How to use the CNN Output Size Calculator
- 1. Set the inputEnter the image height, width and channels.
- 2. List the layersOne per line, like conv 32 3 p=same, pool 2, flatten, dense 10.
- 3. CalculateSee each layer's output shape and parameters, and the total.
The formulas
For a convolution or pooling layer, each spatial size becomes ⌊(input + 2·padding − dilation·(kernel − 1) − 1) ÷ stride⌋ + 1, the formula PyTorch uses. p=same follows Keras: output = ⌈input ÷ stride⌉. A convolution has (kernel × kernel × input channels + 1) × filters parameters (weights plus one bias per filter). A dense layer has (inputs + 1) × units. Batch norm has 4 values per channel, 2 of them trainable. Pooling, flatten and dropout have none.
Worked example
A small CIFAR-10-style network on 32 × 32 × 3 images: conv 32 3 p=same keeps 32 × 32 and adds 896 parameters (3·3·3 + 1 = 28, × 32). Pool 2 halves it to 16 × 16. Conv 64 3 p=same adds 18,496 parameters, and pool 2 gives 8 × 8 × 64. Flatten makes 4,096 values, dense 128 adds 5,24,416 parameters and dense 10 adds 1,290. The total is 5,45,098 parameters, about 2.08 MB as 32-bit floats. Most of them sit in the first dense layer, which is why many modern CNNs use global average pooling (gap) instead of flatten.
Checks
- Shapes and counts were checked against an independent implementation of the same formulas on 80 random networks, including stride, dilation and padding.
- The tool flags layers whose output would shrink below 1 × 1.
- Grouped and depthwise convolutions, transposed convolutions and non-square kernels aren't supported.
Frequently asked questions
Why is my output size a fraction in my notes?
Frameworks round down (floor) for valid or explicit padding. If the numbers don't divide evenly, the last rows or columns are simply not covered.
Does 'same' padding work with stride 2?
Yes. As in Keras, the output is the input divided by the stride, rounded up.
Is the CNN output size calculator free?
Yes. You can open the CNN output size calculator and enter everything without signing up. When you calculate shapes for the first time, we ask once for your name and WhatsApp number; after that it, and every other Uncodemy tool, opens straight away.