Training calculation results
Results update after each valid calculation.
Step-by-step calculation
Reverse calculations
Batch-size comparison
Training workload profile
Configuration comparison table
Training configuration
Enter dataset, batching, timing, and distributed settings.
Formula used
How to use this calculator
Enter the dataset size, batch size, and planned epochs. Choose whether the final incomplete batch remains. Then add distributed, timing, and scheduler settings.
Press Calculate to generate training and evaluation workload. Review warnings before using the estimates. Export the configuration when you need repeatable experiments.
- Set the dataset and train-validation-test percentages.
- Enter per-device batching and distributed device counts.
- Add accumulation, timing, evaluation, and checkpoint settings.
- Use reverse inputs for alternative planning questions.
- Compare batch sizes before selecting a final configuration.
Example configurations
| Use case | Dataset | Batch | Epochs | Devices | Accumulation |
|---|---|---|---|---|---|
| Image classification | 50,000 | 64 | 20 | 1 | 1 |
| Transformer fine-tuning | 120,000 | 8 | 3 | 2 | 4 |
| Object detection | 118,000 | 16 | 100 | 4 | 2 |
| Distributed training | 1,280,000 | 32 | 30 | 8 | 2 |
Saved calculation history
History remains in this browser until cleared.
Frequently asked questions
What is an epoch?
An epoch is one complete pass through the training dataset.
What is an iteration?
An iteration processes one batch and usually performs one forward-backward cycle.
How are epochs and iterations related?
Total iterations equal iterations per epoch multiplied by planned epochs.
Why does ceiling division appear?
Ceiling division includes the final smaller batch when samples remain.
What happens when drop-last is enabled?
The incomplete batch is ignored, reducing processed samples each epoch.
What is effective batch size?
It combines per-device batching, device count, and gradient accumulation.
Do multiple GPUs reduce iterations?
They reduce iterations when the global batch increases proportionally.
What are optimizer update steps?
They count parameter updates after accumulated gradients are applied.
Can validation use another batch size?
Yes. Validation and testing workloads are calculated separately.
How is training time estimated?
The estimate combines iteration, validation, checkpoint, and loading time.
What is warm-up progress?
It shows the early training period before the full learning rate.
Why compare batch sizes?
Batch changes affect memory, iteration count, updates, and estimated duration.
Can fractional epochs be used?
Yes. Fractional epochs are useful for partial dataset passes.
Does more training always improve results?
No. Excessive epochs can increase overfitting and wasted computation.
Are time estimates exact?
No. Hardware, data loading, caching, and evaluation behaviour vary.