Submitted by **rlresearcher** t3_xvjfj4
in **MachineLearning**

https://arxiv.org/abs/2209.14981

Abstract: 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.

IdentifiableParamt1_ir31zjs wroteWeird that this paper didn't seem to cite https://arxiv.org/abs/1409.4842v1 which also used Polyak averaging on models trained on ImageNet.