cosine decay schedule

The learning rate is the single most consequential number in a training run: it sets how far each optimizer step moves the weights. Set it too high and the loss diverges; set it too low and training crawls or settles into a poor minimum. A *learning-rate schedule* is the recognition that no single value is right for the whole run — the ideal step size early in training, when the weights are random and gradients are large, is not the ideal step size late in training, when the model is fine-tuning its way into a minimum. The canonical modern recipe, warmup followed by cosine decay, encodes exactly this intuition.\n\n**Warmup starts the learning rate near zero and ramps it up over the first few percent of training.** This looks wasteful but is essential for large models, and for two reasons. At initialization the weights are random, so gradients are large and pointing in inconsistent directions; a full-size step here can knock the model into a bad region it never recovers from. And adaptive optimizers like Adam estimate a running variance of the gradients that is unreliable for the first few hundred steps, so their effective step size is erratic until those statistics settle. A linear warmup holds the step size small while both problems resolve, then hands off to the peak learning rate once training is on stable footing. Large-batch training makes warmup even more important.\n\n**Decay then walks the learning rate back down toward zero over the rest of training.** The logic is explore-then-settle: a high learning rate covers ground quickly and escapes shallow traps, but you cannot converge to a sharp minimum while taking large steps, so you gradually shrink the step size to let the model settle. *Cosine decay* is the dominant choice — it follows a smooth half-cosine from the peak down to near zero, spending a lot of the run at a moderately high rate and only slowing sharply at the very end. Its smoothness avoids the abrupt loss jumps that hard step-decay schedules can cause.\n\n**Warmup plus cosine decay is the default for essentially all large-model training.** You pick a peak learning rate, a warmup length (often 1-4% of total steps), and a total step budget the cosine decays across; that budget coupling is why you generally must know your total training length up front. Other schedules still have their places: the original Transformer used an inverse-square-root decay tied to warmup; step decay (cut the rate by a factor at fixed milestones) remains common in vision; and a constant rate with a short decay at the end is used when the total length is not known in advance. The through-line is always the same shape of idea — ramp up carefully, run hot, then cool down to converge.\n\n| Schedule | Shape | Needs total steps? | Typical home |\n|---|---|---|---|\n| Constant | Flat | No | Debugging, small jobs |\n| Step decay | Cut at milestones | No | Classic vision (ResNets) |\n| Inverse sqrt | 1/sqrt(step) after warmup | No | Original Transformer |\n| Warmup + linear | Ramp up, linear down | Yes | Fine-tuning (BERT-style) |\n| Warmup + cosine | Ramp up, cosine down | Yes | LLM pretraining (default) |\n\n```svg\n\n```\n\nIt is tempting to treat the learning rate as one number you sweep for and forget. The schedule reframes it as a story the training run tells over time: begin timidly because the model is fragile and the optimizer's own statistics are still forming, open up to a high rate once things are stable to make fast progress, then quiet down to converge cleanly. Read a schedule through an explore-then-settle lens rather than a set-and-forget lens, and warmup, cosine decay, and the coupling to your total step budget stop being ritual and become a direct expression of what the model needs at each phase of its training.