Cosine Annealing Scheduler Calculator

Calculator

Schedule type

Pick a smooth decay, or periodic restarts.

Max learning rate (η_max)

Starting peak rate for the schedule.

Min learning rate (η_min)

Lower bound reached near the cosine trough.

Current epoch (t)

Epoch where you want η(t).

Total epochs for classic (T_max)

Used only for the classic schedule.

Total epochs for restarts (T_total)

Cap: 5000 epochs for speed.

Initial cycle length (T₀)

First restart period length.

Cycle multiplier (T_mult)

Cycle length grows by this factor each restart.

Peak decay per restart (γ)

1 keeps peaks constant. Less than 1 reduces peaks.

Display decimals

Precision for rates in the table.

Rows to display

Downloads always include the full schedule.

Use case	η_max	η_min	Main setting	Note
Classic training run	0.10	0.0001	T_max = 100	Best for one continuous training horizon.
Exploration with restarts	0.05	0.0001	T_0 = 10, T_mult = 2	Periodic peaks can escape sharp minima.
Restart peaks decay	0.05	0.0001	γ = 0.9	Gradually reduces peak learning rate over cycles.

Use these as starting points, then validate against your optimizer and batch size.

Classic cosine annealing

η(t) = η_min + 0.5(η_max − η_min) · (1 + cos(π · t / T_max))

Warm restarts

η(t) = η_min + 0.5(η_max(cycle) − η_min) · (1 + cos(π · t_cur / T_i))

t_cur is the epoch index within the active cycle.
T_i is the active cycle length, starting at T_0 and multiplied by T_mult after each restart.
η_max(cycle) = η_max · γ^cycle applies optional peak decay per restart.

Choose Classic cosine annealing for one training horizon, or Warm restarts for periodic cycles.
Enter η_max and η_min that match your optimizer scale.
Set t to the epoch you want to inspect.
For classic mode, set T_max to total epochs.
For restarts, set T_total, T_0, and T_mult. Use γ if you want peaks to shrink.
Click Calculate. The results appear above the form, with a downloadable schedule.