I am encountering issues while training the YOLOv8 model on a dataset with 35,000 images and 10,000 classes. I am trying to train the model using images with a resolution of 1280×1280 but facing not enough memory error in my 8xV100 GPUs.
My current parameters:
- task=detect
- mode=train
- model=/home/e/yolov8x.pt
- imgsz=1216
- val=False
- data=/home/e/dataset/small_21/data.yaml
- epochs=60
- batch=8
- name=small_21
- exist_ok=True
- device=0,1,2,3,4,5,6,7
Are there any specific tips for setting hyperparameters, using data augmentation techniques, or other strategies to improve the model’s performance on such a large and complex dataset?
Can I improve memory usage performance?