Scheduler function while fine-tuning ASR

[evilc3]

While fine-tuning ctc asr model the lr used is in range of 0.025 to 0.001 (according to examples In notebooks). But the scheduler further reduces the lr.

when I set lr = 0.025 the max_lr I see after warmup is 1 e -6, its not 0.025

so should we only use scheduler while training from scratch and not fine-tuning ?