Calculator Inputs
Example Data Table
| Scenario | Dimensions | Channels | Bit Depth | Raw Size | Stored % | Stored Size |
|---|---|---|---|---|---|---|
| Classification thumbnails | 512 × 512 | 3 | 8-bit | 768.00 KiB | 30% | 230.40 KiB |
| Medical grayscale scans | 1024 × 1024 | 1 | 16-bit | 2.00 MiB | 50% | 1.00 MiB |
| Remote sensing multispectral set | 2048 × 2048 | 13 | 16-bit | 104.00 MiB | 70% | 72.80 MiB |
Formula Used
Width × Height × Channels × Bit Depth ÷ 8
Raw Bytes × (Stored Size Percentage ÷ 100)
Dataset Images × (1 + Augmentation Factor)
Stored Bytes per Image × Effective Dataset Images
Stored Dataset Bytes × (1 + Metadata Overhead ÷ 100)
Width : Height reduced by their greatest common divisor
This model helps estimate image memory usage for training pipelines, dataset versioning, augmentation planning, storage forecasting, and deployment packaging.
How to Use This Calculator
- Choose a preset or enter custom image characteristics.
- Enter width, height, channels, and bit depth.
- Set the stored size percentage to reflect compression or file format savings.
- Enter batch size, dataset count, augmentation factor, and overhead.
- Set train, validation, and test percentages so they total 100.
- Click the calculate button to view single image, batch, and full dataset estimates.
- Use the CSV and PDF buttons to export the calculated report.
Frequently Asked Questions
1. What does stored size percentage mean?
It represents how much space the saved image uses compared with raw memory. A value of 35 means the file is estimated at 35% of raw size.
2. Why do channels matter so much?
Each channel stores separate pixel information. RGB usually uses three channels, while multispectral and scientific images can use many more, increasing memory sharply.
3. Does this calculator estimate GPU memory?
It estimates image storage and dataset size, not full GPU usage. Training memory also depends on tensors, model weights, activations, precision, and framework overhead.
4. What is augmentation factor?
It multiplies the effective image count. A factor of 1 means one extra generated sample per original image, doubling the dataset count for planning.
5. Can I use decimal or unusual bit depths?
Yes. The calculator accepts values like 10, 12, or 16 bits per channel. This helps when estimating RAW, HDR, or sensor-driven machine vision data.
6. Why add metadata overhead?
Real datasets often include labels, masks, indexes, manifests, caches, previews, and directory structures. Overhead helps create more realistic storage budgets.
7. Is raw size the same as file size?
No. Raw size is the uncompressed in-memory footprint derived from pixels and channels. Actual file size depends on encoding, compression, and container format.
8. Why are train, validation, and test splits included?
They help estimate storage allocation across pipeline stages. This is useful when planning separate folders, object storage tiers, or distributed training workflows.