RAdam

On the Variance of the Adaptive Learning Rate and Beyond

We are in an early-release beta. Expect some adventures and rough edges.

Introduction
Motivation
Questions and Discussions
Quick Start Guide
Related Posts and Repos
Citation

Introduction

If warmup is the answer, what is the question?

The learning rate warmup for Adam is a must-have trick for stable training in certain situations. But the underlying mechanism is largely unknown. In our study, we suggest one fundamental cause is the large variance of the adaptive learning rates, and provide both theoretical and empirical support evidence.

In addition to explaining why we should use warmup, we also purpose RAdam, a theoretically sound variant of Adam.

Motivation

As in Figure 8, we assume gradients follows a normal distribution (mean: \mu, variance: 1). The variance of the adaptive learning rate is simulated and plotted in Figure 8 (blue curve). We can see that the adaptive learning rate has a large variance in the early stage of training.

When using the Transformer for NMT, a warmup stage is usually required to avoid convergence problems (Adam-vanilla converges around 500 PPL in Figure 1, while Adam-warmup successfully converges under 10 PPL). In further explorations, we notice that, if we use additional 2000 samples to estimate the adaptive learning rate, the convergence problems are avoided (Adam-2k); or, if we increase the value of eps, the convergence problems are also relieved (Adam-eps).

Therefore, we conjecture that the large variance in the early stage causes the convergence problem, and further propose Rectified Adam by analytically reducing the large variance. More details can be found in our paper.

Questions and Discussions

Do I need to tune learning rate?

Yes, the robustness of RAdam is not infinity. In our experiments, it works for a broader range of learning rates, but not all learning rates.

Notes on Transformer

Choice of the Original Transformer

We choose the original Transformer as our main study object because, without warmup, it suffers from the most serious convergence problems in our experiments. With such serious problems, our controlled experiments can better verify our hypothesis (i.e., we demonstrate that Adam-2k / Adam-eps can avoid spurious local optima by minimal changes).

Sensitivity

We observe the Transformer is sensitive to the architecture configuration, despite its efficiency and effectiveness. For example, by changing the position of the layer norm, the model may / may not require the warmup to get a good performance. Intuitively, since the gradient of the attention layer could be more sparse and the adaptive learning rates for smaller gradients have a larger variance, they are more sensitive. Nevertheless, we believe this problem deserves more in-depth analysis and is beyond the scope of our study.

Why warmup has a bigger impact on some model than others?

Although the adaptive learning rate has a larger variance in the early stage, the exact magnitude is subject to the model design. Thus, the convergent problem could be more serious for some models/tasks than others. In our experiments, we observe RAdam achieves consistent improvements over the vanilla Adam. It verifies the variance issue widely exists (since we can get better performance by fixing it).

Why SGD needs warmup?

To the best of our knowledge, the warmup heuristic is originally designed for large minibatch SGD [0], based on the intuition that the network changes rapidly in the early stage. However, we find that it does not explain why Adam requires warmup. Notice that, Adam-2k uses the same large learning rate but with a better estimation of the adaptive learning rate can also avoid the convergence problems.

The reason why sometimes warmup also helps SGD still lacks of theoretical support. FYI, when optimizing a simple 2-layer CNN with gradient descent, the thoery of [1] could be used to show the benifits of warmup. Specifically, the lr must be $O(cos \phi)$, where $\phi$ is the angle between the current weight and the ground true weight and $cos \phi$ could be very small due to high dimensional space and random initialization. And thus lr must be very small at the beginning to guarentee the convergence. $cos \phi$ however can be improved in the later stage, and thus the learning rate is allowed to be larger. Their theory somehow can justify why warmup is needed by gradient descend and neural networks. But it is still far-fetched for the real scenario.

[0] Goyal et al, Accurate, Large Minibatch SGD: Training Imagenet in 1 Hour, 2017
[1] Du et al, Gradient Descent Learns One-hidden-layer CNN: Don’t be Afraid of Spurious Local Minima, 2017

Quick Start Guide

Directly replace the vanilla Adam with RAdam without changing any settings.
Further tune hyper-parameters (including the learning rate) for a better performance.

Note that in our paper, our major contribution is to identify why we need the warmup for Adam. Although some researchers successfully improve their model performance (user comments), directly plugging in RAdam may not result in an immediate performance boost. Based on our experience, replacing the vanilla Adam with RAdam usually results in a better performance; however, if warmup has already been employed and tuned in the baseline method, it is necessary to tune hyper-parameters for RAdam.

Citation

Please cite the following paper if you found our model useful. Thanks!

Liu, Liyuan, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. "On the Variance of the Adaptive Learning Rate and Beyond." arXiv preprint arXiv:1908.03265 (2019).

@article{liu2019radam,
  title={On the Variance of the Adaptive Learning Rate and Beyond},
  author={Liu, Liyuan and Jiang, Haoming and He, Pengcheng and Chen, Weizhu and Liu, Xiaodong and Gao, Jianfeng and Han, Jiawei},
  journal={arXiv preprint arXiv:1908.03265},
  year={2019}
}

jabogithub/RAdam