Really hard RAM usage for large number of classes.

[mazatov]

I'm training the model on my own dataset and had no issues training even Yolov10x when I was detecting just one class. However ever since I expanded to many classes (320 classes) I do not have enough RAM to train even Yolov10n. What could be the issue?

I remember I had a similar problem with Yolov7 and there it was the OTA loss that was increasing the RAM usage with the number of classes and by turning that off I could decrease the RAM usage significantly. Is something similar going on here?

[mazatov]

Also the RAM seems to go up before the first epoch starts. Once it starts everything is fine and it drops.

For example I lowered my image size to 320 pixels, workers to 0, batch to 1 and it finally went through. Once the first epoch kicked of on GPU the RAM dropped to super low percentage. It seems like it's something to do with initialization.