deeplearning4j/deeplearning4j · error · IllegalArgumentException
Kernel argument should be a [dim]d array, got [kernelSize]
Error message
Kernel argument should be a [dim]d array, got [kernelSize]
What it means
Thrown by the ConvolutionLayer constructor (deeplearning4j-nn, ConvolutionLayer.java:130) as an IllegalArgumentException when the kernelSize array's length does not match the layer's convolution dimension (2 for Convolution2D, 3 for Convolution3D). DL4J validates all shape-configuration arrays up front so misconfiguration fails at network construction rather than during training. The message shows the expected dimensionality and the actual array contents via Arrays.toString.
Solutions
- Count the values passed to kernelSize(...) and make the length match the layer's convolution dimension: 2 for 2D convolution (e.g. kernelSize(3,3)), 3 for 3D (e.g. kernelSize(3,3,3)).
- If you build configs programmatically, check kernelSize.length against the intended dim before constructing the layer.
- Use the single-value convenience form (e.g. kernelSize(3)) which DL4J expands to a square kernel of the right dimensionality.
Example fix
// before
new ConvolutionLayer.Builder(1, 2, 3) // 3 values on a 2D layer
.nIn(1).nOut(32).build();
// after
new ConvolutionLayer.Builder(3, 3) // 2 values for 2D convolution
.nIn(1).nOut(32).build(); Defensive patterns
Strategy: validation
Validate before calling
long[] kernelSize = ...;
int convDim = 2; // 3 for Convolution3D
if (kernelSize.length != convDim)
throw new IllegalArgumentException("kernelSize must have " + convDim + " values, got " + kernelSize.length); Try / catch
try {
layer = new ConvolutionLayer.Builder(kernelSize).nOut(32).build();
} catch (IllegalArgumentException e) {
if (e.getMessage().contains("Kernel argument")) {
kernelSize = Arrays.copyOf(kernelSize, convDim);
layer = new ConvolutionLayer.Builder(kernelSize).nOut(32).build();
} else throw e;
} Prevention
- Match kernelSize/stride/padding/dilation lengths to the convolution dimension (2 or 3).
- Use the scalar convenience forms (kernelSize(3)) that expand per dimension.
- Unit-test network construction so config errors surface at build time, not fit time.
When it happens
Trigger: Calling the ConvolutionLayer.Builder with kernelSize(...) whose array length differs from the dimension implied by the builder variant, e.g. new ConvolutionLayer.Builder(1,2,3) (3 values) on a 2D convolution where dim==2, or a 2-element kernelSize on a 3D layer.
Common situations: Copy-pasting config between Convolution2D and Convolution3D; specifying per-dimension values like {height,width,depth} for a 2D layer; programmatic generation of layer configs with variable-length arrays; upgrading code where the old Builder(int...) convenience constructor expected a different arity.
Understand the failure class
Background: "Must be a positive integer", "Invalid value", "Unsupported": the invalid-argument-value error family, when a library rejects the value you pass — this error's family across 35 libraries.
Related errors
- Dilation argument should be a [dim]d array, got [dilation]
- Padding argument should be a [dim]d array, got [padding]
- Strides argument should be a [dim]d array, got [stride]
- Invalid input for Convolution layer (layer…
- Kernel size has to be either two or three, got
AI-assisted analysis of deeplearning4j/deeplearning4j@4c22ac5fe4 (2026-09-07).
Data as JSON: /api/errors/f846552c4cbac3d3.
Report an issue: GitHub.
Appendix: source
Thrown at deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/ConvolutionLayer.java:130
protected AlgoMode cudnnAlgoMode = AlgoMode.PREFER_FASTEST;
protected FwdAlgo cudnnFwdAlgo;
protected BwdFilterAlgo cudnnBwdFilterAlgo;
protected BwdDataAlgo cudnnBwdDataAlgo;
/**
* ConvolutionLayer nIn in the input layer is the number of channels nOut is the number of filters to be used in the
* net or in other words the channels The builder specifies the filter/kernel size, the stride and padding The
* pooling layer takes the kernel size
*/
protected ConvolutionLayer(BaseConvBuilder<?> builder) {
super(builder);
long dim = builder.convolutionDim;
this.hasBias = builder.hasBias;
this.convolutionMode = builder.convolutionMode;
this.dilation = builder.dilation;
if (builder.kernelSize.length != dim) {
throw new IllegalArgumentException("Kernel argument should be a " + dim + "d array, got " + Arrays.toString(builder.kernelSize));
}
this.kernelSize = builder.kernelSize;
if (builder.stride.length != dim) {
throw new IllegalArgumentException("Strides argument should be a " + dim + "d array, got " + Arrays.toString(builder.stride));
}
this.stride = builder.stride;
if (builder.padding.length != dim) {
throw new IllegalArgumentException("Padding argument should be a " + dim + "d array, got " + Arrays.toString(builder.padding));
}
this.padding = builder.padding;
if (builder.dilation.length != dim) {
throw new IllegalArgumentException("Dilation argument should be a " + dim + "d array, got " + Arrays.toString(builder.dilation));
}
this.dilation = builder.dilation;
this.cudnnAlgoMode = builder.cudnnAlgoMode;
this.cudnnFwdAlgo = builder.cudnnFwdAlgo;
this.cudnnBwdFilterAlgo = builder.cudnnBwdFilterAlgo;
this.cudnnBwdDataAlgo = builder.cudnnBwdDataAlgo;View on GitHub (pinned to 4c22ac5fe4)