Receptive Field Calculator
The receptive field of a convolutional network — how much of the input one output pixel sees — from the kernel sizes and strides of each layer, with the effective stride (jump) and the output size for a given input.
Each layer widens the receptive field by (kernel − 1) times the product of all earlier strides, because a stride of 2 means each of this layer’s pixels is two input pixels apart.
How the receptive field calculator works
Each layer widens the receptive field by (kernel − 1) times the product of all earlier strides, because a stride of 2 means each of this layer’s pixels is two input pixels apart. Stack five 3×3 layers with two stride-2 layers among them and each output pixel sees a 25-pixel-wide patch. The jump is the input distance between neighbouring output pixels; the output size follows from the input size and the strides with same-padding.
Formula: RF_l = RF_{l−1} + (k_l − 1) × jump_{l−1}; jump_l = jump_{l−1} × s_l
Worked examples
| Inputs | Receptive field | Note |
|---|---|---|
| ResNet-style stem: 7/2, 3/2, 3/1, 3/2, 3/1 | 43 × 43 pixels | 43 pixels, jump 8, 28 × 28 out |
| Five 3×3 layers, stride 1 | 11 × 11 pixels | 11 pixels |
| VGG block: two 3×3 then 2×2 pool, twice | 16 × 16 pixels | 16 pixels, jump 4 |
FAQFrequently asked questions
Why does the receptive field matter?
A pixel cannot be classified using context it cannot see. Detecting large objects or global structure needs a receptive field comparable to the object; too small and the network is guessing from texture.
Does padding change it?
No — padding changes the output size and alignment, not how far each output looks. The formula counts the theoretical field; the effective field, where most of the influence lies, is smaller and roughly Gaussian.
What about dilated convolutions and pooling?
A dilated kernel of size k and dilation d acts like size d(k − 1) + 1; a pooling layer is a kernel of its window size with its stride. Enter those equivalents.
How do residual and transformer blocks fit in?
Residual connections do not change the field of the convolutions inside them. A self-attention layer sees the whole input at once — its receptive field is global, which is one reason vision transformers differ.
Where these figures come from
- Vaswani et al. (2017) — Attention Is All You Need — the transformer architecture the memory arithmetic follows
- Kaplan et al. (2020) — Scaling Laws for Neural Language Models — the compute relationship used for training estimates
- Hoffmann et al. (2022) — Training Compute-Optimal Large Language Models — the tokens-per-parameter guidance ("Chinchilla")
- IEEE 754 — Standard for Floating-Point Arithmetic — the numeric formats behind bytes per parameter
- NIST AI Risk Management Framework — the US federal AI framework
Last checked: September 2026. The relationships here are architectural, not vendor-specific: bytes per parameter follow the numeric format, KV-cache size follows the transformer definition, and token-per-word ratios come from published tokeniser behaviour.