Neural Network Architecture Diagram Calculator

Architecture Summary

Ready to analyse the current architecture.

Not checked

Project Settings

Architecture name

Network category

Precision

Bytes per value

Batch size

Diagram layout

Detail level

Warning limit

Layer Builder

0 layers

Drag layers to reorder. Changes update the diagram automatically.

Diagram Styling

Block width

Block height

Layer spacing

Font size

Arrow thickness

Corner radius

Diagram background

Show tensor shapes

Show parameter counts

Show activations

Transparent export background

Interactive Architecture Diagram

Use the wheel or trackpad to scroll. Select a block to edit its layer.

#	Layer	Type	Input	Output	Activation	Parameters	Memory	Status

Architecture A

Architecture B

Metric	A	B	Difference

Architecture JSON

Export tools

Formula Used

Dense layer parameters

Parameters = (Input units × Output units) + Output units

Convolution output size

Output = floor(((Input + 2P − D(K−1) − 1) / S) + 1)

Convolution parameters

Parameters = (Kernel H × Kernel W × Input channels × Filters) + Filters

How to Use

Choose a preset or build an architecture from scratch.
Add layers and configure each layer’s dimensions and options.
Drag layer cards to change their execution order.
Review calculated shapes, parameters, memory, FLOPs, and warnings.
Adjust diagram styling and export the finished architecture.

Example Architecture Data

Layer	Configuration	Typical output	Purpose
Input	28 × 28 × 1	28 × 28 × 1	Receives a grayscale image.
Conv2D	32 filters, 3 × 3, same	28 × 28 × 32	Extracts local spatial features.
Max Pool	2 × 2, stride 2	14 × 14 × 32	Reduces spatial dimensions.
Dense	128 units, ReLU	128	Combines extracted features.
Output	10 units, Softmax	10	Returns class probabilities.

Frequently Asked Questions

What does the calculator produce?

It creates a visual architecture diagram and a numerical model summary. It also checks common dimension errors.

Are parameter counts exact?

Common dense, convolutional, recurrent, embedding, normalisation, and attention layers use standard formulas. Custom layers use an editable estimate.

How are FLOPs estimated?

The calculator applies practical operation estimates for supported layers. Framework profilers may report slightly different totals.

Can I create residual connections?

Yes. Set a skip-source index on a layer. The diagram displays a curved residual connection.

Can I compare architectures?

Capture the current model as Architecture A or B. The comparison table shows parameter, memory, and FLOP differences.

Does the calculator save projects?

Yes. Browser storage preserves the current architecture locally. JSON files provide portable backups.

Which precision settings are available?

FP64, FP32, FP16, BF16, INT8, and custom byte sizes are supported. Memory estimates update immediately.

Can diagrams be exported?

SVG, PNG, PDF, JSON, and CSV exports are included. Printing is also supported.

Does it require a database?

No database is required. The calculator runs in the browser after the PHP page loads.