I used two approaches to train a YOLOv8X detection model. In the first approach, I split the dataset into three parts, trained the model from scratch on the first part, then fine-tuned it on the second part, and finally on the third part (this was done to save memory on the GPUs). In the second approach, I simply trained the model from scratch on the combined dataset. Validation showed that the second approach performed better. Should the second approach perform better, or did I make a mistake somewhere? Combined dataset is 30000 photos and 10000 classes.
