extends Convolution {
private static final LayoutType[] EXPECTED_LAYOUT = {
LayoutType.BATCH, LayoutType.CHANNEL, LayoutType.HEIGHT, LayoutType.WIDTH
};
private static final String STRING_LAYOUT = "NCHW";
private static final int NUM_DIMENSIONS = 4;
Conv2d(Builder builder) {
super(builder);
}
/** {@inheritDoc} */
@Override
protected LayoutType[] getExpectedLayout() {
return EXPECTED_LAYOUT;
}
/** {@inheritDoc} */
@Override
protected String getStringLayout() {
return STRING_LAYOUT;
}
/** {@inheritDoc} */
@Override
protected int numDimensions() {
return NUM_DIMENSIONS;
}
/**
* Applies 2D convolution over an input signal composed of several input planes.
*
* @param input the input {@code NDArray} of shape (batchSize, inputChannel, height, width)
* @param weight filters {@code NDArray} of shape (outChannel, inputChannel/groups, height,
* width)
* @return the output of the conv2d operation
*/
public static NDList conv2d(NDArray input, NDArray weight) {
return conv2d(input, weight, null, new Shape(1, 1), new Shape(0, 0), new Shape(1, 1));
}
/**
* Applies 2D convolution over an input signal composed of several input planes.
*
* @param input the input {@code NDArray} of shape (batchSize, inputChannel, height, width)
* @param weight filters {@code NDArray} of shape (outChannel, inputChannel/groups, height,
* width)
* @param bias bias {@code NDArray} of shape (outChannel)
* @return the output of the conv2d operation
*/
public static NDList conv2d(NDArray input, NDArray weight, NDArray bias) {
return conv2d(input, weight, bias, new Shape(1, 1), new Shape(0, 0), new Shape(1, 1));
}
/**
* Applies 2D convolution over an input signal composed of several input planes.
*
* @param input the input {@code NDArray} of shape (batchSize, inputChannel, height, width)
* @param weight filters {@code NDArray} of shape (outChannel, inputChannel/groups, height,
* width)
* @param bias bias {@code NDArray} of shape (outChannel)
* @param stride the stride of the convolving kernel: Shape(height, width)
* @return the output of the conv2d operation
*/
public static NDList conv2d(NDArray input, NDArray weight, NDArray bias, Shape stride) {
return conv2d(input, weight, bias, stride, new Shape(0, 0), new Shape(1, 1));
}
/**
* Applies 2D convolution over an input signal composed of several input planes.
*
* @param input the input {@code NDArray} of shape (batchSize, inputChannel, height, width)
* @param weight filters {@code NDArray} of shape (outChannel, inputChannel/groups, height,
* width)
* @param bias bias {@code NDArray} of shape (outChannel)
* @param stride the stride of the convolving kernel: Shape(height, width)
* @param padding implicit paddings on both sides of the input: Shape(height, width)
* @return the output of the conv2d operation
*/
public static NDList conv2d(
NDArray input, NDArray weight, NDArray bias, Shape stride, Shape padding) {
return conv2d(input, weight, bias, stride, padding, new Shape(1, 1));
}
/**
* Applies 2D convolution over an input signal composed of several input planes.
*
* @param input the input {@code NDArray} of shape (batchSize, inputChannel, height, width)
* @param weight filters {@code NDArray} of shape (outChannel, inputChannel/groups, height,
* width)
* @param bias bias {@code NDArray} of shape (outChannel)
* @param stride the stride of the convolving kernel: Shape(height, width)
* @param padding implicit paddings on both sides of the input: Shape(height, width)
* @param dilation the spacing between kernel elements: Shape(height, width)
* @return the output of the conv2d operation
*/
public static NDList conv2d(
NDArray input,
NDArray weight,
NDArray bias,
Shape stride,
Shape padding,
Shape dilation) {
return conv2d(input, weight, bias, stride, padding, dilation, 1);
}
/**
* Applies 2D convolution over an input signal composed of several input planes.
*
* @param input the input {@code NDArray} of shape (batchSize, inputChannel, height, width)
* @param weight filters {@code NDArray} of shape (outChannel, inputChannel/groups, height,
* width)
* @param bias bias {@code NDArray} of shape (outChannel)
* @param stride the stride of the convolving kernel: Shape(height, width)
* @param padding implicit paddings on both sides of the input: Shape(height, width)
* @param dilation the spacing between kernel elements: Shape(height, width)
* @param groups split input into groups: input channel(input.size(1)) should be divisible by
* the number of groups
* @return the output of the conv2d operation
*/
public static NDList conv2d(
NDArray input,
NDArray weight,
NDArray bias,
Shape stride,
Shape padding,
Shape dilation,
int groups) {
Preconditions.checkArgument(
input.getShape().dimension() == 4 && weight.getShape().dimension() == 4,
"the shape of input or weight doesn't match the conv2d");
Preconditions.checkArgument(
stride.dimension() == 2 && padding.dimension() == 2 && dilation.dimension() == 2,
"the shape of stride or padding or dilation doesn't match the conv2d");
return Convolution.convolution(input, weight, bias, stride, padding, dilation, groups);
}
/**
* Creates a builder to build a {@code Conv2d}.
*
* @return a new builder
*/
public static Builder builder() {
return new Builder();
}
/** The Builder to construct a {@link Conv2d} type of {@link Block}. */
public static final class Builder extends ConvolutionBuilder {
/** Creates a builder that can build a {@link Conv2d} block. */
Builder() {
stride = new Shape(1, 1);
padding = new Shape(0, 0);
dilation = new Shape(1, 1);
}
/** {@inheritDoc} */
@Override
protected Builder self() {
return this;
}
/**
* Builds a {@link Conv2d} block.
*
* @return the {@link Conv2d} block
*/
public Conv2d build() {
validate();
return new Conv2d(this);
}
}
}