Model soups to increase inference without increasing compute time

Leo Breiman published Bagging predictors, where he presented a method that generates multiple versions of a predictor In this paper, we compare Model Soups performances and uses them to get an aggregated predictor. Then, on three different models (ResNet, ViT and in 1999, Eric Bauer and Ron Kohavi published An Empirical EfficientNet) using three Soup Recipes (Greedy Comparison of Voting Classification Algorithms: Bagging, Soup Sorted, Greedy Soup Random and Uniform Boosting, and Variants, where they reviewed many voting soup) from [1], and reproduce the results of the classification algorithms, like Bagging and AdaBoost, and authors. We then introduce a new Soup Recipe showed interesting results. In 2000, Thomas G. Dietterich called Pruned Soup. Results from the soups were published Ensemble Methods in Machine Learning, where better than the best individual model for the pretrained he explains why ensembles can outperform single classifiers.