How to use the Neural Network Parameter Count Calculator
- 1. Pick the network typeChoose a dense MLP or a CNN.
- 2. Describe the layersList layer sizes, or write CNN layers one per line.
- 3. CountSee each layer's output shape, formula and parameter count.
Formulas
- Dense: inputs × units + units (one bias per unit).
- Conv2D: (kernel height × kernel width × input channels + 1) × filters. Image size does not change the count.
- Pooling and flatten: no parameters; they only change the shape.
Worked example (default MLP 784 → 128 → 64 → 10, like an MNIST classifier): 784 × 128 + 128 = 100,480, then 128 × 64 + 64 = 8,256, then 64 × 10 + 10 = 650. The total is 109,386 parameters, which take about 427 KiB at 4 bytes each.
The CNN example
Switch to CNN mode for the classic small MNIST convnet: input 28×28×1, conv 32 filters 3×3 → 26×26×32 with (9 × 1 + 1) × 32 = 320 parameters; pool 2 → 13×13×32; conv 64 3×3 → 11×11×64 with (9 × 32 + 1) × 64 = 18,496; pool 2 → 5×5×64; flatten → 1,600; dense 10 → 16,010. Total: 34,826, about a third of the dense network, because convolution shares weights across the image.
Convolutions use valid padding unless you add the word same. Totals match Keras model.summary() for these layer types. Batch normalisation, attention, embeddings and recurrent layers are not covered. To estimate memory for large language models, see the LLM GPU VRAM calculator; to see a filter at work, try the convolution visualizer.
Frequently asked questions
Why do biases matter?
Each unit or filter has one bias. Turning them off removes that many parameters, as some layers before batch norm do.
Does padding change the parameter count?
No. Padding changes output size, which affects later dense layers, but not the conv layer's own parameters.
Is the Neural Network Parameter Count Calculator free?
Yes, the Neural Network Parameter Count Calculator is free. The first time you see your first count, you are asked for your name and WhatsApp number once. After that, this and every other Uncodemy tool opens straight away without the form.