Object detection model inject

MRFI can directly inject errors into a object detection model. We use Ultralytics YOLOv8 as an example.

YOLOv8n official model has 168 layers with complex structure. We test it on coco128 dataset with default_fi.yaml configuration.

# Validate fault injection on yolov8 by one line command.
# Following yolov8's official docs https://docs.ultralytics.com/modes/val/.
from ultralytics import YOLO
from mrfi import MRFI, EasyConfig

model = YOLO('experiments/yolov8n.pt')  # load an YOLO official model

# Just use MRFI like this.
# Note YOLO.model is true PyTorch model, so provide model.model to MRFI here.
fi_model = MRFI(model.model, EasyConfig.load_preset('default_fi'))

# fi_model.save_config('detailconfigs/yolov8n_dt.yaml') # Optional, save detail config

# Validate the model using YOLO's function, nothing needs to change.
metrics = model.val(data='coco128.yaml') # 128 images
print('mAP0.5:', metrics.box.map50) # got 0.03~0.05

with fi_model.golden_run():
    metrics = model.val(data='coco128.yaml')
    print('mAP0.5:', metrics.box.map50) # got 0.60

The mAP0.5 metric drop to 0.035 from 0.605.

Golden run prediction:

Fault inject prediction: