Averaging checkpoints along the training trajectory is a simple yet powerful approach to improve the generalization performance of Machine Learning models and reduce training time. Motivated by these potential gains, and in an effort to fairly and thoroughly benchmark this technique, we present an extensive evaluation of averaging techniques in modern Deep Learning, which we perform using AlgoPerf \citep{dahl_benchmarking_2023}, a large-scale benchmark for optimization algorithms. We investigate whether weight averaging can reduce training time, improve generalization, and replace learning rate decay, as suggested by recent literature. Our evaluation across seven architectures and datasets reveals that averaging significantly accelerates training and yields considerable efficiency gains, at the price of a minimal implementation and memory cost, while mildly improving generalization across all considered workloads. Finally, we explore the relationship between averaging and learning rate annealing and show how to optimally combine the two to achieve the best performances.

Averaging iterations of Stochastic Gradient Descent (SGD) have achieved empirical success in training deep learning models, such as Stochastic Weight Averaging (SWA), Exponential Moving Average (EMA), and LAtest Weight Averaging (LAWA). Especially, with a finite weight averaging method, LAWA can attain faster convergence and better generalization. However, its theoretical explanation is still less explored since there are fundamental differences between finite and infinite settings. In this work, we first generalize SGD and LAWA as Finite Weight Averaging (FWA) and explain their advantages compared to SGD from the perspective of optimization and generalization. A key challenge is the inapplicability of traditional methods in the sense of expectation or optimal values for infinite-dimensional settings in analyzing FWA's convergence. Second, the cumulative gradients introduced by FWA introduce additional confusion to the generalization analysis, especially making it more difficult to discuss them under different assumptions. Extending the final iteration convergence analysis to the FWA, this paper, under a convexity assumption, establishes a convergence bound $\mathcal{O}(\log\left(\frac{T}{k}\right)/\sqrt{T})$, where $k\in[1, T/2]$ is a constant representing the last $k$ iterations. Compared to SGD with $\mathcal{O}(\log(T)/\sqrt{T})$, we prove theoretically that FWA has a faster convergence rate and explain the effect of the number of average points. In the generalization analysis, we find a recursive representation for bounding the cumulative gradient using mathematical induction. We provide bounds for constant and decay learning rates and the convex and non-convex cases to show the good generalization performance of FWA. Finally, experimental results on several benchmarks verify our theoretical results.

Training Large Language Models (LLMs) incurs significant cost; hence, any strategy that accelerates model convergence is helpful. In this paper, we investigate the ability of a simple idea checkpoint averaging along the trajectory of a training run to improve both convergence and generalization quite early on during training. Here we show that models trained with high learning rates observe higher gains due to checkpoint averaging. Furthermore, these gains are amplified when checkpoints are sampled with considerable spacing in training steps. Our training recipe outperforms conventional training and popular checkpoint averaging baselines such as exponential moving average (EMA) and stochastic moving average (SWA). We evaluate our training recipe by pre-training LLMs, where high learning rates are inherently preferred due to extremely large batch sizes. Specifically, we pre-trained nanoGPT-2 models of varying sizes, small (125M), medium (335M), and large (770M)on the OpenWebText dataset, comprised of 9B tokens. Additionally, we present results for publicly available Pythia LLMs, ranging from 1B to 12B, which were trained on the PILE-deduped dataset containing 207B tokens.

Training vision or language models on large datasets can take days, if not weeks. We show that averaging the weights of the k latest checkpoints, each collected at the end of an epoch, can speed up the training progression in terms of loss and accuracy by dozens of epochs, corresponding to time savings up to ~68 and ~30 GPU hours when training a ResNet50 on ImageNet and RoBERTa-Base model on WikiText-103, respectively. We also provide the code and model checkpoint trajectory to reproduce the results and facilitate research on reusing historical weights for faster convergence.