CNN Receptive Field Calculator

Calculated Results

Results appear here after calculation.

Not calculated

#	Layer	Type	Input	Channels	Kernel	Stride	Dilation	Padding	Output	Jump	Receptive field	Start	Parameters	MACs	Activation	Coverage
Configure the network and calculate.

Output Coordinate Analysis

X index

Y index

Z index

Calculate the network first.

Input Coverage Map

The highlighted rectangle shows one selected output cell projected onto the original input.

Network Configuration

Build sequential, residual, branched, one-dimensional, two-dimensional, or three-dimensional CNN paths.

CNN mode

Input width / length

Input height

Input depth / frames

Input channels

Activation data type

Output rounding

Decimal precision

Architecture preset

Import JSON

Architecture Comparison

Capture two configurations and compare final receptive field, output resolution, jump, depth, and coverage.

Metric	Architecture A	Architecture B
Capture two calculated architectures.

Target Receptive Field Helper

Target receptive field

Preferred kernel

Estimate how many stride-one layers are needed for a target field.

Formula Used

effectiveKernel = dilation × (kernel − 1) + 1

Dilation expands the kernel footprint without adding sampled weights.

jumpₗ = jumpₗ₋₁ × strideₗ

Jump describes spacing between neighbouring output centres on the original input.

receptiveFieldₗ = receptiveFieldₗ₋₁ + (effectiveKernel − 1) × jumpₗ₋₁

Each spatial layer enlarges the previous theoretical receptive field.

output = floor((input + padBefore + padAfter − effectiveKernel) / stride) + 1

Transposed convolution and upsampling use their own reverse-size rules.

How to Use

Select one-dimensional, two-dimensional, or three-dimensional mode.
Enter the original input dimensions.
Add layers and choose kernels, strides, dilation, padding, and optional graph connections.
Press calculate to inspect every intermediate output and receptive field.
Select an output coordinate to project its coverage onto the original input.
Export results or capture two architectures for comparison.

Example Data

Layer	Kernel	Stride	Dilation	Padding	Expected effect
Conv 3×3	3	1	1	Same	Adds two input pixels to receptive field.
Max pool	2	2	1	Valid	Doubles jump and reduces feature-map size.
Dilated conv	3	1	4	Same	Acts like an effective 9×9 kernel.
Upsample	1	2	1	Valid	Increases resolution while reducing effective jump.

Understanding the Results

Theoretical field

The maximum original-input region capable of influencing an output unit.

Effective field estimate

A simplified concentration estimate showing the central region likely to contribute most strongly.

Valid field

The portion remaining inside real input boundaries after padded coordinates are clipped.

Projected field

The selected output unit mapped back to coordinates on the original image, sequence, or volume.

Frequently Asked Questions

What is a CNN receptive field?

The receptive field is the input region influencing one output activation.

Does same padding stop receptive-field growth?

No. It preserves output size but still enlarges the theoretical field.

How does stride affect receptive fields?

Stride increases the jump between neighbouring output centres on the original input.

Why does dilation increase coverage?

Dilation spaces kernel samples apart and increases the effective kernel size.

Can the receptive field exceed the image?

Yes. Theoretical coverage may extend into padded coordinates beyond the image.

How are residual connections handled?

The calculator combines referenced paths and reports mismatched shapes as warnings.

Is transposed convolution exact here?

Output size is exact, while receptive-field growth is reported as a practical approximation.

What is effective jump?

It is the spacing between adjacent feature-map positions measured on the original input.

Why compare architectures?

Comparison reveals trade-offs between context, resolution, downsampling, and boundary dependence.