Stop Wasting My Time! Saving Days of ImageNet and BERT Training with Latest Weight Averaging
This addresses the time and resource costs for researchers and practitioners training vision or language models, though it is incremental as it builds on existing checkpointing and averaging techniques.
The paper tackles the problem of long training times for large models by showing that averaging the latest checkpoints can speed up training progression, saving up to ~68 GPU hours for ResNet50 on ImageNet and ~30 GPU hours for RoBERTa-Base on WikiText-103.
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.