Empezar

Evaluación comparativa de modelos con Ultralytics YOLO

Ultralytics YOLO ecosystem and integrations

Visualización de la evaluación comparativa#

Introducción#

Una vez que hayas entrenado y validado tu modelo, el siguiente paso es evaluar su rendimiento en producción. El modo Benchmark de Ultralytics YOLO26 mide la velocidad y la precisión de tu modelo en distintos formatos de exportación.

Consulta la versión preliminar de YOLO27, aún no publicada para ver los resultados preliminares de velocidad.



Ver: Compara el rendimiento de los modelos Ultralytics YOLO26 | ¿Cómo comparar el rendimiento de los modelos en distintos equipos?

¿Por qué es crucial la evaluación comparativa?#

  • Decisiones fundamentadas: Conoce las compensaciones entre velocidad y precisión.
  • Asignación de recursos: Comprende cómo funcionan los distintos formatos de exportación en diferentes equipos.
  • Optimización: Averigua qué formato de exportación ofrece el mejor rendimiento para tu caso de uso concreto.
  • Eficiencia de costes: Aprovecha mejor los recursos de hardware a partir de los resultados de la evaluación comparativa.

Métricas clave del modo de evaluación comparativa#

Formatos de exportación compatibles#

  • ONNX: Para un rendimiento óptimo de la CPU
  • TensorRT: Para la máxima eficiencia de la GPU
  • OpenVINO: Para optimizar el hardware Intel
  • CoreML, TensorFlow SavedModel y más: Para distintas necesidades de implementación.
Consejo
  • Exporta a ONNX u OpenVINO para lograr hasta 3 veces más velocidad en la CPU.
  • Exporta a TensorRT para lograr hasta 5 veces más velocidad en la GPU.

Ejemplos de uso#

Instalación recomendada

Instala Ultralytics con las dependencias de exportación antes de ejecutar la evaluación comparativa para evitar que falten paquetes.

pip install "ultralytics[export]"

Ejecuta pruebas de rendimiento de YOLO26n en todos los formatos de exportación compatibles (ONNX, TensorRT, etc.).

Ejemplo
from ultralytics.utils.benchmarks import benchmark

# Evaluar el rendimiento en la GPU
benchmark(model="yolo26n.pt", data="coco8.yaml", imgsz=640, device=0)

# Evaluar un formato de exportación específico
benchmark(model="yolo26n.pt", data="coco8.yaml", imgsz=640, format="onnx")

Argumentos#

Argumentos como model, data, imgsz, quantize, device, verbose y format te permiten adaptar las pruebas de rendimiento a tus necesidades y comparar el rendimiento de distintos formatos de exportación.

ClaveValor predeterminadoDescripción
modelNoneEspecifica la ruta al archivo del modelo. Acepta los formatos .pt y .yaml, por ejemplo, "yolo26n.pt" para modelos preentrenados o archivos de configuración.
dataNoneRuta al YAML del conjunto de datos para la evaluación comparativa, que suele incluir rutas y ajustes para los datos de validación. Ejemplo: "coco8.yaml". En cambio, la clasificación utiliza un directorio de conjunto de datos o el nombre de un conjunto de datos integrado (por ejemplo, imagenet10).
imgsz640Tamaño de la imagen de entrada del modelo. Debe ser un único entero para imágenes cuadradas (por ejemplo, 640); benchmark() solo admite tamaños de imagen cuadrados.
quantizeNonePrecisión solicitada: 16 (FP16) o 8 (INT8; los métodos PTQ necesitan datos de calibración data/fraction, pero los métodos que solo cuantizan los pesos no); 32/sin definir es FP32 cuando el formato lo admite. Si un formato no puede exportar en FP32, rechaza un 32 explícito o usa la precisión que requiere. La fila de PyTorch nativo no se exporta: solo 16 la afecta y selecciona la inferencia FP16 fuera de la CPU; de lo contrario, se ejecuta en FP32. Después, cada formato ejecuta la inferencia con su propia precisión en tiempo de ejecución. Sustituye las marcas obsoletas half/int8.
deviceNoneDefine el dispositivo de cálculo para la evaluación comparativa, como "cpu" o "cuda:0"; si no se configura, se usa la primera GPU disponible y, si no, la CPU. Con format="openvino" o "litert", "xpu" y "npu" seleccionan la GPU y la NPU Intel; consulta dispositivos de inferencia.
verboseFalseSi se especifica True o un número decimal, los errores de la evaluación comparativa (salvo las aserciones) se propagan en lugar de registrarse; un número decimal como 0.25 también comprueba que la métrica de cada formato que se ejecute correctamente supere ese umbral.
eps0.001Pequeño épsilon (milisegundos) que se añade al tiempo de inferencia por imagen antes de convertirlo a FPS para evitar la división por cero. Rara vez se modifica.
format''Evalúa únicamente el formato de exportación especificado (por ejemplo, format=onnx). Déjalo en blanco para probar automáticamente todos los formatos compatibles.
Valores predeterminados de la función independiente `benchmark()`

La función independiente benchmark() (from ultralytics.utils.benchmarks import benchmark) utiliza sus propios valores predeterminados de firma, en lugar de los valores de la tabla anterior, en particular model="yolo26n.pt", imgsz=160 y device="cpu"; pasa explícitamente imgsz y device para usar los mismos valores que la CLI de yolo benchmark.

Formatos de exportación#

Las evaluaciones comparativas intentarán ejecutarse automáticamente en todos los formatos de exportación posibles que se indican a continuación. También puedes ejecutarlas para un formato concreto mediante el argumento format, que acepta cualquiera de los formatos mencionados a continuación.

FormatoArgumento formatModeloMetadatosArgumentos
PyTorch-yolo26n.pt✅-
TorchScripttorchscriptyolo26n.torchscript✅imgsz, quantize, dynamic, nms, batch, device
ONNXonnxyolo26n.onnx✅imgsz, quantize, dynamic, simplify, opset, nms, batch, data, fraction, device
OpenVINOopenvinoyolo26n_openvino_model/✅imgsz, quantize, dynamic, nms, batch, data, fraction, device
TensorRTengineyolo26n.engine✅imgsz, quantize, dynamic, simplify, opset, workspace, nms, batch, data, fraction, device
CoreMLcoremlyolo26n.mlpackage✅imgsz, dynamic, quantize, nms, batch, device
Apple Core AIcoreaiyolo26n.aimodel✅imgsz, batch, quantize
TF SavedModelsaved_modelyolo26n_saved_model/✅imgsz, quantize, opset, nms, batch, data, fraction, device
TF GraphDefpbyolo26n.pb❌imgsz, opset, batch, device
TF Edge TPUedgetpuyolo26n_edgetpu.tflite✅imgsz, quantize, opset, data, fraction, device
LiteRTlitertyolo26n.tflite✅imgsz, quantize, batch, data, fraction, device
PaddlePaddlepaddleyolo26n_paddle_model/✅imgsz, batch, device
MNNmnnyolo26n.mnn✅imgsz, batch, dynamic, quantize, simplify, opset, nms, device
NCNNncnnyolo26n_ncnn_model/✅imgsz, quantize, device
IMX500imxyolo11n_imx_model/✅imgsz, quantize, data, fraction, nms, device
RKNNrknnyolo26n_rknn_model/✅imgsz, batch, name, quantize, simplify, opset, data, fraction, device
ExecuTorchexecutorchyolo26n_executorch_model/✅imgsz, quantize, batch, data, fraction, device
Axeleraaxelerayolo26n_axelera_model/✅imgsz, batch, quantize, data, fraction, device
DEEPXdeepxyolo26n_deepx_model/✅imgsz, quantize, simplify, opset, data, optimize, device
Qualcomm QNNqnnyolo26n_qnn.onnx✅imgsz, batch, name, quantize, simplify, opset, data, fraction, device
Hailohailoyolo26n_hailo_model/✅imgsz, name, quantize, data, fraction, simplify, conf, iou, device
Huawei Ascendascendyolo26n_ascend_model/✅imgsz, batch, name, quantize, opset, simplify, nms, device
AMD Xilinxxilinxyolo26n_xilinx_model/✅imgsz, name, quantize, data, fraction, opset, simplify, device
NXPnxpyolo26n_nxp.tflite✅imgsz, quantize, data, fraction, device

nms=None genera salidas sin procesar de forma predeterminada para aplicar NMS externamente (YOLO26 mantiene su cabezal sin NMS en Axelera). Establece nms=False para seleccionar un cabezal sin NMS disponible; los formatos no compatibles recurren a su ruta de salida nativa. Las entradas nms anteriores identifican los formatos que pueden integrar NMS con nms=True.

Consulta todos los detalles de export en la página Exportar.

Preguntas frecuentes#

  • Ultralytics YOLO26 ofrece un modo Benchmark para evaluar el rendimiento de tu modelo en distintos formatos de exportación. Este modo informa de métricas clave, como la precisión media (mAP50-95), la precisión y el tiempo de inferencia en milisegundos. Puedes ejecutar benchmarks con Python o con comandos CLI. Por ejemplo, para ejecutar un benchmark en una GPU:

    Ejemplo
    from ultralytics.utils.benchmarks import benchmark
    
    # Evaluar el rendimiento en la GPU
    benchmark(model="yolo26n.pt", data="coco8.yaml", imgsz=640, device=0)

    Para obtener más información sobre los argumentos de evaluación comparativa, visita la sección Argumentos.

  • Exportar los modelos YOLO26 a distintos formatos, como ONNX, TensorRT y OpenVINO, te permite optimizar el rendimiento en función del entorno de implementación. Por ejemplo:

    • ONNX: Proporciona hasta 3 veces más velocidad en la CPU.
    • TensorRT: Ofrece hasta 5 veces más velocidad en la GPU.
    • OpenVINO: Está optimizado específicamente para el hardware Intel.

    Estos formatos mejoran la velocidad de inferencia en el hardware para el que están diseñados, lo que hace que tus modelos sean más eficientes en diversas aplicaciones del mundo real. Visita la página Exportar para obtener todos los detalles.

  • Evaluar tus modelos YOLO26 es fundamental por varios motivos:

    • Decisiones fundamentadas: Comprende las compensaciones entre velocidad y precisión.
    • Asignación de recursos: Evalúa el rendimiento en distintas opciones de hardware.
    • Optimización: Determina qué formato de exportación ofrece el mejor rendimiento para casos de uso específicos.
    • Eficiencia de costes: Optimiza el uso del hardware a partir de los resultados de la evaluación comparativa.

    Métricas clave como mAP50-95, la precisión Top-1 y el tiempo de inferencia ayudan a realizar estas evaluaciones. Consulta la sección Métricas clave para obtener más información.

  • YOLO26 admite diversos formatos de exportación, cada uno adaptado a un hardware y unos casos de uso específicos:

    • ONNX: La mejor opción para el rendimiento de la CPU.
    • TensorRT: Ideal para la eficiencia de la GPU.
    • OpenVINO: Optimizado para el hardware Intel.
    • CoreML y TensorFlow: CoreML para dispositivos Apple; formatos de TensorFlow para implementaciones móviles, web y de aprendizaje automático en general.

    Para ver la lista completa de formatos compatibles y sus respectivas ventajas, consulta la sección Formatos de exportación compatibles.

  • Al ejecutar evaluaciones comparativas, puedes personalizar varios argumentos según tus necesidades:

    • model: Ruta al archivo del modelo (por ejemplo, "yolo26n.pt").
    • data: Ruta al YAML del conjunto de datos (por ejemplo, "coco8.yaml"); en cambio, la clasificación utiliza un directorio de conjunto de datos o el nombre de un conjunto de datos integrado (por ejemplo, imagenet10).
    • imgsz: Tamaño de la imagen de entrada cuadrada, indicado como un único entero, por ejemplo, 640. El modo de evaluación comparativa utiliza el mismo tamaño de imagen cuadrada en PyTorch y en los formatos exportados para garantizar una comparación justa.
    • quantize: Precisión solicitada: 16 para FP16, 8 para INT8 (útil para dispositivos edge); 32/sin definir es FP32 cuando el formato lo admite. Si un formato no puede exportar en FP32, rechaza un 32 explícito o usa la precisión que requiere. La fila de PyTorch nativo no se exporta: solo 16 la afecta y selecciona la inferencia FP16 fuera de la CPU; de lo contrario, se ejecuta en FP32. Después, cada formato ejecuta la inferencia con su propia precisión en tiempo de ejecución.
    • device: especifica el dispositivo de cálculo (p. ej., "cpu", "cuda:0") o "xpu" y "npu" para la GPU y la NPU Intel con format="openvino" o "litert".
    • verbose: Propaga los errores de la evaluación comparativa en lugar de registrarlos; pasa un número decimal para exigir también que la métrica de cada formato supere ese umbral.

    Para ver la lista completa de argumentos, consulta la sección Argumentos.

Comentarios39

Haz una pregunta, comparte tus resultados o ayuda a otra persona.

  1. bayda369GitHub

    Hi, I want to customize the yolov8 to detect and track only the objects determined by the user, like selecting 3 persons in different positions out of 10 persons in the video appears in the first frame of the video and ignore any other persons, what are your suggestions?

    1
    1. Paula DerrengerUltralytics

      @bayda369 hello,

      For customizing YOLOv8 to detect and track specific objects (like selecting certain individuals from a video), you can approach this by integrating a pre-processing step to identify and tag the objects of interest in the first frame. Here’s a brief outline on how you might proceed:

      1. Object Detection: Use YOLOv8 to detect all persons in the first frame.
      2. Object Selection: Implement a method to select the desired objects. This could be manual selection via bounding box coordinates or an automated method based on certain criteria.
      3. Tracking: Once the objects are tagged, you can use the tracking functionality to follow these objects across frames. You might need to modify the tracking logic to focus only on the tagged objects.

      For the tracking part, you can utilize the existing tracking capabilities of YOLOv8, which supports multiple tracking algorithms. You may need to customize the tracking code to ensure it only tracks the initially selected objects.

      This approach requires some custom coding, especially for the selection and modified tracking logic. If you need further guidance on the specific code modifications, feel free to ask!

    2. bayda369GitHub

      yes please, I need further guidance on the specific code modifications

    3. Paula DerrengerUltralytics(editado)

      Hello @bayda369,

      Absolutely, I'd be happy to guide you through the code needed to track only selected objects in a video. The simplest way is to use the built-in tracker, which gives every person a persistent ID, and then keep only the IDs you selected in the first frame.

      Step-by-Step Guide
      1. Track persons: run tracking with classes=[0] (person).
      2. Select IDs: choose the IDs of the persons you want from the first frame (manually or by criteria).
      3. Filter: ignore every other ID in later frames.
      Code Example
      from ultralytics import YOLO
      
      model = YOLO("yolo26n.pt")
      
      selected_ids = None
      for r in model.track("path/to/your/video/file.mp4", classes=[0], persist=True, stream=True):
          if r.boxes.id is None:
              continue
          ids = r.boxes.id.int().tolist()
      
          if selected_ids is None:
              # For simplicity, select the first 3 persons; replace with your own selection method
              selected_ids = set(ids[:3])
      
          keep = [i for i, t in enumerate(ids) if t in selected_ids]
          boxes = r.boxes.xyxy[keep]  # boxes of the selected persons only
          track_ids = [ids[i] for i in keep]
          print(track_ids, boxes)

      Feel free to customize the selection method as per your requirements. If you need further assistance, don't hesitate to ask! 😊

  2. AA278-HUBGitHub

    it seems benchmark only run on Linux System ,It couldn't run on Window Enviroment?

    1. Glenn JocherUltralytics(editado)

      It runs on Windows too—only some export formats are OS‑specific (e.g., CoreML, EdgeTPU and RKNN aren’t supported on Windows); try a single supported format to verify your setup:

      yolo benchmark model=yolo26n.pt data=coco8.yaml format=onnx device=cpu

      If it still fails, please paste the yolo checks output and the full traceback here; format/OS notes are summarized in the benchmarks reference at ultralytics.utils.benchmarks.

    2. Glenn JocherUltralytics(editado)

      Glad it helped! On Windows you can also benchmark a fast CPU path with OpenVINO:

      yolo benchmark model=yolo26n.pt data=coco8.yaml format=openvino

      Setup and tips are in the OpenVINO export guide.

  3. andyg2GitHub

    Which GPUs are used in these benchmarks? I imagine the result would be different between a GTX 1650 and an H100?

    1. Glenn JocherUltralytics(editado)

      @andyg2 you're absolutely right that benchmark results vary significantly between different GPU models. The documentation doesn't specify which exact GPUs were used for the benchmark results shown. For accurate performance metrics on your specific hardware, I'd recommend running the benchmarks directly on your system using:

      from ultralytics.utils.benchmarks import benchmark
      benchmark(model="yolo26n.pt", data="coco8.yaml", imgsz=640, device=0)

      This will give you hardware-specific results for your GTX 1650 or any other GPU you're using, which will be much more relevant for your deployment needs than generic benchmark numbers.

    2. Glenn JocherUltralytics

      @andyg2 that's a very valid point! Benchmark results without hardware specifications are indeed not very meaningful for comparison purposes. You're absolutely right that performance metrics should always include the testing hardware context. For meaningful benchmarks, it's best to run tests on your specific hardware setup to get accurate performance expectations for your deployment environment.

    3. andyg2GitHub(editado)

      @glenn-jocher Thanks for the reply. My problem is I need to purchase over 120 small GPUs for a computer vision project.

      I guess I need to buy 10 GPUs and see how each of them do, because nobody at Ultralytics can remember how they tested their own code.

      Sorry if I'm coming across a little sarcastic here, I'm just so flabbergasted, not that this information is not included - everyone makes mistakes in documentation, but that nobody there thinks it should be included.

    4. Glenn JocherUltralytics

      @andyg2 you're absolutely right, and I completely understand your frustration. This is a valid oversight on our part - benchmark results without hardware specifications are essentially meaningless for making informed purchasing decisions, especially at your scale.

      For your project requiring 120 GPUs, I'd suggest focusing on a few key considerations: look for GPUs with sufficient VRAM for your model size (typically 8GB+ for larger YOLO models), prioritize newer architectures that support TensorRT optimization well (RTX 30/40 series, A-series), and consider power efficiency for large-scale deployments.

      Your approach of testing with 10 GPUs first is actually the most reliable path forward given the current documentation gap. This will give you real-world performance data specific to your models, datasets, and deployment constraints.

      Thank you for the feedback - this highlights an important area where our documentation needs improvement to better serve users making hardware decisions.

      1
  4. josiahkim29GitHub

    How do i proceed to calculate the mean Average Precision of the predictions? I have noted the predictions = model(10 sample images from COCO val2017 dataset)

    is a bit different especially when you want to get the classes and scores of the predictions

    for pred in predictions[0]:
    x, y, w, h, class_id, score = pred[:6]

    1. Glenn JocherUltralytics

      To calculate mean Average Precision (mAP) for your predictions, you can use the val method in Ultralytics YOLO, which evaluates the model on a dataset and provides mAP metrics. For more details, refer to the benchmarking guide.

    2. josiahkim29GitHub

      @glenn-jocher thank you for the response. Based on the benchmarking guide I see that I need a custom.yaml file for this specific images I am using. Is this the case? If so how do you proceed to create a custom.yaml file.

    3. Glenn JocherUltralytics

      To create a custom yaml file for your specific images, you'll need to define the dataset paths and class names. You can start by copying an existing dataset file like coco128.yaml and modifying it to suit your dataset structure. For detailed guidance, please refer to the dataset preparation section in the documentation.

  5. xpeedent289GitHub

    Hi there, I have a question that, which dataset is use to benchmark, train, val or test ?

  6. xpeedent289GitHub

    Hi sir! when I use benchmark for "yolov8n.pt", it auto exports TorchScript, ONNX,... and auto benchmark for these too, but I just wanna benchmark for yolov8n.pt, so how can I adjust it ?

    1. Glenn JocherUltralytics(editado)

      To benchmark only the yolov8n.pt model without exporting to other formats, you can use the format argument to limit the benchmark to the PyTorch model. Here's how you can do it:

      from ultralytics.utils.benchmarks import benchmark
      
      # Benchmark only the PyTorch model ("-" is the PyTorch format)
      benchmark(model="yolov8n.pt", data="coco8.yaml", imgsz=640, device=0, format="-")

      Or using the CLI:

      yolo benchmark model=yolov8n.pt data=coco8.yaml imgsz=640 device=0 format=-

      This will ensure that only the PyTorch model is benchmarked. For more details, you can visit the benchmark documentation.

    2. themeowsketeerGitHub

      @glenn-jocher I got this error. TypeError: benchmark() got an unexpected keyword argument 'format' I don't think by default benchmark can accept custom format argument?

    3. Glenn JocherUltralytics(editado)

      The benchmark function supports a format argument that limits the run to one export format, e.g. format=onnx. To benchmark only the PyTorch model, use format="-", for example yolo benchmark model=yolov8n.pt data=coco8.yaml format=-. Leave format unset to benchmark all formats. See the benchmark documentation for details.

  7. themeowsketeerGitHub

    Greetings, I'm facing the following error when trying to benchmark using trained weight from RT-DeTR-L model, which is known to be fully supported by Ultralytics:

    KeyError Traceback (most recent call last) File /opt/conda/lib/python3.10/site-packages/ultralytics/utils/benchmarks.py:133, in benchmark(model, data, imgsz, half, int8, device, verbose) 132 assert i != 5 or platform.system() == "Darwin", "inference only supported on macOS>=10.13" # CoreML --> 133 exported_model.predict(ASSETS / "bus.jpg", imgsz=imgsz, device=device, half=half) 135 # Validate

    File /opt/conda/lib/python3.10/site-packages/ultralytics/engine/model.py:563, in Model.predict(self, source, stream, predictor, **kwargs) 562 self.predictor.set_prompts(prompts) --> 563 return self.predictor.predict_cli(source=source) if is_cli else self.predictor(source=source, stream=stream)

    File /opt/conda/lib/python3.10/site-packages/ultralytics/engine/predictor.py:168, in BasePredictor.call(self, source, model, stream, *args, **kwargs) 167 else: --> 168 return list(self.stream_inference(source, model, *args, **kwargs))

    File /opt/conda/lib/python3.10/site-packages/torch/utils/_contextlib.py:36, in _wrap_generator..generator_context(*args, **kwargs) 35 with ctx_factory(): ---> 36 response = gen.send(None) 38 while True:

    File /opt/conda/lib/python3.10/site-packages/ultralytics/engine/predictor.py:274, in BasePredictor.stream_inference(self, source, model, *args, **kwargs) 273 if self.args.verbose or self.args.save or self.args.save_txt or self.args.show: --> 274 s[i] += self.write_results(i, Path(paths[i]), im, s) 276 # Print batch results

    File /opt/conda/lib/python3.10/site-packages/ultralytics/engine/predictor.py:334, in BasePredictor.write_results(self, i, p, im, s) 333 result.save_dir = self.save_dir.str() # used in other locations --> 334 string += f"{result.verbose()}{result.speed['inference']:.1f}ms" 336 # Add predictions to image

    File /opt/conda/lib/python3.10/site-packages/ultralytics/engine/results.py:633, in Results.verbose(self) 632 n = (boxes.cls == c).sum() # detections per class --> 633 log_string += f"{n} {self.names[int(c)]}{'s' * (n > 1)}, " 634 return log_string

    KeyError: 25

    During handling of the above exception, another exception occurred:

    AssertionError Traceback (most recent call last) Cell In[11], line 3 1 from ultralytics.utils.benchmarks import benchmark ----> 3 benchmark( 4 model = "/kaggle/working/weight_rtdetr/last_rtdetr.pt", 5 data = "Yolo_path_fixed.yaml", 6 imgsz = 640, 7 device = 0, 8 half = False, 9 verbose = True 10 )

    File /opt/conda/lib/python3.10/site-packages/ultralytics/utils/benchmarks.py:146, in benchmark(model, data, imgsz, half, int8, device, verbose) 144 except Exception as e: 145 if verbose: --> 146 assert type(e) is AssertionError, f"Benchmark failure for {name}: {e}" 147 LOGGER.warning(f"ERROR ❌️ Benchmark failure for {name}: {e}") 148 y.append([name, emoji, round(file_size(filename), 1), None, None, None]) # mAP, t_inference

    AssertionError: Benchmark failure for PyTorch: 25

    Any idea how this exception is caused? My code snippet is as follows: from ultralytics.utils.benchmarks import benchmark

    benchmark( model = "/kaggle/working/weight_rtdetr/last_rtdetr.pt", data = "Yolo_path_fixed.yaml", imgsz = 640, device = 0, half = False, verbose = True )

    1. Glenn JocherUltralytics

      @themeowsketeer thank you for reporting this issue. It seems like the error is related to a missing class name in the model's names attribute. Please ensure that your model and dataset are correctly configured and that the latest versions of the Ultralytics package are installed. If the issue persists, try running the benchmark with a different model or dataset to isolate the problem. For more details on benchmarking, you can refer to our documentation.

    2. themeowsketeerGitHub

      YOLO model works flawlessly with the command, while RT-DeTR models just do not work at all. AssertionError returns different error code everytime: 8, 13 or 25. Any other method I could try?

    3. Glenn JocherUltralytics

      It seems like the issue might be related to the RT-DeTR model's class names. Please ensure your dataset and model are correctly configured, and try running the benchmark with the latest version of the Ultralytics package. If the problem persists, consider testing with a different model or dataset to isolate the issue. For more details, refer to our benchmarking documentation.

  8. WooFy11GitHub

    @title Speed Tracker

    import cv2 import numpy as np from ultralytics import YOLO

    from ultralytics.utils.checks import check_imshow from ultralytics.utils.plotting import Annotator, colors

    from collections import defaultdict

    track_history = defaultdict(lambda: []) model = YOLO("/content/drive/MyDrive/DATASET_TUGASAKHIR/PlayerResults/weights/best.pt") names = model.model.names video_path = "/content/drive/MyDrive/DATASET_TUGASAKHIR/testvid/BDEvid1.mp4" cap = cv2.VideoCapture(video_path) assert cap.isOpened(), "Error reading video file"

    w, h, fps = (int(cap.get(x)) for x in (cv2.CAP_PROP_FRAME_WIDTH, cv2.CAP_PROP_FRAME_HEIGHT, cv2.CAP_PROP_FPS))

    result = cv2.VideoWriter("SpeedTracker.avi", cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))

    time_per_frame = 1.0 / fps real_world_scale = 0.05 # Example: 0.05 meters per pixel

    while cap.isOpened(): success, frame = cap.read() if success: results = model.track(frame, persist=True, verbose=False, tracker="/content/drive/MyDrive/DATASET_TUGASAKHIR/Tracker/bytetrack.yaml") boxes = results[0].boxes.xyxy.cpu()

        if results[0].boxes.id is not None:
    
            # Extract prediction results
            clss = results[0].boxes.cls.cpu().tolist()
            track_ids = results[0].boxes.id.int().cpu().tolist()
            confs = results[0].boxes.conf.float().cpu().tolist()
    
            # Annotator Init
            annotator = Annotator(frame, line_width=2)
    
            for box, cls, track_id in zip(boxes, clss, track_ids):
                # Store tracking history
                track = track_history[track_id]
                track.append((int((box[0] + box[2]) / 2), int((box[1] + box[3]) / 2)))
                if len(track) > 30:
                    track.pop(0)
    
                if len(track) >= 2:
                    # Calculate pixel distance between the last and second to last points
                    dx = track[-1][0] - track[-2][0]
                    dy = track[-1][1] - track[-2][1]
                    pixel_distance = (dx ** 2 + dy ** 2) ** 0.5
                    real_distance = pixel_distance * real_world_scale
                    speed = real_distance / time_per_frame  # speed in meters per second
                    label = f"{names[int(cls)]} {track_id} {speed:.2f} m/s"
                    annotator.box_label(box, color=colors(int(cls), True), label=label)
    
                    # Plot tracks
                    points = np.array(track, dtype=np.int32).reshape((-1, 1, 2))
                    cv2.circle(frame, (track[-1]), 7, colors(int(cls), True), -1)
                    cv2.polylines(frame, [points], isClosed=False, color=colors(int(cls), True), thickness=2)
                else:
                    label = f"{names[int(cls)]} {track_id}"
                    annotator.box_label(box, color=colors(int(cls), True), label=label)
    
        result.write(frame)
        if cv2.waitKey(1) & 0xFF == ord("q"):
            break
    else:
        break

    result.release() cap.release() cv2.destroyAllWindows()

    how to collect the ground truth ID's from this code

  9. WooFy11GitHub

    @title Speed Tracker

    import cv2 import numpy as np from ultralytics import YOLO

    from ultralytics.utils.checks import check_imshow from ultralytics.utils.plotting import Annotator, colors

    from collections import defaultdict

    track_history = defaultdict(lambda: []) model = YOLO("/content/drive/MyDrive/DATASET_TUGASAKHIR/PlayerResults/weights/best.pt") names = model.model.names video_path = "/content/drive/MyDrive/DATASET_TUGASAKHIR/testvid/BDEvid1.mp4" cap = cv2.VideoCapture(video_path) assert cap.isOpened(), "Error reading video file"

    w, h, fps = (int(cap.get(x)) for x in (cv2.CAP_PROP_FRAME_WIDTH, cv2.CAP_PROP_FRAME_HEIGHT, cv2.CAP_PROP_FPS))

    result = cv2.VideoWriter("SpeedTracker.avi", cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))

    time_per_frame = 1.0 / fps real_world_scale = 0.05 # Example: 0.05 meters per pixel

    while cap.isOpened(): success, frame = cap.read() if success: results = model.track(frame, persist=True, verbose=False, tracker="/content/drive/MyDrive/DATASET_TUGASAKHIR/Tracker/bytetrack.yaml") boxes = results[0].boxes.xyxy.cpu()

        if results[0].boxes.id is not None:
    
            # Extract prediction results
            clss = results[0].boxes.cls.cpu().tolist()
            track_ids = results[0].boxes.id.int().cpu().tolist()
            confs = results[0].boxes.conf.float().cpu().tolist()
    
            # Annotator Init
            annotator = Annotator(frame, line_width=2)
    
            for box, cls, track_id in zip(boxes, clss, track_ids):
                # Store tracking history
                track = track_history[track_id]
                track.append((int((box[0] + box[2]) / 2), int((box[1] + box[3]) / 2)))
                if len(track) > 30:
                    track.pop(0)
    
                if len(track) >= 2:
                    # Calculate pixel distance between the last and second to last points
                    dx = track[-1][0] - track[-2][0]
                    dy = track[-1][1] - track[-2][1]
                    pixel_distance = (dx ** 2 + dy ** 2) ** 0.5
                    real_distance = pixel_distance * real_world_scale
                    speed = real_distance / time_per_frame  # speed in meters per second
                    label = f"{names[int(cls)]} {track_id} {speed:.2f} m/s"
                    annotator.box_label(box, color=colors(int(cls), True), label=label)
    
                    # Plot tracks
                    points = np.array(track, dtype=np.int32).reshape((-1, 1, 2))
                    cv2.circle(frame, (track[-1]), 7, colors(int(cls), True), -1)
                    cv2.polylines(frame, [points], isClosed=False, color=colors(int(cls), True), thickness=2)
                else:
                    label = f"{names[int(cls)]} {track_id}"
                    annotator.box_label(box, color=colors(int(cls), True), label=label)
    
        result.write(frame)
        if cv2.waitKey(1) & 0xFF == ord("q"):
            break
    else:
        break

    result.release() cap.release() cv2.destroyAllWindows()

    how do we benchmark the bytetrack perfomance per player ID?

  10. Zency-SunGitHub

    Hi, what should I do if I want to benchmark using only Pytorch, ONNX, openvino and tensorRT?

    1. Paula DerrengerUltralytics(editado)

      Hey there! 😊 To benchmark only PyTorch, ONNX, OpenVINO, and TensorRT, use the format argument of benchmark, which limits the run to one export format. Call it once per format ("-" is the PyTorch model):

      from ultralytics.utils.benchmarks import benchmark
      
      # "-" = PyTorch, then ONNX, OpenVINO (CPU only) and TensorRT (GPU only)
      for fmt, device in [("-", 0), ("onnx", 0), ("openvino", "cpu"), ("engine", 0)]:
          benchmark(model="yolo26n.pt", data="coco8.yaml", imgsz=640, device=device, format=fmt)

      And for CLI:

      yolo benchmark model=yolo26n.pt data=coco8.yaml imgsz=640 device=0 format=onnx
      yolo benchmark model=yolo26n.pt data=coco8.yaml imgsz=640 device=cpu format=openvino
      yolo benchmark model=yolo26n.pt data=coco8.yaml imgsz=640 device=0 format=engine

      Benchmark exports the model for you, so you don't need to export it first. Hope this helps! If there are more questions, feel free to ask.

      1
    2. Zency-SunGitHub

      when I use the under code to deploy: ' model = YOLO("yolov8s.engine") metrics = model.val(data="ultralytics/cfg/datasets/coco128.yaml",device='cuda:0') ' but I meet the error,What should I do?,Thanks your reply!!!

      Traceback (most recent call last): File "/media/zency/SSD-Data/Codes/yolov8_v8.1.0/ultralytics/nn/autobackend.py", line 220, in init import tensorrt as trt # noqa https://developer.nvidia.com/nvidia-tensorrt-download File "/home/zency/anaconda3/envs/yolov8plus/lib/python3.9/site-packages/tensorrt/init.py", line 68, in from .tensorrt import * ImportError: libnvinfer.so.8: cannot open shared object file: No such file or directory

      During handling of the above exception, another exception occurred:

      Traceback (most recent call last): File "/media/zency/SSD-Data/Codes/yolov8_v8.1.0/run_test.py", line 190, in yolo_mytest() File "/media/zency/SSD-Data/Codes/yolov8_v8.1.0/run_test.py", line 116, in yolo_mytest metrics = model.val(data="ultralytics/cfg/datasets/coco128.yaml",device='cuda:0') # cpu cuda:0 File "/media/zency/SSD-Data/Codes/yolov8_v8.1.0/ultralytics/engine/model.py", line 513, in val validator(model=self.model) File "/home/zency/anaconda3/envs/yolov8plus/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/media/zency/SSD-Data/Codes/yolov8_v8.1.0/ultralytics/engine/validator.py", line 124, in call model = AutoBackend( File "/home/zency/anaconda3/envs/yolov8plus/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/media/zency/SSD-Data/Codes/yolov8_v8.1.0/ultralytics/nn/autobackend.py", line 224, in init import tensorrt as trt # noqa File "/home/zency/anaconda3/envs/yolov8plus/lib/python3.9/site-packages/tensorrt/init.py", line 68, in from .tensorrt import * ImportError: libnvinfer.so.8: cannot open shared object file: No such file or directory Ultralytics YOLOv8.1.24 🚀 Python-3.9.18 torch-1.12.1 CUDA:0 (NVIDIA GeForce RTX 3090, 24268MiB) Loading yolov8s.engine for TensorRT inference...

    3. Zency-SunGitHub

      I solved the problem above, but there is a new problem:

      [TRT] [W] CUDA initialization failure with error: 35

    4. Paula DerrengerUltralytics

      Hey there! 🌟 It looks like you've encountered the CUDA initialization failure error with code 35, which usually indicates a CUDA driver/runtime version mismatch or an issue with accessing your GPU.

      Here are a couple of steps you can try to resolve the issue:

      1. Check CUDA Compatibility: Ensure that the installed CUDA version is compatible with the version required by TensorRT.
      2. Update Drivers: Sometimes, simply updating your NVIDIA drivers can resolve such issues.
      3. Environment Check: Make sure your environment variables (like PATH and LD_LIBRARY_PATH for Linux) are correctly set to include paths to CUDA toolkit and libraries.

      If these steps don't help, could you provide a bit more context about your setup? Specifically:

      • The versions of CUDA, TensorRT, and PyTorch you're using.
      • The output of nvidia-smi to check if the GPU is being recognized properly.

      This information can help pinpoint the issue more accurately. Thanks for reaching out, and looking forward to getting you past this hurdle! 🚀

    5. Paula DerrengerUltralytics(editado)

      @Zency-Sun i'm glad to hear you've resolved your previous issue! Regarding your question about model benchmarking, it's a great way to understand your model's performance across different formats. Here's a quick overview of what you might find useful:

      • Why Benchmark?: Helps you make informed decisions by understanding the trade-off between speed and accuracy across different formats and on varied hardware.
      • Key Metrics: Look at mAP50-95 for detection accuracy, and inference time for speed. This gives you a holistic view of how well your model performs.

      For benchmarking specific formats like PyTorch, ONNX, OpenVINO, and TensorRT, use the benchmark utility with the format argument. It exports the model for you:

      from ultralytics.utils.benchmarks import benchmark
      
      # "-" = PyTorch; OpenVINO runs on CPU, TensorRT on GPU
      for fmt, device in [("-", "cuda:0"), ("onnx", "cuda:0"), ("openvino", "cpu"), ("engine", "cuda:0")]:
          benchmark(model="yolo26n.pt", data="coco8.yaml", imgsz=640, device=device, format=fmt)

      And for CLI:

      yolo benchmark model=yolo26n.pt data=coco8.yaml imgsz=640 device=0 format=onnx
      # Repeat with format=openvino (device=cpu) and format=engine

      Remember, exporting to ONNX or OpenVINO can speed up CPU inference, and TensorRT is excellent for maximizing GPU efficiency. 🚀

      Let me know if there's anything else you need help with!