Open Models Almanac
Methodsoptimizationlearning-rate schedule

single mention · optimization

Linear warmup followed by step decay

After a warmup period, the learning rate is reduced in discrete steps, in the evidence by a fixed factor at regular intervals.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See learning-rate schedule for the methods filed beside it.

Also called a composite learning rate schedule.

Filed under optimization › learning-rate schedule