Feature Map Size Calculator for CNN and Pooling Layers

Plan CNN layers with fast reliable size estimates. Test kernels, strides, padding, and dilation quickly. Build cleaner architectures using dependable dimensional calculations every time.

Feature Map Size Calculator

Use separate controls for width, height, stride, dilation, padding, and output channels. This layout stays single column overall, while the form fields switch between 3, 2, and 1 columns responsively.

Example Data Table

These examples help compare common CNN and feature map patterns.

Case Operation Input Kernel Stride Padding Dilation Output
Classic same-size conv Convolution 224 × 224 × 3 3 × 3 1 × 1 1 all sides 1 × 1 224 × 224 × 64
Downsampling conv Convolution 128 × 128 × 64 5 × 5 2 × 2 2 all sides 1 × 1 64 × 64 × 128
Max pooling style Pooling 56 × 56 × 128 2 × 2 2 × 2 0 all sides 1 × 1 28 × 28 × 128
Transposed upsampling Transposed Convolution 28 × 28 × 128 4 × 4 2 × 2 1 all sides 1 × 1 56 × 56 × 64

Formula Used

Convolution or Pooling
Output Width = floor or chosen rounding of ((Input Width + Left Padding + Right Padding - Effective Kernel Width) / Stride X) + 1
Output Height = floor or chosen rounding of ((Input Height + Top Padding + Bottom Padding - Effective Kernel Height) / Stride Y) + 1
Effective Kernel
Effective Kernel Width = ((Kernel Width - 1) × Dilation X) + 1
Effective Kernel Height = ((Kernel Height - 1) × Dilation Y) + 1
Transposed Convolution
Output Width = ((Input Width - 1) × Stride X) - Left Padding - Right Padding + Effective Kernel Width + Output Padding X
Output Height = ((Input Height - 1) × Stride Y) - Top Padding - Bottom Padding + Effective Kernel Height + Output Padding Y
Parameter Estimate
Parameters = (Kernel Height × Kernel Width × Input Channels × Output Channels) + Optional Biases

These formulas are commonly used for CNN design, pooling analysis, feature extractor planning, encoder-decoder layouts, and dimensional debugging across AI workflows.

How to Use This Calculator

  1. Choose the layer operation: convolution, pooling, or transposed convolution.
  2. Enter input width, input height, input channels, and batch size.
  3. Set kernel size, stride, dilation, and separate padding values.
  4. Enter filters for convolution style layers, or keep pooling channels unchanged.
  5. Use output padding only for transposed convolution layers.
  6. Select a rounding mode if you want to test floor, ceil, or round behavior.
  7. Set how many repeated layers you want in the Plotly projection graph.
  8. Click the calculate button to view output dimensions, tensor shape, estimates, chart, and export options.

FAQs

1) What does feature map size mean in deep learning?

Feature map size describes the height, width, and channels produced by a layer. It helps you track spatial resolution and tensor growth across a model.

2) Why do stride values reduce output size?

A larger stride moves the kernel farther each step. Fewer placements fit across the input, so the resulting feature map becomes smaller.

3) What does dilation change in convolution?

Dilation expands the effective kernel without adding more kernel weights. It increases the receptive coverage while preserving the original parameter count pattern.

4) Why can padding preserve spatial dimensions?

Padding adds border values around the input. With a suitable kernel and stride, it offsets shrinkage and can keep output size equal to input size.

5) Does pooling change the number of channels?

Standard pooling usually changes height and width only. The channel count normally stays the same unless another operation changes depth afterward.

6) What is output padding in transposed convolution?

Output padding helps fine-tune final upsampled dimensions in transposed convolution. It does not add learned weights, but it changes the resulting spatial size.

7) Why estimate parameters and MACs here?

Parameter and MAC estimates help compare memory cost and compute demand. They are useful while planning CNN architectures before training or deployment.

8) Can I use this calculator for rectangular inputs?

Yes. Width and height are entered separately, so you can test non-square images, kernels, stride choices, and asymmetric padding combinations easily.

Related Calculators

batch size calculatorstride calculatorimage size calculatorintersection over unionimage resolution calculatorpixel density calculatorbounding box areaimage resize scaleanchor box sizereceptive field calculator

Important Note: All the Calculators listed in this site are for educational purpose only and we do not guarentee the accuracy of results. Please do consult with other sources as well.