Zum Inhalt springen

Referenz fĂŒr ultralytics/utils/benchmarks.py

Hinweis

Diese Datei ist verfĂŒgbar unter https://github.com/ultralytics/ ultralytics/blob/main/ ultralytics/utils/benchmarks .py. Wenn du ein Problem entdeckst, hilf bitte mit, es zu beheben, indem du einen Pull Request đŸ› ïž einreichst. Vielen Dank 🙏!



ultralytics.utils.benchmarks.ProfileModels

ProfileModels Klasse fĂŒr die Profilierung verschiedener Modelle auf ONNX und TensorRT.

Diese Klasse erstellt ein Profil der Leistung verschiedener Modelle und liefert Ergebnisse wie die Modellgeschwindigkeit und die FLOPs.

Attribute:

Name Typ Beschreibung
paths list

Pfade der zu profilierenden Modelle.

num_timed_runs int

Anzahl der zeitlich begrenzten DurchlĂ€ufe fĂŒr die Profilerstellung. Der Standardwert ist 100.

num_warmup_runs int

Anzahl der AufwÀrmlÀufe vor der Profilerstellung. Standard ist 10.

min_time float

Mindestanzahl von Sekunden, fĂŒr die ein Profil erstellt werden soll. Der Standardwert ist 60.

imgsz int

BildgrĂ¶ĂŸe, die in den Modellen verwendet wird. Standard ist 640.

Methoden:

Name Beschreibung
profile

Profiliert die Modelle und druckt das Ergebnis aus.

Beispiel
from ultralytics.utils.benchmarks import ProfileModels

ProfileModels(['yolov8n.yaml', 'yolov8s.yaml'], imgsz=640).profile()
Quellcode in ultralytics/utils/benchmarks.py
class ProfileModels:
    """
    ProfileModels class for profiling different models on ONNX and TensorRT.

    This class profiles the performance of different models, returning results such as model speed and FLOPs.

    Attributes:
        paths (list): Paths of the models to profile.
        num_timed_runs (int): Number of timed runs for the profiling. Default is 100.
        num_warmup_runs (int): Number of warmup runs before profiling. Default is 10.
        min_time (float): Minimum number of seconds to profile for. Default is 60.
        imgsz (int): Image size used in the models. Default is 640.

    Methods:
        profile(): Profiles the models and prints the result.

    Example:
        ```python
        from ultralytics.utils.benchmarks import ProfileModels

        ProfileModels(['yolov8n.yaml', 'yolov8s.yaml'], imgsz=640).profile()
        ```
    """

    def __init__(
        self,
        paths: list,
        num_timed_runs=100,
        num_warmup_runs=10,
        min_time=60,
        imgsz=640,
        half=True,
        trt=True,
        device=None,
    ):
        """
        Initialize the ProfileModels class for profiling models.

        Args:
            paths (list): List of paths of the models to be profiled.
            num_timed_runs (int, optional): Number of timed runs for the profiling. Default is 100.
            num_warmup_runs (int, optional): Number of warmup runs before the actual profiling starts. Default is 10.
            min_time (float, optional): Minimum time in seconds for profiling a model. Default is 60.
            imgsz (int, optional): Size of the image used during profiling. Default is 640.
            half (bool, optional): Flag to indicate whether to use half-precision floating point for profiling.
            trt (bool, optional): Flag to indicate whether to profile using TensorRT. Default is True.
            device (torch.device, optional): Device used for profiling. If None, it is determined automatically.
        """
        self.paths = paths
        self.num_timed_runs = num_timed_runs
        self.num_warmup_runs = num_warmup_runs
        self.min_time = min_time
        self.imgsz = imgsz
        self.half = half
        self.trt = trt  # run TensorRT profiling
        self.device = device or torch.device(0 if torch.cuda.is_available() else "cpu")

    def profile(self):
        """Logs the benchmarking results of a model, checks metrics against floor and returns the results."""
        files = self.get_files()

        if not files:
            print("No matching *.pt or *.onnx files found.")
            return

        table_rows = []
        output = []
        for file in files:
            engine_file = file.with_suffix(".engine")
            if file.suffix in (".pt", ".yaml", ".yml"):
                model = YOLO(str(file))
                model.fuse()  # to report correct params and GFLOPs in model.info()
                model_info = model.info()
                if self.trt and self.device.type != "cpu" and not engine_file.is_file():
                    engine_file = model.export(
                        format="engine", half=self.half, imgsz=self.imgsz, device=self.device, verbose=False
                    )
                onnx_file = model.export(
                    format="onnx", half=self.half, imgsz=self.imgsz, simplify=True, device=self.device, verbose=False
                )
            elif file.suffix == ".onnx":
                model_info = self.get_onnx_model_info(file)
                onnx_file = file
            else:
                continue

            t_engine = self.profile_tensorrt_model(str(engine_file))
            t_onnx = self.profile_onnx_model(str(onnx_file))
            table_rows.append(self.generate_table_row(file.stem, t_onnx, t_engine, model_info))
            output.append(self.generate_results_dict(file.stem, t_onnx, t_engine, model_info))

        self.print_table(table_rows)
        return output

    def get_files(self):
        """Returns a list of paths for all relevant model files given by the user."""
        files = []
        for path in self.paths:
            path = Path(path)
            if path.is_dir():
                extensions = ["*.pt", "*.onnx", "*.yaml"]
                files.extend([file for ext in extensions for file in glob.glob(str(path / ext))])
            elif path.suffix in {".pt", ".yaml", ".yml"}:  # add non-existing
                files.append(str(path))
            else:
                files.extend(glob.glob(str(path)))

        print(f"Profiling: {sorted(files)}")
        return [Path(file) for file in sorted(files)]

    def get_onnx_model_info(self, onnx_file: str):
        """Retrieves the information including number of layers, parameters, gradients and FLOPs for an ONNX model
        file.
        """
        return 0.0, 0.0, 0.0, 0.0  # return (num_layers, num_params, num_gradients, num_flops)

    def iterative_sigma_clipping(self, data, sigma=2, max_iters=3):
        """Applies an iterative sigma clipping algorithm to the given data times number of iterations."""
        data = np.array(data)
        for _ in range(max_iters):
            mean, std = np.mean(data), np.std(data)
            clipped_data = data[(data > mean - sigma * std) & (data < mean + sigma * std)]
            if len(clipped_data) == len(data):
                break
            data = clipped_data
        return data

    def profile_tensorrt_model(self, engine_file: str, eps: float = 1e-3):
        """Profiles the TensorRT model, measuring average run time and standard deviation among runs."""
        if not self.trt or not Path(engine_file).is_file():
            return 0.0, 0.0

        # Model and input
        model = YOLO(engine_file)
        input_data = np.random.rand(self.imgsz, self.imgsz, 3).astype(np.float32)  # must be FP32

        # Warmup runs
        elapsed = 0.0
        for _ in range(3):
            start_time = time.time()
            for _ in range(self.num_warmup_runs):
                model(input_data, imgsz=self.imgsz, verbose=False)
            elapsed = time.time() - start_time

        # Compute number of runs as higher of min_time or num_timed_runs
        num_runs = max(round(self.min_time / (elapsed + eps) * self.num_warmup_runs), self.num_timed_runs * 50)

        # Timed runs
        run_times = []
        for _ in TQDM(range(num_runs), desc=engine_file):
            results = model(input_data, imgsz=self.imgsz, verbose=False)
            run_times.append(results[0].speed["inference"])  # Convert to milliseconds

        run_times = self.iterative_sigma_clipping(np.array(run_times), sigma=2, max_iters=3)  # sigma clipping
        return np.mean(run_times), np.std(run_times)

    def profile_onnx_model(self, onnx_file: str, eps: float = 1e-3):
        """Profiles an ONNX model by executing it multiple times and returns the mean and standard deviation of run
        times.
        """
        check_requirements("onnxruntime")
        import onnxruntime as ort

        # Session with either 'TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider'
        sess_options = ort.SessionOptions()
        sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
        sess_options.intra_op_num_threads = 8  # Limit the number of threads
        sess = ort.InferenceSession(onnx_file, sess_options, providers=["CPUExecutionProvider"])

        input_tensor = sess.get_inputs()[0]
        input_type = input_tensor.type

        # Mapping ONNX datatype to numpy datatype
        if "float16" in input_type:
            input_dtype = np.float16
        elif "float" in input_type:
            input_dtype = np.float32
        elif "double" in input_type:
            input_dtype = np.float64
        elif "int64" in input_type:
            input_dtype = np.int64
        elif "int32" in input_type:
            input_dtype = np.int32
        else:
            raise ValueError(f"Unsupported ONNX datatype {input_type}")

        input_data = np.random.rand(*input_tensor.shape).astype(input_dtype)
        input_name = input_tensor.name
        output_name = sess.get_outputs()[0].name

        # Warmup runs
        elapsed = 0.0
        for _ in range(3):
            start_time = time.time()
            for _ in range(self.num_warmup_runs):
                sess.run([output_name], {input_name: input_data})
            elapsed = time.time() - start_time

        # Compute number of runs as higher of min_time or num_timed_runs
        num_runs = max(round(self.min_time / (elapsed + eps) * self.num_warmup_runs), self.num_timed_runs)

        # Timed runs
        run_times = []
        for _ in TQDM(range(num_runs), desc=onnx_file):
            start_time = time.time()
            sess.run([output_name], {input_name: input_data})
            run_times.append((time.time() - start_time) * 1000)  # Convert to milliseconds

        run_times = self.iterative_sigma_clipping(np.array(run_times), sigma=2, max_iters=5)  # sigma clipping
        return np.mean(run_times), np.std(run_times)

    def generate_table_row(self, model_name, t_onnx, t_engine, model_info):
        """Generates a formatted string for a table row that includes model performance and metric details."""
        layers, params, gradients, flops = model_info
        return f"| {model_name:18s} | {self.imgsz} | - | {t_onnx[0]:.2f} ± {t_onnx[1]:.2f} ms | {t_engine[0]:.2f} ± {t_engine[1]:.2f} ms | {params / 1e6:.1f} | {flops:.1f} |"

    def generate_results_dict(self, model_name, t_onnx, t_engine, model_info):
        """Generates a dictionary of model details including name, parameters, GFLOPS and speed metrics."""
        layers, params, gradients, flops = model_info
        return {
            "model/name": model_name,
            "model/parameters": params,
            "model/GFLOPs": round(flops, 3),
            "model/speed_ONNX(ms)": round(t_onnx[0], 3),
            "model/speed_TensorRT(ms)": round(t_engine[0], 3),
        }

    def print_table(self, table_rows):
        """Formats and prints a comparison table for different models with given statistics and performance data."""
        gpu = torch.cuda.get_device_name(0) if torch.cuda.is_available() else "GPU"
        header = f"| Model | size<br><sup>(pixels) | mAP<sup>val<br>50-95 | Speed<br><sup>CPU ONNX<br>(ms) | Speed<br><sup>{gpu} TensorRT<br>(ms) | params<br><sup>(M) | FLOPs<br><sup>(B) |"
        separator = "|-------------|---------------------|--------------------|------------------------------|-----------------------------------|------------------|-----------------|"

        print(f"\n\n{header}")
        print(separator)
        for row in table_rows:
            print(row)

__init__(paths, num_timed_runs=100, num_warmup_runs=10, min_time=60, imgsz=640, half=True, trt=True, device=None)

Initialisiere die Klasse ProfileModels fĂŒr Profilierungsmodelle.

Parameter:

Name Typ Beschreibung Standard
paths list

Liste der Pfade der Modelle, die profiliert werden sollen.

erforderlich
num_timed_runs int

Anzahl der zeitlich begrenzten DurchlĂ€ufe fĂŒr die Profilerstellung. Der Standardwert ist 100.

100
num_warmup_runs int

Anzahl der AufwÀrmlÀufe, bevor die eigentliche Profilerstellung beginnt. Standard ist 10.

10
min_time float

Mindestzeit in Sekunden fĂŒr die Profilerstellung eines Modells. Der Standardwert ist 60.

60
imgsz int

GrĂ¶ĂŸe des Bildes, das bei der Profilerstellung verwendet wird. Standard ist 640.

640
half bool

Flag, das angibt, ob fĂŒr die Profilerstellung halbgenaue Fließkommazahlen verwendet werden sollen.

True
trt bool

Flagge, die angibt, ob das Profil mit TensorRT erstellt werden soll. Die Voreinstellung ist True.

True
device device

Das fĂŒr die Profilerstellung verwendete GerĂ€t. Wenn Keine, wird es automatisch ermittelt.

None
Quellcode in ultralytics/utils/benchmarks.py
def __init__(
    self,
    paths: list,
    num_timed_runs=100,
    num_warmup_runs=10,
    min_time=60,
    imgsz=640,
    half=True,
    trt=True,
    device=None,
):
    """
    Initialize the ProfileModels class for profiling models.

    Args:
        paths (list): List of paths of the models to be profiled.
        num_timed_runs (int, optional): Number of timed runs for the profiling. Default is 100.
        num_warmup_runs (int, optional): Number of warmup runs before the actual profiling starts. Default is 10.
        min_time (float, optional): Minimum time in seconds for profiling a model. Default is 60.
        imgsz (int, optional): Size of the image used during profiling. Default is 640.
        half (bool, optional): Flag to indicate whether to use half-precision floating point for profiling.
        trt (bool, optional): Flag to indicate whether to profile using TensorRT. Default is True.
        device (torch.device, optional): Device used for profiling. If None, it is determined automatically.
    """
    self.paths = paths
    self.num_timed_runs = num_timed_runs
    self.num_warmup_runs = num_warmup_runs
    self.min_time = min_time
    self.imgsz = imgsz
    self.half = half
    self.trt = trt  # run TensorRT profiling
    self.device = device or torch.device(0 if torch.cuda.is_available() else "cpu")

generate_results_dict(model_name, t_onnx, t_engine, model_info)

Erzeugt ein Wörterbuch mit Modelldetails wie Name, Parameter, GFLOPS und Geschwindigkeitsmetriken.

Quellcode in ultralytics/utils/benchmarks.py
def generate_results_dict(self, model_name, t_onnx, t_engine, model_info):
    """Generates a dictionary of model details including name, parameters, GFLOPS and speed metrics."""
    layers, params, gradients, flops = model_info
    return {
        "model/name": model_name,
        "model/parameters": params,
        "model/GFLOPs": round(flops, 3),
        "model/speed_ONNX(ms)": round(t_onnx[0], 3),
        "model/speed_TensorRT(ms)": round(t_engine[0], 3),
    }

generate_table_row(model_name, t_onnx, t_engine, model_info)

Erzeugt eine formatierte Zeichenkette fĂŒr eine Tabellenzeile, die Details zur Modellleistung und Metrik enthĂ€lt.

Quellcode in ultralytics/utils/benchmarks.py
def generate_table_row(self, model_name, t_onnx, t_engine, model_info):
    """Generates a formatted string for a table row that includes model performance and metric details."""
    layers, params, gradients, flops = model_info
    return f"| {model_name:18s} | {self.imgsz} | - | {t_onnx[0]:.2f} ± {t_onnx[1]:.2f} ms | {t_engine[0]:.2f} ± {t_engine[1]:.2f} ms | {params / 1e6:.1f} | {flops:.1f} |"

get_files()

Gibt eine Liste der Pfade fĂŒr alle relevanten Modelldateien zurĂŒck, die der Benutzer angegeben hat.

Quellcode in ultralytics/utils/benchmarks.py
def get_files(self):
    """Returns a list of paths for all relevant model files given by the user."""
    files = []
    for path in self.paths:
        path = Path(path)
        if path.is_dir():
            extensions = ["*.pt", "*.onnx", "*.yaml"]
            files.extend([file for ext in extensions for file in glob.glob(str(path / ext))])
        elif path.suffix in {".pt", ".yaml", ".yml"}:  # add non-existing
            files.append(str(path))
        else:
            files.extend(glob.glob(str(path)))

    print(f"Profiling: {sorted(files)}")
    return [Path(file) for file in sorted(files)]

get_onnx_model_info(onnx_file)

Ruft die Informationen ab, einschließlich der Anzahl der Schichten, Parameter, Gradienten und FLOPs fĂŒr eine ONNX Modell Datei.

Quellcode in ultralytics/utils/benchmarks.py
def get_onnx_model_info(self, onnx_file: str):
    """Retrieves the information including number of layers, parameters, gradients and FLOPs for an ONNX model
    file.
    """
    return 0.0, 0.0, 0.0, 0.0  # return (num_layers, num_params, num_gradients, num_flops)

iterative_sigma_clipping(data, sigma=2, max_iters=3)

Wendet einen iterativen Sigma-Clipping-Algorithmus auf die angegebenen Daten mal Anzahl der Iterationen an.

Quellcode in ultralytics/utils/benchmarks.py
def iterative_sigma_clipping(self, data, sigma=2, max_iters=3):
    """Applies an iterative sigma clipping algorithm to the given data times number of iterations."""
    data = np.array(data)
    for _ in range(max_iters):
        mean, std = np.mean(data), np.std(data)
        clipped_data = data[(data > mean - sigma * std) & (data < mean + sigma * std)]
        if len(clipped_data) == len(data):
            break
        data = clipped_data
    return data

print_table(table_rows)

Formatiert und druckt eine Vergleichstabelle fĂŒr verschiedene Modelle mit vorgegebenen Statistiken und Leistungsdaten.

Quellcode in ultralytics/utils/benchmarks.py
def print_table(self, table_rows):
    """Formats and prints a comparison table for different models with given statistics and performance data."""
    gpu = torch.cuda.get_device_name(0) if torch.cuda.is_available() else "GPU"
    header = f"| Model | size<br><sup>(pixels) | mAP<sup>val<br>50-95 | Speed<br><sup>CPU ONNX<br>(ms) | Speed<br><sup>{gpu} TensorRT<br>(ms) | params<br><sup>(M) | FLOPs<br><sup>(B) |"
    separator = "|-------------|---------------------|--------------------|------------------------------|-----------------------------------|------------------|-----------------|"

    print(f"\n\n{header}")
    print(separator)
    for row in table_rows:
        print(row)

profile()

Protokolliert die Benchmarking-Ergebnisse eines Modells, prĂŒft die Metriken gegen den Boden und gibt die Ergebnisse zurĂŒck.

Quellcode in ultralytics/utils/benchmarks.py
def profile(self):
    """Logs the benchmarking results of a model, checks metrics against floor and returns the results."""
    files = self.get_files()

    if not files:
        print("No matching *.pt or *.onnx files found.")
        return

    table_rows = []
    output = []
    for file in files:
        engine_file = file.with_suffix(".engine")
        if file.suffix in (".pt", ".yaml", ".yml"):
            model = YOLO(str(file))
            model.fuse()  # to report correct params and GFLOPs in model.info()
            model_info = model.info()
            if self.trt and self.device.type != "cpu" and not engine_file.is_file():
                engine_file = model.export(
                    format="engine", half=self.half, imgsz=self.imgsz, device=self.device, verbose=False
                )
            onnx_file = model.export(
                format="onnx", half=self.half, imgsz=self.imgsz, simplify=True, device=self.device, verbose=False
            )
        elif file.suffix == ".onnx":
            model_info = self.get_onnx_model_info(file)
            onnx_file = file
        else:
            continue

        t_engine = self.profile_tensorrt_model(str(engine_file))
        t_onnx = self.profile_onnx_model(str(onnx_file))
        table_rows.append(self.generate_table_row(file.stem, t_onnx, t_engine, model_info))
        output.append(self.generate_results_dict(file.stem, t_onnx, t_engine, model_info))

    self.print_table(table_rows)
    return output

profile_onnx_model(onnx_file, eps=0.001)

Profiliert ein ONNX Modell, indem es mehrfach ausgefĂŒhrt wird, und gibt den Mittelwert und die Standardabweichung der Lauf Zeiten zurĂŒck.

Quellcode in ultralytics/utils/benchmarks.py
def profile_onnx_model(self, onnx_file: str, eps: float = 1e-3):
    """Profiles an ONNX model by executing it multiple times and returns the mean and standard deviation of run
    times.
    """
    check_requirements("onnxruntime")
    import onnxruntime as ort

    # Session with either 'TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider'
    sess_options = ort.SessionOptions()
    sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
    sess_options.intra_op_num_threads = 8  # Limit the number of threads
    sess = ort.InferenceSession(onnx_file, sess_options, providers=["CPUExecutionProvider"])

    input_tensor = sess.get_inputs()[0]
    input_type = input_tensor.type

    # Mapping ONNX datatype to numpy datatype
    if "float16" in input_type:
        input_dtype = np.float16
    elif "float" in input_type:
        input_dtype = np.float32
    elif "double" in input_type:
        input_dtype = np.float64
    elif "int64" in input_type:
        input_dtype = np.int64
    elif "int32" in input_type:
        input_dtype = np.int32
    else:
        raise ValueError(f"Unsupported ONNX datatype {input_type}")

    input_data = np.random.rand(*input_tensor.shape).astype(input_dtype)
    input_name = input_tensor.name
    output_name = sess.get_outputs()[0].name

    # Warmup runs
    elapsed = 0.0
    for _ in range(3):
        start_time = time.time()
        for _ in range(self.num_warmup_runs):
            sess.run([output_name], {input_name: input_data})
        elapsed = time.time() - start_time

    # Compute number of runs as higher of min_time or num_timed_runs
    num_runs = max(round(self.min_time / (elapsed + eps) * self.num_warmup_runs), self.num_timed_runs)

    # Timed runs
    run_times = []
    for _ in TQDM(range(num_runs), desc=onnx_file):
        start_time = time.time()
        sess.run([output_name], {input_name: input_data})
        run_times.append((time.time() - start_time) * 1000)  # Convert to milliseconds

    run_times = self.iterative_sigma_clipping(np.array(run_times), sigma=2, max_iters=5)  # sigma clipping
    return np.mean(run_times), np.std(run_times)

profile_tensorrt_model(engine_file, eps=0.001)

Erstelle ein Profil des Modells TensorRT und messe die durchschnittliche Laufzeit und die Standardabweichung zwischen den LĂ€ufen.

Quellcode in ultralytics/utils/benchmarks.py
def profile_tensorrt_model(self, engine_file: str, eps: float = 1e-3):
    """Profiles the TensorRT model, measuring average run time and standard deviation among runs."""
    if not self.trt or not Path(engine_file).is_file():
        return 0.0, 0.0

    # Model and input
    model = YOLO(engine_file)
    input_data = np.random.rand(self.imgsz, self.imgsz, 3).astype(np.float32)  # must be FP32

    # Warmup runs
    elapsed = 0.0
    for _ in range(3):
        start_time = time.time()
        for _ in range(self.num_warmup_runs):
            model(input_data, imgsz=self.imgsz, verbose=False)
        elapsed = time.time() - start_time

    # Compute number of runs as higher of min_time or num_timed_runs
    num_runs = max(round(self.min_time / (elapsed + eps) * self.num_warmup_runs), self.num_timed_runs * 50)

    # Timed runs
    run_times = []
    for _ in TQDM(range(num_runs), desc=engine_file):
        results = model(input_data, imgsz=self.imgsz, verbose=False)
        run_times.append(results[0].speed["inference"])  # Convert to milliseconds

    run_times = self.iterative_sigma_clipping(np.array(run_times), sigma=2, max_iters=3)  # sigma clipping
    return np.mean(run_times), np.std(run_times)



ultralytics.utils.benchmarks.benchmark(model=WEIGHTS_DIR / 'yolov8n.pt', data=None, imgsz=160, half=False, int8=False, device='cpu', verbose=False)

Vergleiche ein YOLO Modell in verschiedenen Formaten auf Geschwindigkeit und Genauigkeit.

Parameter:

Name Typ Beschreibung Standard
model str | Path | optional

Pfad zu der Modelldatei oder dem Verzeichnis. Standard ist Path(SETTINGS['weights_dir']) / 'yolov8n.pt'.

WEIGHTS_DIR / 'yolov8n.pt'
data str

Datensatz, der ausgewertet werden soll. Wird von TASK2DATA geerbt, wenn er nicht ĂŒbergeben wird. Der Standardwert ist None.

None
imgsz int

BildgrĂ¶ĂŸe fĂŒr den Benchmark. Standard ist 160.

160
half bool

Verwende die halbe Genauigkeit fĂŒr das Modell, wenn True. Standard ist False.

False
int8 bool

Verwende int8-PrĂ€zision fĂŒr das Modell, wenn True. Standard ist False.

False
device str

GerĂ€t, auf dem der Benchmark ausgefĂŒhrt werden soll, entweder "cpu" oder "cuda". Standard ist "cpu".

'cpu'
verbose bool | float | optional

Wenn True oder eine Fließkommazahl, werden die Benchmarks mit der angegebenen Metrik bestĂ€tigt. Standard ist False.

False

Retouren:

Name Typ Beschreibung
df DataFrame

Ein Pandas DataFrame mit Benchmark-Ergebnissen fĂŒr jedes Format, einschließlich DateigrĂ¶ĂŸe, Metrik und Inferenzzeit.

Beispiel
from ultralytics.utils.benchmarks import benchmark

benchmark(model='yolov8n.pt', imgsz=640)
Quellcode in ultralytics/utils/benchmarks.py
def benchmark(
    model=WEIGHTS_DIR / "yolov8n.pt", data=None, imgsz=160, half=False, int8=False, device="cpu", verbose=False
):
    """
    Benchmark a YOLO model across different formats for speed and accuracy.

    Args:
        model (str | Path | optional): Path to the model file or directory. Default is
            Path(SETTINGS['weights_dir']) / 'yolov8n.pt'.
        data (str, optional): Dataset to evaluate on, inherited from TASK2DATA if not passed. Default is None.
        imgsz (int, optional): Image size for the benchmark. Default is 160.
        half (bool, optional): Use half-precision for the model if True. Default is False.
        int8 (bool, optional): Use int8-precision for the model if True. Default is False.
        device (str, optional): Device to run the benchmark on, either 'cpu' or 'cuda'. Default is 'cpu'.
        verbose (bool | float | optional): If True or a float, assert benchmarks pass with given metric.
            Default is False.

    Returns:
        df (pandas.DataFrame): A pandas DataFrame with benchmark results for each format, including file size,
            metric, and inference time.

    Example:
        ```python
        from ultralytics.utils.benchmarks import benchmark

        benchmark(model='yolov8n.pt', imgsz=640)
        ```
    """

    import pandas as pd

    pd.options.display.max_columns = 10
    pd.options.display.width = 120
    device = select_device(device, verbose=False)
    if isinstance(model, (str, Path)):
        model = YOLO(model)

    y = []
    t0 = time.time()
    for i, (name, format, suffix, cpu, gpu) in export_formats().iterrows():  # index, (name, format, suffix, CPU, GPU)
        emoji, filename = "❌", None  # export defaults
        try:
            # Checks
            if i == 9:
                assert LINUX, "Edge TPU export only supported on Linux"
            elif i == 7:
                assert model.task != "obb", "TensorFlow GraphDef not supported for OBB task"
            elif i in {5, 10}:  # CoreML and TF.js
                assert MACOS or LINUX, "export only supported on macOS and Linux"
            if i in {3, 5}:  # CoreML and OpenVINO
                assert not IS_PYTHON_3_12, "CoreML and OpenVINO not supported on Python 3.12"
            if "cpu" in device.type:
                assert cpu, "inference not supported on CPU"
            if "cuda" in device.type:
                assert gpu, "inference not supported on GPU"

            # Export
            if format == "-":
                filename = model.ckpt_path or model.cfg
                exported_model = model  # PyTorch format
            else:
                filename = model.export(imgsz=imgsz, format=format, half=half, int8=int8, device=device, verbose=False)
                exported_model = YOLO(filename, task=model.task)
                assert suffix in str(filename), "export failed"
            emoji = "❎"  # indicates export succeeded

            # Predict
            assert model.task != "pose" or i != 7, "GraphDef Pose inference is not supported"
            assert i not in (9, 10), "inference not supported"  # Edge TPU and TF.js are unsupported
            assert i != 5 or platform.system() == "Darwin", "inference only supported on macOS>=10.13"  # CoreML
            exported_model.predict(ASSETS / "bus.jpg", imgsz=imgsz, device=device, half=half)

            # Validate
            data = data or TASK2DATA[model.task]  # task to dataset, i.e. coco8.yaml for task=detect
            key = TASK2METRIC[model.task]  # task to metric, i.e. metrics/mAP50-95(B) for task=detect
            results = exported_model.val(
                data=data, batch=1, imgsz=imgsz, plots=False, device=device, half=half, int8=int8, verbose=False
            )
            metric, speed = results.results_dict[key], results.speed["inference"]
            y.append([name, "✅", round(file_size(filename), 1), round(metric, 4), round(speed, 2)])
        except Exception as e:
            if verbose:
                assert type(e) is AssertionError, f"Benchmark failure for {name}: {e}"
            LOGGER.warning(f"ERROR ❌ Benchmark failure for {name}: {e}")
            y.append([name, emoji, round(file_size(filename), 1), None, None])  # mAP, t_inference

    # Print results
    check_yolo(device=device)  # print system info
    df = pd.DataFrame(y, columns=["Format", "Status❔", "Size (MB)", key, "Inference time (ms/im)"])

    name = Path(model.ckpt_path).name
    s = f"\nBenchmarks complete for {name} on {data} at imgsz={imgsz} ({time.time() - t0:.2f}s)\n{df}\n"
    LOGGER.info(s)
    with open("benchmarks.log", "a", errors="ignore", encoding="utf-8") as f:
        f.write(s)

    if verbose and isinstance(verbose, float):
        metrics = df[key].array  # values to compare to floor
        floor = verbose  # minimum metric floor to pass, i.e. = 0.29 mAP for YOLOv5n
        assert all(x > floor for x in metrics if pd.notna(x)), f"Benchmark failure: metric(s) < floor {floor}"

    return df





Erstellt am 2023-11-12, Aktualisiert am 2023-11-25
Autoren: glenn-jocher (3)