Problem face when using Train/Test Split

  1. After I Rebalance the dataset and save it, it do not make any changes base on my allocation of dataset. For example, I want to allocate 600 images to my train set, 50 images to my validation set, and 50 images to my test set by using the rebalance feature. But, after I clicked save, there is nothing changes to my allocation of dataset.

  2. After I export into the yolo pytorch format, the validation set and test set of my dataset is changes. For example, my validation set should be 100 images, and test set should be 100 images, but after I exported it, the validation set become 30 images, and test set become 45 images, it do not tally with the preset number of validation and test set of my data.

Appreciate it!

Hi @Lik_Seng_Fong :wave: ! Could you send me the link to your project?

@Lik_Seng_Fong - for some datasets, if there are larger image sizes or a lot of images, or if the process queue is backed up, it can take a moment to load the train/valid/test split you just reset

  • it is best to give it about 30 seconds to fully refresh the splits if there are more than 500 images you are rebalancing

@SkalskiP - no need, I ran a recalculation on it

@Lik_Seng_Fong - I also took a look at your dataset. You’ll want to delete all of the labels that say Good from the dataset. You will keep the images, but just delete the labels.

You can do this manually - or use the Modify Classes feature to drop the labels for Good (un-checked box) entirely:

You will receive better training results this way.

@Mohamed - Thank You So Much Bro, I think is my dataset more than 500 images, so it will have lag issues during the rebalancing. For removing the ‘Good’ label, actually I decided to create a new project on object detection, however, if I do not make any label on the images, Roboflow will not allow me to put these ‘images with no annotation’ to the dataset. How can I solve it.

@Lik_Seng_Fong - no problem at all, you’re very welcome!

To add images without labels to the dataset, label them as Null images: The Difference Between Missing and Null Annotations