Convolutional Layers

CNN layers in Deeplearning4j — Conv1D/2D/3D, pooling, deconvolution, depthwise, separable, and upsampling layers

Overview

Deeplearning4j provides a complete set of convolutional layers covering 1D (sequences), 2D (images), and 3D (video/volumetric) convolutions. All CNN layers share common builder patterns and integrate with setInputType() for automatic shape inference and pre-processor insertion.

---

ConvolutionMode

ConvolutionMode controls how output spatial dimensions are computed relative to input size. Set it with .convolutionMode(ConvolutionMode.X) on any convolutional layer.

Mode	Description
Truncated (default)	Output may be smaller than input if padding doesn't divide evenly; no truncation warning
Same	Output has the same spatial size as the input (appropriate zero-padding is added automatically)
Full	Output is larger than input; full convolution with padding
Causal	Causal (1D only): no future information leaks into past time steps

import org.deeplearning4j.nn.conf.ConvolutionMode;

new ConvolutionLayer.Builder(3, 3)
    .convolutionMode(ConvolutionMode.Same)
    .nIn(64).nOut(128)
    .build()

---

ConvolutionLayer (Conv2D)

Class: org.deeplearning4j.nn.conf.layers.ConvolutionLayer Source: ConvolutionLayer.java

Standard 2D convolution layer. Input shape: [minibatch, channels, height, width].

Builder Parameters

Parameter	Type	Default	Description
constructor (int, int)	int, int	required	Kernel height and width
nIn	int	auto	Number of input channels
nOut	int	required	Number of output filters
stride(int, int)	int, int	1, 1	Stride in height and width
padding(int, int)	int, int	0, 0	Zero-padding in height and width
dilation(int, int)	int, int	1, 1	Dilation (atrous convolution)
activation	Activation	RELU	Activation function
hasBias	boolean	true	Include bias
convolutionMode	ConvolutionMode	Truncated	How to handle border effects
cudnnAlgoMode	AlgoMode	PREFER_FASTEST	CuDNN algorithm selection

Example — Basic Conv2D

new ConvolutionLayer.Builder(3, 3)
    .nIn(3)
    .nOut(64)
    .stride(1, 1)
    .padding(1, 1)
    .activation(Activation.RELU)
    .build()

Example — Dilated (Atrous) Convolution

new ConvolutionLayer.Builder(3, 3)
    .nIn(64).nOut(64)
    .dilation(2, 2)          // receptive field = 5x5 without extra parameters
    .convolutionMode(ConvolutionMode.Same)
    .activation(Activation.RELU)
    .build()

Full Network with setInputType

int channels = 3, height = 32, width = 32, numClasses = 10;

MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
    .seed(42)
    .dataType(DataType.FLOAT)
    .updater(new Adam(1e-3))
    .weightInit(WeightInit.XAVIER)
    .list()
    .layer(new ConvolutionLayer.Builder(3, 3)
        .nIn(channels).nOut(32)
        .stride(1, 1).padding(1, 1)
        .activation(Activation.RELU).build())
    .layer(new SubsamplingLayer.Builder(PoolingType.MAX)
        .kernelSize(2, 2).stride(2, 2).build())
    .layer(new ConvolutionLayer.Builder(3, 3)
        .nOut(64).stride(1, 1).padding(1, 1)
        .activation(Activation.RELU).build())
    .layer(new SubsamplingLayer.Builder(PoolingType.MAX)
        .kernelSize(2, 2).stride(2, 2).build())
    .layer(new DenseLayer.Builder().nOut(256).activation(Activation.RELU).build())
    .layer(new OutputLayer.Builder(LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD)
        .nOut(numClasses).activation(Activation.SOFTMAX).build())
    // Automatically infers nIn for every layer and inserts CnnToFeedForwardPreProcessor
    .setInputType(InputType.convolutional(height, width, channels))
    .build();

---

Convolution1DLayer (Conv1D)

Class: org.deeplearning4j.nn.conf.layers.Convolution1D (alias: Convolution1DLayer) Source: Convolution1D.java

1D convolution for sequence data. Input shape: [minibatch, channels, sequenceLength].

Builder Parameters

Parameter	Type	Default	Description
constructor (int)	int	required	Kernel size (length)
nIn	int	auto	Input channels (embedding dim)
nOut	int	required	Number of filters
stride(int)	int	1	Stride along sequence
padding(int)	int	0	Zero-padding on each side
dilation(int)	int	1	Dilation along sequence
convolutionMode	ConvolutionMode	Truncated	Border handling

Example — Text Classification

.layer(new EmbeddingSequenceLayer.Builder()
    .nIn(vocabSize).nOut(128).inputLength(maxLen).build())
.layer(new Convolution1D.Builder(3)      // trigram-style kernel
    .nOut(128).activation(Activation.RELU)
    .convolutionMode(ConvolutionMode.Same).build())
.layer(new GlobalPoolingLayer.Builder(PoolingType.MAX).build())
.layer(new OutputLayer.Builder(LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD)
    .nIn(128).nOut(numClasses).activation(Activation.SOFTMAX).build())

---

Convolution3D

Class: org.deeplearning4j.nn.conf.layers.Convolution3D Source: Convolution3D.java

3D convolution for volumetric data (e.g., video, medical imaging). Supports two data formats.

Builder Parameters

Parameter	Type	Default	Description
kernelSize(int...)	int[3]	required	Kernel in (depth, height, width)
stride(int...)	int[3]	1,1,1	Stride in (depth, height, width)
padding(int...)	int[3]	0,0,0	Padding in (depth, height, width)
dilation(int...)	int[3]	1,1,1	Dilation in (depth, height, width)
dataFormat	DataFormat	NCDHW	NCDHW (channels first) or NDHWC (channels last)
hasBias	boolean	true	Include bias

Example

new Convolution3D.Builder()
    .kernelSize(3, 3, 3)
    .stride(1, 1, 1)
    .padding(1, 1, 1)
    .nIn(16).nOut(32)
    .dataFormat(Convolution3D.DataFormat.NCDHW)
    .activation(Activation.RELU)
    .build()

---

SubsamplingLayer (Pooling2D)

Class: org.deeplearning4j.nn.conf.layers.SubsamplingLayer Source: SubsamplingLayer.java

Standard 2D pooling layer (max, average, sum, or p-norm). Input shape: [minibatch, channels, height, width]. No learnable parameters.

Builder Parameters

Parameter	Type	Default	Description
constructor PoolingType	PoolingType	required	MAX, AVG, SUM, PNORM, NONE
kernelSize(int, int)	int, int	required	Pooling window height, width
stride(int, int)	int, int	same as kernelSize	Stride
padding(int, int)	int, int	0, 0	Padding
convolutionMode	ConvolutionMode	Truncated	Border handling
pnorm	int	2	P-norm exponent (only for PNORM)

Example

// Max pooling 2x2 with stride 2 (halves spatial dimensions)
new SubsamplingLayer.Builder(SubsamplingLayer.PoolingType.MAX)
    .kernelSize(2, 2)
    .stride(2, 2)
    .build()

// Average pooling 3x3 same-padded
new SubsamplingLayer.Builder(SubsamplingLayer.PoolingType.AVG)
    .kernelSize(3, 3)
    .stride(1, 1)
    .convolutionMode(ConvolutionMode.Same)
    .build()

---

Subsampling1DLayer (Pooling1D)

Class: org.deeplearning4j.nn.conf.layers.Subsampling1DLayer

1D temporal pooling. Input shape: [minibatch, channels, sequenceLength]. Supports MAX, AVG, SUM, PNORM.

new Subsampling1DLayer.Builder(SubsamplingLayer.PoolingType.MAX)
    .kernelSize(2)
    .stride(2)
    .build()

---

GlobalPoolingLayer

See GlobalPoolingLayer in the Layers Reference for full documentation. In the CNN context:

// Replace flatten + dense with global average pooling
.layer(new GlobalPoolingLayer.Builder(PoolingType.AVG).build())
.layer(new OutputLayer.Builder(LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD)
    .nIn(numFilters).nOut(numClasses).activation(Activation.SOFTMAX).build())

---

Deconvolution2D (Transposed Convolution)

Class: org.deeplearning4j.nn.conf.layers.Deconvolution2D Source: Deconvolution2D.java

Transposed convolution (also called fractionally strided convolution or deconvolution). Swaps the forward and backward passes of a regular ConvolutionLayer, producing outputs that are typically larger than the input. Used in decoders, segmentation networks, and generative models.

Reference: Zeiler et al., "Deconvolutional Networks" (CVPR 2010).

Builder Parameters

Same as ConvolutionLayer — kernelSize, stride, padding, nIn, nOut, activation, convolutionMode, hasBias.

Example — Upsampling Decoder Block

// Encoder output: [mb, 256, 4, 4]
// After Deconvolution2D with kernel 4, stride 2: [mb, 128, 8, 8]
new Deconvolution2D.Builder(4, 4)
    .nIn(256).nOut(128)
    .stride(2, 2)
    .activation(Activation.RELU)
    .build()

---

DepthwiseConvolution2D

Class: org.deeplearning4j.nn.conf.layers.DepthwiseConvolution2D Source: DepthwiseConvolution2D.java

Applies a separate convolutional filter to each input channel independently (no cross-channel mixing). Used as the first step in MobileNet-style depthwise separable convolutions.

Builder Parameters

Parameter	Type	Description
nIn	int	Number of input channels
depthMultiplier	int	Number of output channels per input channel (nOut = nIn * depthMultiplier)
kernelSize(int, int)	int, int	Kernel size
stride(int, int)	int, int	Stride
convolutionMode	ConvolutionMode	Border handling

Example

// 64 input channels, depthMultiplier=1 -> 64 output channels
new DepthwiseConvolution2D.Builder()
    .nIn(64)
    .depthMultiplier(1)
    .kernelSize(3, 3)
    .stride(1, 1)
    .convolutionMode(ConvolutionMode.Same)
    .activation(Activation.RELU)
    .build()

---

SeparableConvolution2D

Class: org.deeplearning4j.nn.conf.layers.SeparableConvolution2D Source: SeparableConvolution2D.java

Combines a depthwise convolution with a 1x1 pointwise convolution for cross-channel mixing. Approximates a standard convolution with far fewer parameters — the MobileNet building block.

Builder Parameters

Parameter	Type	Description
nIn	int	Input channels
nOut	int	Output channels (after pointwise conv)
kernelSize(int, int)	int, int	Depthwise kernel size
depthMultiplier	int	Intermediate channels = nIn * depthMultiplier
stride(int, int)	int, int	Stride for depthwise conv
convolutionMode	ConvolutionMode	Border handling

Example

// MobileNet-style block
new SeparableConvolution2D.Builder(3, 3)
    .nIn(64).nOut(128)
    .depthMultiplier(1)
    .convolutionMode(ConvolutionMode.Same)
    .activation(Activation.RELU)
    .build()

---

Upsampling2D

Class: org.deeplearning4j.nn.conf.layers.Upsampling2D Source: Upsampling2D.java

Nearest-neighbor upsampling — repeats each value size times in both the height and width dimensions. No learnable parameters. Used in U-Net decoders and image segmentation.

Input [mb, C, H, W] with size=2 -> output [mb, C, 2H, 2W].

Builder Parameters

Parameter	Type	Description
size(int)	int	Upsampling factor (applied to both H and W)
size(int[])	int[2]	Separate height and width factors

Example

new Upsampling2D.Builder()
    .size(2)    // double spatial resolution
    .build()

// Non-uniform upsampling
new Upsampling2D.Builder()
    .size(new int[]{2, 3})   // 2x height, 3x width
    .build()

---

Upsampling1D

Class: org.deeplearning4j.nn.conf.layers.Upsampling1D

Repeats each value along the sequence dimension. Input [mb, C, L] -> output [mb, C, size*L].

new Upsampling1D.Builder().size(2).build()

---

Upsampling3D

Class: org.deeplearning4j.nn.conf.layers.Upsampling3D

Repeats values in depth, height, and width. Input [mb, C, D, H, W] -> output [mb, C, size*D, size*H, size*W].

new Upsampling3D.Builder().size(2).build()          // uniform
new Upsampling3D.Builder().size(new int[]{2,2,2}).build()

---

ZeroPaddingLayer (2D)

Class: org.deeplearning4j.nn.conf.layers.ZeroPaddingLayer Source: ZeroPaddingLayer.java

Adds zero-padding explicitly around 2D spatial inputs. Useful when you need precise control over padding at a specific point in a ComputationGraph, separate from a convolution layer's built-in padding.

// Pad 2 rows top, 2 rows bottom, 2 cols left, 2 cols right
new ZeroPaddingLayer.Builder(2, 2, 2, 2).build()

// Symmetric padding
new ZeroPaddingLayer.Builder(1, 1).build()   // 1 top/bottom, 1 left/right

---

ZeroPadding1DLayer

Adds zero-padding to the start and end of 1D sequences.

new ZeroPadding1DLayer.Builder(2, 2).build()   // 2 steps left, 2 steps right

---

ZeroPadding3DLayer

Adds zero-padding to volumetric data. Accepts a 6-element array: [padLeftD, padRightD, padLeftH, padRightH, padLeftW, padRightW].

new ZeroPadding3DLayer.Builder(1, 1, 1, 1, 1, 1).build()

---

Cropping Layers

Cropping1D

Removes n time steps from the start and/or end of sequences.

new Cropping1D.Builder(2, 2).build()   // remove 2 steps from each end

Cropping2D

Removes rows/columns from 2D spatial inputs.

new Cropping2D.Builder(1, 1, 1, 1).build()   // top, bottom, left, right

Cropping3D

new Cropping3D.Builder(1, 1, 1).build()       // depth, height, width (symmetric)
new Cropping3D.Builder(1, 1, 2, 2, 0, 0).build()  // per-side for D, H, W

---

SpaceToDepthLayer

Class: org.deeplearning4j.nn.conf.layers.SpaceToDepthLayer

Rearranges spatial blocks into the channel (depth) dimension. Downsamples spatial size while increasing channel count without any learned parameters. Used in certain detection architectures (YOLO passthrough layer).

Input [mb, C, H, W] with block size b -> output [mb, C * b * b, H/b, W/b].

new SpaceToDepthLayer.Builder()
    .blocks(2)    // 2x2 spatial blocks -> 4x channel expansion
    .dataFormat(SpaceToDepthLayer.DataFormat.NCHW)
    .build()

---

Yolo2OutputLayer

For object detection with the YOLOv2 architecture:

// boundingBoxPriors: shape [numBoxes, 2] with (w, h) as fraction of grid cell
INDArray priors = Nd4j.create(new double[][] {{1.3221, 1.73145}, {3.19275, 4.00944}});

new Yolo2OutputLayer.Builder()
    .boundingBoxPriors(priors)
    .build()

See the DL4J object detection examples for a full YOLO pipeline.

---

Complete CNN Architecture Example

A simple VGG-style classifier for CIFAR-10:

int h = 32, w = 32, c = 3, numClasses = 10;

MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
    .seed(42)
    .dataType(DataType.FLOAT)
    .updater(new Adam(1e-3))
    .weightInit(WeightInit.XAVIER)
    .l2(1e-4)
    .list()
    // Block 1
    .layer(new ConvolutionLayer.Builder(3, 3).nIn(c).nOut(64)
        .padding(1,1).activation(Activation.RELU).build())
    .layer(new ConvolutionLayer.Builder(3, 3).nOut(64)
        .padding(1,1).activation(Activation.RELU).build())
    .layer(new SubsamplingLayer.Builder(PoolingType.MAX)
        .kernelSize(2,2).stride(2,2).build())
    // Block 2
    .layer(new ConvolutionLayer.Builder(3, 3).nOut(128)
        .padding(1,1).activation(Activation.RELU).build())
    .layer(new ConvolutionLayer.Builder(3, 3).nOut(128)
        .padding(1,1).activation(Activation.RELU).build())
    .layer(new SubsamplingLayer.Builder(PoolingType.MAX)
        .kernelSize(2,2).stride(2,2).build())
    // Classifier head
    .layer(new DenseLayer.Builder().nOut(512).activation(Activation.RELU).build())
    .layer(new DropoutLayer.Builder(0.5).build())
    .layer(new OutputLayer.Builder(LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD)
        .nOut(numClasses).activation(Activation.SOFTMAX).build())
    .setInputType(InputType.convolutional(h, w, c))
    .build();