In Predict Mode, I have save enabled. Since I run my Python scripts in an IDE, I often encounter result displays from prior runs. I’ve quickly glanced at the Results object, but I could not understand which attribute will enable me to get a fresh object without the images of the previous runs. I’m probably not understanding a simple setting. Please advise. Thanks.
Regards.
Hello!
It sounds like you’re having a bit of trouble with residual data from previous runs in your Results object. To ensure you’re working with fresh results each time, you can clear any stored results by reinitializing the Results object or simply ensuring that your script does not retain any references to old results between runs.
Here’s a quick tip: make sure to reinitialize your model or results variable at the start of each run. For example:
from ultralytics import YOLO
# Load a model
model = YOLO("yolo11n.pt")
# Run inference on a new source
results = model.predict(source="new_image.jpg", save=True)
This approach ensures that each run starts with a clean slate. If you are using an IDE, consider restarting the kernel or clearing variables to prevent any carryover from previous executions.
For more detailed information, you might find the Predict Mode documentation helpful. It provides insights into managing results and other useful settings.
If you have any more questions, feel free to ask. Happy coding! 
Please share the code you’re using to run inference.
Dear All, I have similar issue. If the next inference has no object, the result will be the previous bounding boxes with the same location and confidence, but different label. This happens when running on Intel GPU but the CPU is fine.
The result for the first and second inference are as follow.
Following is my code.
from pathlib import Path
from ultralytics import YOLO
import openvino as ov
IMAGE_PATH = "./data/coco_bike.jpg"
IMAGE_BLANK_PATH = "./data/blank.jpg"
def compile_model(det_model_path, device):
core = ov.Core()
det_ov_model = core.read_model(det_model_path)
ov_config = {}
#if device != "CPU":
# det_ov_model.reshape({0: [1, 3, 640, 640]})
if "GPU" in device or ("AUTO" in device and "GPU" in core.available_devices):
ov_config = {"GPU_DISABLE_WINOGRAD_CONVOLUTION": "YES"}
det_compiled_model = core.compile_model(det_ov_model, device, ov_config)
det_model = YOLO(det_model_path.parent, task="detect")
if det_model.predictor is None:
custom = {"conf": 0.25, "batch": 1, "save": False, "mode": "predict"} # method defaults
args = {**det_model.overrides, **custom}
det_model.predictor = det_model._smart_load("predictor")(overrides=args, _callbacks=det_model.callbacks)
det_model.predictor.setup_model(model=det_model.model)
det_model.predictor.model.ov_compiled_model = det_compiled_model
return det_model
def run_e2e(det_model, image_path):
res = det_model(image_path)
det_model_path = Path("yolo11s_openvino_model/yolo11s.xml")
for device in ["CPU", "GPU.0", "GPU.1"]:
det_model = compile_model(det_model_path, device)
for x in range(10):
print("\n\n###", device, " - object:")
run_e2e(det_model, IMAGE_PATH)
print("\n\n###", device, " - blank:")
run_e2e(det_model, IMAGE_BLANK_PATH)
>python inference.py
Ultralytics 8.3.142 Python-3.12.10 torch-2.8.0+cpu CPU (Intel Core(TM) Ultra 9 285K)
Loading yolo11s_openvino_model for OpenVINO inference...
Using OpenVINO LATENCY mode for batch=1 inference...
### CPU - object:
image 1/1 C:\Users\ekaak\Workspace\yolov11_blank\data\coco_bike.jpg: 480x640 1 bicycle, 2 cars, 1 parking meter, 1 dog, 61.5ms
Speed: 2.2ms preprocess, 61.5ms inference, 0.4ms postprocess per image at shape (1, 3, 480, 640)
### CPU - blank:
image 1/1 C:\Users\ekaak\Workspace\yolov11_blank\data\blank.jpg: 480x640 (no detections), 19.7ms
Speed: 1.3ms preprocess, 19.7ms inference, 0.2ms postprocess per image at shape (1, 3, 480, 640)
### GPU.0 - object:
image 1/1 C:\Users\ekaak\Workspace\yolov11_blank\data\coco_bike.jpg: 480x640 1 bicycle, 2 cars, 1 parking meter, 1 dog, 490.1ms
Speed: 1.7ms preprocess, 490.1ms inference, 0.2ms postprocess per image at shape (1, 3, 480, 640)
### GPU.0 - blank:
image 1/1 C:\Users\ekaak\Workspace\yolov11_blank\data\blank.jpg: 480x640 1 car, 1 bus, 1 truck, 1 giraffe, 1 baseball bat, 19.9ms
Speed: 1.4ms preprocess, 19.9ms inference, 0.2ms postprocess per image at shape (1, 3, 480, 640)
### GPU.1 - object:
image 1/1 C:\Users\ekaak\Workspace\yolov11_blank\data\coco_bike.jpg: 480x640 1 bicycle, 2 cars, 1 parking meter, 1 dog, 2759.1ms
Speed: 1.8ms preprocess, 2759.1ms inference, 0.2ms postprocess per image at shape (1, 3, 480, 640)
### GPU.1 - blank:
image 1/1 C:\Users\ekaak\Workspace\yolov11_blank\data\blank.jpg: 480x640 1 car, 1 bus, 1 truck, 1 giraffe, 1 baseball bat, 6.9ms
Speed: 1.7ms preprocess, 6.9ms inference, 0.2ms postprocess per image at shape (1, 3, 480, 640)
But if I re-setup the model before each inference, the GPU result is as expected. This method increases total inference time form 6ms to 40ms.
det_model.predictor.setup_model(model=det_model.model)
New code is as follow.
from pathlib import Path
from ultralytics import YOLO
import openvino as ov
IMAGE_PATH = "./data/coco_bike.jpg"
IMAGE_BLANK_PATH = "./data/blank.jpg"
def compile_model(det_model_path, device):
core = ov.Core()
det_ov_model = core.read_model(det_model_path)
ov_config = {}
#if device != "CPU":
# det_ov_model.reshape({0: [1, 3, 640, 640]})
if "GPU" in device or ("AUTO" in device and "GPU" in core.available_devices):
ov_config = {"GPU_DISABLE_WINOGRAD_CONVOLUTION": "YES"}
det_compiled_model = core.compile_model(det_ov_model, device, ov_config)
det_model = YOLO(det_model_path.parent, task="detect")
if det_model.predictor is None:
custom = {"conf": 0.25, "batch": 1, "save": False, "mode": "predict"} # method defaults
args = {**det_model.overrides, **custom}
det_model.predictor = det_model._smart_load("predictor")(overrides=args, _callbacks=det_model.callbacks)
det_model.predictor.setup_model(model=det_model.model)
det_model.predictor.model.ov_compiled_model = det_compiled_model
return det_model
def run_e2e(det_model, image_path):
results = det_model(image_path)
det_model_path = Path("yolo11s_openvino_model/yolo11s.xml")
for device in ["CPU", "GPU.0", "GPU.1"]:
det_model = compile_model(det_model_path, device)
for x in range(10):
print("\n\n###", device, " - object:")
det_model.predictor.setup_model(model=det_model.model)
run_e2e(det_model, IMAGE_PATH)
print("\n\n###", device, " - blank:")
det_model.predictor.setup_model(model=det_model.model)
run_e2e(det_model, IMAGE_BLANK_PATH)
>python inference.py
Ultralytics 8.3.142 Python-3.12.10 torch-2.8.0+cpu CPU (Intel Core(TM) Ultra 9 285K)
Loading yolo11s_openvino_model for OpenVINO inference...
Using OpenVINO LATENCY mode for batch=1 inference...
### CPU - object:
Ultralytics 8.3.142 Python-3.12.10 torch-2.8.0+cpu CPU (Intel Core(TM) Ultra 9 285K)
Loading yolo11s_openvino_model for OpenVINO inference...
Using OpenVINO LATENCY mode for batch=1 inference...
image 1/1 C:\Users\ekaak\Workspace\yolov11_blank\data\coco_bike.jpg: 480x640 1 bicycle, 2 cars, 1 parking meter, 1 dog, 2459.3ms
Speed: 4.0ms preprocess, 2459.3ms inference, 0.4ms postprocess per image at shape (1, 3, 480, 640)
### CPU - blank:
Ultralytics 8.3.142 Python-3.12.10 torch-2.8.0+cpu CPU (Intel Core(TM) Ultra 9 285K)
Loading yolo11s_openvino_model for OpenVINO inference...
Using OpenVINO LATENCY mode for batch=1 inference...
image 1/1 C:\Users\ekaak\Workspace\yolov11_blank\data\blank.jpg: 480x640 (no detections), 67.8ms
Speed: 1.5ms preprocess, 67.8ms inference, 0.2ms postprocess per image at shape (1, 3, 480, 640)
### GPU.0 - object:
Ultralytics 8.3.142 Python-3.12.10 torch-2.8.0+cpu CPU (Intel Core(TM) Ultra 9 285K)
Loading yolo11s_openvino_model for OpenVINO inference...
Using OpenVINO LATENCY mode for batch=1 inference...
image 1/1 C:\Users\ekaak\Workspace\yolov11_blank\data\coco_bike.jpg: 480x640 1 bicycle, 2 cars, 1 parking meter, 1 dog, 43.8ms
Speed: 1.6ms preprocess, 43.8ms inference, 0.2ms postprocess per image at shape (1, 3, 480, 640)
### GPU.0 - blank:
Ultralytics 8.3.142 Python-3.12.10 torch-2.8.0+cpu CPU (Intel Core(TM) Ultra 9 285K)
Loading yolo11s_openvino_model for OpenVINO inference...
Using OpenVINO LATENCY mode for batch=1 inference...
image 1/1 C:\Users\ekaak\Workspace\yolov11_blank\data\blank.jpg: 480x640 (no detections), 45.5ms
Speed: 1.6ms preprocess, 45.5ms inference, 0.2ms postprocess per image at shape (1, 3, 480, 640)
### GPU.1 - object:
Ultralytics 8.3.142 Python-3.12.10 torch-2.8.0+cpu CPU (Intel Core(TM) Ultra 9 285K)
Loading yolo11s_openvino_model for OpenVINO inference...
Using OpenVINO LATENCY mode for batch=1 inference...
image 1/1 C:\Users\ekaak\Workspace\yolov11_blank\data\coco_bike.jpg: 480x640 1 bicycle, 2 cars, 1 parking meter, 1 dog, 38.8ms
Speed: 1.6ms preprocess, 38.8ms inference, 0.2ms postprocess per image at shape (1, 3, 480, 640)
### GPU.1 - blank:
Ultralytics 8.3.142 Python-3.12.10 torch-2.8.0+cpu CPU (Intel Core(TM) Ultra 9 285K)
Loading yolo11s_openvino_model for OpenVINO inference...
Using OpenVINO LATENCY mode for batch=1 inference...
image 1/1 C:\Users\ekaak\Workspace\yolov11_blank\data\blank.jpg: 480x640 (no detections), 45.5ms
Speed: 1.7ms preprocess, 45.5ms inference, 0.2ms postprocess per image at shape (1, 3, 480, 640)
Is there a method to reset just the output especially the one in GPU?
Thanks.
Additional information: I use Windows 11, converted the model to OpenVINO with dynamic input and NMS. Is it possible to reset the confidence result before NMS?
Does the issue occur if you run prediction normally as described in the docs?
Ultralytics already supports passing the specific device for inference for OpenVINO. You don’t need to perform all these monkey-patching.
@Toxite , it works as expected if the model is exported without NMS as follow.
model.export(format="openvino")
The issue occurs if the model is exported with NMS.
model.export(format="openvino", nms=True)
And it only occurs on GPU?
@Toxite , yes only on GPU (iGPU/dGPU), with NMS head in the model, as well as static/dynamic input.
If that’s the case, then it seems like OpenVINO bug. You cna open an issue in their GitHub Ultralytics doesn’t have different handling for CPU and GPU OpenVINO inference, so there shouldn’t be any difference unless it’s due to OpenVINO bug.
This issue happens on OpenVINO 2025.2 and 2025.3. The fix has already in place (PR-32067). I have tested successfully on both iGPU and dGPU using OpenVINO nightly 20250917 targeting for OpenVINO 2025.4 release. Thanks.
Thanks for the update, great to see there’s a fix in the works!