Managing Memory by Splitting Test set predictions

I was wondering the following could work.

  1. Splitting the test folder into smaller folders
  2. Running predictions on these smaller folders
  3. Combining these predictions

I have been encountering kernel deaths whenever I try to run TTA on the test set using larger, more memory intensive architectures. Not too sure whether this would work so just posting this question, hoping for expert insights.

Any tips on how to execute this plan would be much appreciated too! Just looking for the technical terms to research or resources to look at to learn how to do this.