The Warmup Dilemma: How Learning Rate Strategies Impact Speech-to-Text Model Convergence

TL;DR

This paper investigates how different learning rate warmup strategies affect the convergence and final performance of large-scale speech-to-text models, revealing that sub-exponential warmup is optimal and higher warmup LR accelerates initial training but doesn't improve ultimate results.

Contribution

It systematically compares various learning rate warmup schedules for speech-to-text models, highlighting the effectiveness of sub-exponential warmup and clarifying the impact of warmup LR on convergence.

Findings

Large-scale S2T training requires sub-exponential LR warmup.

Higher warmup LR speeds up initial convergence.

Warmup LR does not improve final model performance.

Abstract

Training large-scale models presents challenges not only in terms of resource requirements but also in terms of their convergence. For this reason, the learning rate (LR) is often decreased when the size of a model is increased. Such a simple solution is not enough in the case of speech-to-text (S2T) trainings, where evolved and more complex variants of the Transformer architecture -- e.g., Conformer or Branchformer -- are used in light of their better performance. As a workaround, OWSM designed a double linear warmup of the LR, increasing it to a very small value in the first phase before updating it to a higher value in the second phase. While this solution worked well in practice, it was not compared with alternative solutions, nor was the impact on the final performance of different LR warmup schedules studied. This paper fills this gap, revealing that i) large-scale S2T trainings demand a sub-exponential LR warmup, and ii) a higher LR in the warmup phase accelerates initial convergence, but it does not boost final performance.