Splitting Multi-Class Dataset & Rebalancing create severly unbalanced Val/Test sets at Export

Object Detection
FireFox - Chrome
Ubuntu 18.04

Posted feedback several days ago about this Bug, but nothing back from anyone yet.

Upon export of multi-class dataset, Roboflow does NOT generate a balance of images amongst the classes to the Val or Test Sets. In many cases, I only get one object class in my Val Set. Explained that was having to export entire dataset out of Robo, run a script to randomly shuffle then create balanced Train/Val Sets, then re-upload dataset to Robo while maintaining the proper balances.

NOW… without touching the re-balance function, at Export Robo is unwinding both the Val and the Train sets that each have precise numbers of images from each class and is randomly rebuilding the sets. (e.g.) In Val, it now appears to be 80% one class, an small assortment of a few other classes, and is TOTALLY EXCLUDING ONE CLASS ALL TOGETHER. Needless to say, this is a real problem.

Please advise and/or address as soon as possible.


Hi @Sparrowtech

We’re not sure we follow. Can you send us steps to reproduce this error or a video of it being triggered?

Thanks Kelly,
After further review and inspection, there does not appear to be a problem with Roboflow. In our Colab notebook for this experiment, which is where we were examining the Val set composition… it only shows the first ~1000 images and/or labels. I suspect it caches the rest which are hidden; what is visible is largely just one class which is where the concern/confusion originated. Apologies for the mistake on our part.

Best regards,