EMA of weights: choosing the decay
Optional · supplementary note 3 of 4 for DLCV 01A · Training Deep Vision Models in Practice. Nothing in the main text depends on it; citation numbers refer to the reference list at the bottom of this page.
Navigate: this page and this track: Navigate: this page and this track
This note explains how to choose the EMA decay , why the average is biased at the start, and how batch normalization affects it.
Effective window. After updates, the EMA is a weighted sum of past weights, , where are the weights after step and is the initial average. The weight of step falls by a factor per step, so the mean age of the averaged weights is steps, about for near 1. Choose so that this window is a small fraction of the run: with and a run of 885 steps, as in the main text, the average is dominated by early weights.
Bias at the start. If is the initial (random) weights, the term dies out slowly: after steps its weight is still about . Two fixes follow from the formula. Either start the EMA only after warmup, from the current weights, or start from and divide by the total weight, , the same bias correction Adam applies to its moment estimates; the weights then sum to one at every step.
Batch normalization. The EMA averages the learnable weights, but BN layers also store running statistics computed for the current weights. Morales-Brotons et al. found that BN is the limiting factor when choosing the decay: without recomputing the statistics, long averaging windows fail, while recomputing them allows larger windows [29]. SWA, which averages checkpoints with equal weight, recomputes BN statistics with an extra pass over the training data after training [28]. To recompute, reset the running statistics, run the averaged model in training mode without gradient updates over the training data (or a large sample of it), then switch to evaluation mode.
Evaluation. Evaluate both the raw and the averaged weights at every validation step and keep the better one; it costs one extra forward pass over the validation set. EMA models are also often used as “teacher” models in larger training pipelines [29].
References
- T. He, Z. Zhang, H. Zhang, Z. Zhang, J. Xie and M. Li, “Bag of tricks for image classification with convolutional neural networks,” in Proc. IEEE/CVF CVPR, 2019. doi
- A. Karpathy, “A recipe for training neural networks,” blog post, 2019. link
- scikit-learn developers, “load_digits,” scikit-learn documentation, version 1.9.1. link
- M. Buda, A. Maki and M. A. Mazurowski, “A systematic study of the class imbalance problem in convolutional neural networks,” Neural Networks, vol. 106, 2018. doi
- T.-Y. Lin, P. Goyal, R. Girshick, K. He and P. Dollár, “Focal loss for dense object detection,” in Proc. IEEE ICCV, 2017. doi
- A. Buslaev, V. I. Iglovikov, E. Khvedchenya, A. Parinov, M. Druzhinin and A. A. Kalinin, “Albumentations: Fast and flexible image augmentations,” Information, vol. 11, no. 2, 2020. doi
- C. Shorten and T. M. Khoshgoftaar, “A survey on image data augmentation for deep learning,” Journal of Big Data, vol. 6, 2019. doi
- K. He, X. Zhang, S. Ren and J. Sun, “Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification,” in Proc. IEEE ICCV, 2015. doi
- P. Goyal, P. Dollár, R. Girshick, P. Noordhuis, L. Wesolowski, A. Kyrola, A. Tulloch, Y. Jia and K. He, “Accurate, large minibatch SGD: Training ImageNet in 1 hour,” arXiv:1706.02677 (technical report), 2017. arXiv
- S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in Proc. ICML, PMLR 37, 2015. link
- Y. Wu and K. He, “Group normalization,” in Proc. ECCV, 2018. doi
- Y. Bengio, “Practical recommendations for gradient-based training of deep architectures,” in Neural Networks: Tricks of the Trade, 2nd ed., LNCS 7700, Springer, 2012. doi
- L. N. Smith, “Cyclical learning rates for training neural networks,” in Proc. IEEE WACV, 2017. doi
- L. N. Smith, “A disciplined approach to neural network hyper-parameters: Part 1 — learning rate, batch size, momentum, and weight decay,” US Naval Research Laboratory Technical Report 5510-026, arXiv:1803.09820, 2018. arXiv
- I. Loshchilov and F. Hutter, “SGDR: Stochastic gradient descent with warm restarts,” in Proc. ICLR, 2017. arXiv
- L. N. Smith and N. Topin, “Super-convergence: Very fast training of neural networks using large learning rates,” in Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, Proc. SPIE 11006, 2019. doi
- PyTorch contributors, “torch.optim.lr_scheduler.OneCycleLR,” PyTorch 2.15 documentation, 2026. link
- K. He, X. Zhang, S. Ren and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE CVPR, 2016. doi
- J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li and L. Fei-Fei, “ImageNet: A large-scale hierarchical image database,” in Proc. IEEE CVPR, 2009. doi
- D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in Proc. ICLR, 2015. arXiv
- I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in Proc. ICLR, 2019. arXiv
- PyTorch contributors, “torch.optim.Adam,” PyTorch 2.15 documentation, 2026. link
- N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever and R. Salakhutdinov, “Dropout: A simple way to prevent neural networks from overfitting,” Journal of Machine Learning Research, vol. 15, 2014. link
- G. Huang, Y. Sun, Z. Liu, D. Sedra and K. Q. Weinberger, “Deep networks with stochastic depth,” in Proc. ECCV, 2016. doi
- L. Prechelt, “Early stopping — but when?,” in Neural Networks: Tricks of the Trade, 2nd ed., LNCS 7700, Springer, 2012. doi
- V. Godbole, G. E. Dahl, J. Gilmer, C. J. Shallue and Z. Nado, “Deep learning tuning playbook,” Google Research, GitHub repository, 2023. link
- B. T. Polyak and A. B. Juditsky, “Acceleration of stochastic approximation by averaging,” SIAM Journal on Control and Optimization, vol. 30, no. 4, 1992. doi
- P. Izmailov, D. Podoprikhin, T. Garipov, D. Vetrov and A. G. Wilson, “Averaging weights leads to wider optima and better generalization,” in Proc. UAI, 2018. arXiv
- D. Morales-Brotons, T. Vogels and H. Hendrikx, “Exponential moving average of weights in deep learning: Dynamics and benefits,” Transactions on Machine Learning Research, 2024. arXiv
- P. Micikevicius, S. Narang, J. Alben, G. Diamos, E. Elsen, D. Garcia, B. Ginsburg, M. Houston, O. Kuchaiev, G. Venkatesh and H. Wu, “Mixed precision training,” in Proc. ICLR, 2018. arXiv
- PyTorch contributors, “Automatic mixed precision package — torch.amp,” PyTorch 2.15 documentation, 2026. link
- J. Yosinski, J. Clune, Y. Bengio and H. Lipson, “How transferable are features in deep neural networks?,” in Advances in Neural Information Processing Systems 27, 2014. link
- E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang and W. Chen, “LoRA: Low-rank adaptation of large language models,” in Proc. ICLR, 2022. arXiv
- J. Howard and S. Ruder, “Universal language model fine-tuning for text classification,” in Proc. ACL, 2018. doi
- A. Kumar, A. Raghunathan, R. Jones, T. Ma and P. Liang, “Fine-tuning can distort pretrained features and underperform out-of-distribution,” in Proc. ICLR, 2022. arXiv
- C. Zhang, S. Bengio, M. Hardt, B. Recht and O. Vinyals, “Understanding deep learning requires rethinking generalization,” in Proc. ICLR, 2017. arXiv
- R. Geirhos, J.-H. Jacobsen, C. Michaelis, R. Zemel, W. Brendel, M. Bethge and F. A. Wichmann, “Shortcut learning in deep neural networks,” Nature Machine Intelligence, vol. 2, 2020. doi
- PyTorch contributors, “Reproducibility,” PyTorch 2.15 documentation, 2026. link
- Y. Xin, J. Yang, S. Luo, Y. Du, Q. Qin, K. Cen, Y. He, Z. Zhang, B. Fu, X. Yang, G. Zhai, M.-H. Yang and X. Liu, “Parameter-efficient fine-tuning for pre-trained vision models: A survey and benchmark,” arXiv:2402.02242, 2024. arXiv
- G. E. Dahl, F. Schneider, Z. Nado, N. Agarwal, C. S. Sastry, P. Hennig, S. Medapati, R. Eschenhagen, P. Kasimbeg, D. Suo et al., “Benchmarking neural network training algorithms,” arXiv:2306.07179, 2023. arXiv
- P. Kasimbeg, F. Schneider, R. Eschenhagen, J. Bae, C. S. Sastry, M. Saroufim, B. Feng, L. Wright, E. Z. Yang, Z. Nado, S. Medapati, P. Hennig, M. Rabbat and G. E. Dahl, “Accelerating neural network training: An analysis of the AlgoPerf competition,” in Proc. ICLR, 2025. arXiv
- A. Defazio, X. A. Yang, H. Mehta, K. Mishchenko, A. Khaled and A. Cutkosky, “The road less scheduled,” in Advances in Neural Information Processing Systems 37, 2024. link
- S. Hu, Y. Tu, X. Han, C. He, G. Cui, X. Long, Z. Zheng, Y. Fang, Y. Huang, W. Zhao et al., “MiniCPM: Unveiling the potential of small language models with scalable training strategies,” in Proc. Conference on Language Modeling (COLM), 2024. arXiv
- A. Hägele, E. Bakouch, A. Kosson, L. Ben Allal, L. Von Werra and M. Jaggi, “Scaling laws and compute-optimal training beyond fixed training durations,” in Advances in Neural Information Processing Systems 37, 2024. link
- M. Wortsman, G. Ilharco, J. W. Kim, M. Li, S. Kornblith, R. Roelofs, R. Gontijo Lopes, H. Hajishirzi, A. Farhadi, H. Namkoong and L. Schmidt, “Robust fine-tuning of zero-shot models,” in Proc. IEEE/CVF CVPR, 2022. doi
- M. Wortsman, G. Ilharco, S. Y. Gadre, R. Roelofs, R. Gontijo-Lopes, A. S. Morcos, H. Namkoong, A. Farhadi, Y. Carmon, S. Kornblith and L. Schmidt, “Model soups: Averaging weights of multiple fine-tuned models improves accuracy without increasing inference time,” in Proc. ICML, PMLR 162, 2022. link
- PyTorch contributors, “Automatic mixed precision examples,” PyTorch 2.15 documentation, 2026. link
- PyTorch contributors, “DistributedDataParallel,” PyTorch 2.15.0 source documentation, 2026. Official source.