Calculator Inputs
Use the stacked page layout below. The layer input grid shifts to three columns on large screens, two on medium screens, and one on mobile.
Example Data Table
This sample stack mirrors a common vision pipeline and helps verify the calculator quickly.
| Layer | Type | Kernel | Stride | Pad Left | Pad Right | Dilation | Expected Receptive Field After Layer |
|---|---|---|---|---|---|---|---|
| Conv 1 | Convolution | 3 | 1 | 1 | 1 | 1 | 3 |
| Conv 2 | Convolution | 3 | 2 | 1 | 1 | 1 | 5 |
| Pool 1 | Pooling | 2 | 2 | 0 | 0 | 1 | 7 |
| Conv 3 | Convolution | 3 | 1 | 2 | 2 | 2 | 23 |
Formula Used
For each layer, the calculator tracks three core values: effective kernel size, jump, and receptive field. The recursion begins with input values r0 = 1, j0 = 1, and start0 = 0.5.
Effective kernel: k_eff = dilation × (kernel − 1) + 1
Output size: out = floor((in + pad_left + pad_right − k_eff) / stride) + 1
Jump update: j_out = j_in × stride
Receptive field update: r_out = r_in + (k_eff − 1) × j_in
Center start update: start_out = start_in + (((k_eff − 1) / 2) − pad_left) × j_in
The final output center for index n is center = start + n × jump. Its covered input interval becomes [center − (r − 1)/2, center + (r − 1)/2].
How to Use This Calculator
- Enter the input image size for one spatial dimension.
- Choose the output index you want to inspect in the final feature map.
- Add one row per layer and select convolution or pooling.
- Fill in kernel size, stride, left padding, right padding, and dilation.
- Submit the form to display the result above the input form.
- Review the table, summary cards, and Plotly graph.
- Download the layer table as CSV or export the result block as PDF.
8 FAQs
1) What does receptive field mean in vision models?
It is the region of the original input that can influence one output activation. Larger receptive fields capture broader context, while smaller ones focus on local detail.
2) Why does stride affect the final jump?
Stride increases the spacing between neighboring output centers. As stride grows across layers, the sampling step in input coordinates becomes larger, which raises the jump value.
3) How does dilation change receptive field size?
Dilation expands the effective kernel without increasing parameter count proportionally. It lets each layer cover wider input context while keeping the kernel definition compact.
4) Why are left and right padding entered separately?
Asymmetric padding shifts the start location of output centers. Separate values help you inspect architectures where padding differs between both sides.
5) Does pooling use different receptive field recursion?
For spatial growth, pooling follows the same recursion pattern as convolution. Kernel size, stride, and padding still control how the field expands layer by layer.
6) Why is the start value initialized at 0.5?
That convention places the first input pixel center at coordinate 0.5. It makes center tracking consistent and keeps interval boundaries easy to interpret.
7) What does coverage percentage represent?
Coverage percentage compares the current receptive field width with the original input size. It helps estimate how much of the input one activation can theoretically observe.
8) Can this calculator validate full model semantics?
No. It tracks spatial geometry only. Nonlinearities, channel mixing, normalization, and learned weights still affect how much useful information each unit actually uses.