Hyperparameter tuning for very large datasets

Course Q&A Machine Learning Engineering for Production(MLOps) Machine Learning Modeling Pipelines in Production
1

Hello,

There are some lessons which talk about hyperparameter tuning, but normally all the examples use very small neural networks and datasets.

I would like to know the best practices to work with larger datasets and deep learning models for computer vision, like resnet50. What is the to-go approach, considering I don’t want to leave my pipeline running for days? Taking a subset of the dataset is an approach mentioned in the videos, but it does not guarantee best results using the full dataset. And does it really worth to do hyperparameter optimization in these cases? Won’t be small the improvements, around 2-5% accuracy?

Thanks, Bruno

2

I don’t have experience tuning a model over a dataset like coco. That said, it’d help for you to look at research papers and follow up here to get a grasp of training on large datasets.