View in #questions on Slack
@Pranisha_Chaturvedi: Hello Everyone! I am Pranisha and I am a Master’s student at Hochschule Bonn-Rhein-Sieg. I am using Flower for my Master’s thesis for semantic segmentation. I wanted to know about some best practices for setting the hyperparameters in FL. Usually, for a DL model, we can use something like Optuna, but what about the FL parameters like rounds, clients, partitioner (like alpha for Dirichlet partitioning) and different aggregation strategies? I am a beginner and so far I am just changing values manually for every experiment. Thank you for your help 
@Manjil: In Dirichlet partitioning, the alpha value controls how the data is shared among clients. If you want the data to be IID, you can set alpha to a high value like 1000. For highly non-IID data, set it to a lower value, typically less than 1.0. However, the best alpha depends on your dataset and task, so it’s important to experiment and find the right balance between utility and non-IID-ness.
@Chong_Shen: Hello @Pranisha_Chaturvedi, there’s in fact a previous study that used Optuna to select hyperparameters for a Flower app: https://doi.org/10.1145/3555776.3577847. Although the study uses an older version of Flower, I think you can adapt the techniques in the paper for the latest version of Flower. Hope that helps 
.
@Pranisha_Chaturvedi: Thank you @Manjil and @Chong_Shen for your suggestions! I will give this a try.