Step Decay Learning Rate Calculator

Calculator inputs

Initial learning rate

Starting optimizer learning rate before any scheduled drops.

Decay factor

Multiplier applied after each completed step interval.

Step size

Number of epochs between scheduled learning rate drops.

Current epoch index

Use zero-based indexing. Example: the first epoch is 0.

Current batch

Partial progress inside the current epoch.

Batches per epoch

Used to convert batch progress into fractional epochs.

Total planned epochs

Defines the schedule horizon and the output table length.

Minimum learning rate floor

Prevents the scheduled rate from decaying below this floor.

Decimal places

Controls numeric display precision across results and exports.

Reset

Example data table

Scenario	Initial Rate	Decay Factor	Step Size	Current Epoch	Current Batch	Batches per Epoch	Minimum Rate
CNN image training	0.1000	0.50	10	23	40	100	0.0001
Transformer fine-tuning	0.0005	0.80	3	7	120	400	0.00005
Tabular model refresh	0.0200	0.30	5	11	12	60	0.0002

Formula used

Step decay learning rate:

LR(t) = max(LR_min, LR_0 × γ^{floor(t / s)})

Effective training position:

t = current_epoch + (current_batch / batches_per_epoch)

Where: LR_0 is the initial rate, γ is the decay factor, s is step size, and LR_min is the minimum allowed rate.

This schedule keeps the learning rate constant within each interval, then drops it by a fixed multiplier whenever the effective epoch crosses a step boundary.

How to use this calculator

Enter the initial learning rate that your optimizer starts with.
Set the decay factor, such as 0.5 for halving the rate.
Choose the step size in epochs between rate drops.
Provide the current epoch, current batch, and batches per epoch.
Add a minimum learning rate floor if you want bounded decay.
Enter the total epochs to generate the complete schedule table.
Click the calculate button to show the results above the form.
Use the CSV or PDF buttons to export the summary and schedule.

FAQs

1. What does step decay mean?

Step decay lowers the learning rate in sudden stages instead of continuously. The rate stays fixed for a chosen interval, then multiplies by the decay factor when that interval completes.

2. Why use a minimum learning rate floor?

A floor stops the schedule from becoming too small to produce meaningful parameter updates. It helps maintain learning progress late in training and reduces the risk of stagnation.

3. What is a good decay factor?

Common factors are 0.1, 0.5, or 0.8. Smaller values create sharper drops, while larger values decay more gently. The right choice depends on model sensitivity and validation performance.

4. How should I choose the step size?

Choose a step size that matches training stability and dataset complexity. Short steps adapt quickly, while longer steps keep learning aggressive for more epochs before reducing the rate.

5. Why does the calculator use current batch?

Current batch converts partial epoch progress into a fractional effective epoch. That helps you estimate the active learning rate more precisely when training is mid-epoch.

6. Is step decay better than cosine decay?

Neither is always better. Step decay is simple, interpretable, and easy to audit. Cosine decay is smoother. The best option depends on your model, optimizer, and tuning goals.

7. Why is my floor applied early?

An early floor usually means the decay factor is too aggressive, the step size is too short, or the minimum rate is set too high relative to the initial rate.

8. Can I use this for optimizer schedule planning?

Yes. The calculator helps compare current, next, and final rates, identify decay milestones, and export schedules for experiment tracking, reviews, or training documentation.