Data Augmentation Size Calculator

Dataset and Calculation Mode

Calculation mode

Dataset type

Augmentation delivery

Original samples

Training samples

Validation samples

Test samples

Generated sample rounding

Warning multiplier limit

Augment training data only

Retain original samples

Augmented copies per source sample

Target final dataset size

Desired dataset growth

Class-Level Augmentation

Balance target method

Custom shared class target

Class name	Current samples	Per-class target	Action

Class balancing adds samples to underrepresented classes. It never removes larger classes.

Transformation Combination Planner

Use	Transformation	Variations	Probability	Action
			%
			%
			%
			%
			%
			%

Theoretical combinations may include weak or visually similar samples.

Storage Estimation

Average original sample size

Average augmented sample size

Sample size unit

Compression reduction

Metadata overhead

Additional backup copies

Storage warning limit

Limit unit

Calculate uncompressed augmented image size from dimensions

Image width

Image height

Channels

Bits per channel

Online Augmentation and Training Workload

Batch size

Epochs

Steps per epoch

Use zero for automatic calculation.

Samples processed per second

Data workers

GPUs

Measured time per batch

Compute cost per GPU-hour

Formula Used

Fixed copies: Final size = Original size + Augmentation base × Copies per sample.

Target size: Required augmented samples = Target size − Retained baseline.

Percentage growth: Final size = Original size × (1 + Growth percentage ÷ 100).

Class balancing: Required class augmentation = max(0, Target class size − Current class size).

Image storage: Bytes = Width × Height × Channels × Bits per channel ÷ 8.

Training presentations: Batch size × Steps per epoch × Epoch count.

How to Use

Select the calculation mode and dataset type.
Enter dataset counts and choose augmentation scope.
Add class counts when balancing labels.
Configure transformations and variation counts.
Estimate storage, training speed, and compute cost.
Submit the form and review every warning.
Export results as CSV or PDF.

Example Data

Input	Example value	Purpose
Original samples	10,000	Current complete dataset size
Training samples	8,000	Samples eligible for augmentation
Copies per sample	3	Creates 24,000 augmented training samples
Average sample size	250 KB	Estimates physical storage
Batch size	32	Calculates steps and workload
Epochs	20	Calculates total presentations

Practical Guidance

Split the dataset before applying augmentation. Keep validation and test samples unchanged. This reduces leakage across evaluation sets.

Online augmentation creates new batches during training. It usually requires little additional storage. Repeated presentations are not always unique.

Class balancing can improve minority representation. Excessive synthetic duplication may still cause overfitting. Review class quality after augmentation.

Transformation combinations are theoretical upper bounds. Some combinations may be invalid or redundant. Visual inspection remains an important quality check.

Frequently Asked Questions

What does the augmentation multiplier mean?

It compares final dataset size with original size. A multiplier of four means four times more effective samples.

Does online augmentation increase stored dataset size?

Usually it does not create permanent sample files. Storage stays near the original dataset size.

Should validation data be augmented?

Validation data normally remains unchanged. Stable evaluation data gives more trustworthy comparisons.

Can this calculator balance multiple classes?

Yes. Add each class and choose a shared or separate target.

What is a transformation combination?

It is one possible mixture of selected transformation variations. Similar combinations may produce nearly identical samples.

How is storage estimated?

The calculator multiplies sample counts by average file sizes. Compression, metadata, and backups are then applied.

Why can generated samples exceed unique combinations?

Random parameters can create many variations inside one setting. The warning uses a simple discrete combination model.

Does more augmentation always improve accuracy?

No. Poor transformations may distort labels or important features. Test each policy against validation performance.

Can I estimate GPU training cost?

Yes. Enter GPU count, throughput, and hourly cost. The estimate remains approximate.

Which datasets can use this calculator?

It supports image, text, audio, tabular, and multimodal planning. Transformation presets can be customised.