Question About Pre-trained Weights in Custom YOLOv8 Configuration

[khandriod]

Hi,

I’m currently working with a custom YOLOv8 model configuration (yolov8_ResBlock_CBAM.yaml) and have a question regarding the use of pre-trained weights during training. After reviewing the YAML file, I noticed that there is no explicit mention of pre-trained weights. However, I did see that pretrained=True is set in the default.yaml file. If a pre-trained model is used, I am wondering which .pt file or model is being utilized.

My training script initializes the model using the following command:

model = YOLO(args.model) model.train(data=args.data_dir)

From the documentation and the provided YAML file, it's not entirely clear if pre-trained weights are being loaded by default or if the model is training from scratch with randomly initialized weights.

Could you clarify if, in the absence of an explicit weights specification in the YAML file, the YOLOv8 training process automatically uses pre-trained weights (e.g., COCO) by default? If so, is there a way to control this behavior explicitly in the configuration or training script?

Thanks for your great work.
Best regards

[RuiyangJu]

Actually, if you don't set using pre-training, no pre-training weights will be loaded.
You can confirm this by checking if you have automatically downloaded the default pre-training weights (e.g. COCO) before training.