Reference for ultralytics/utils/benchmarks.py

def evaluate(self, yaml_path, val_log_file, eval_log_file, list_ind):
    """
    Evaluate model performance on validation results.

    Args:
        yaml_path (str): Path to the YAML configuration file.
        val_log_file (str): Path to the validation log file.
        eval_log_file (str): Path to the evaluation log file.
        list_ind (int): Index of the current dataset in the list.

    Returns:
        (float): The mean average precision (mAP) value for the evaluated model.

    Examples:
        Evaluate a model on a specific dataset
        >>> benchmark = RF100Benchmark()
        >>> benchmark.evaluate("path/to/data.yaml", "path/to/val_log.txt", "path/to/eval_log.txt", 0)
    """
    skip_symbols = ["🚀", "⚠️", "💡", "❌"]
    with open(yaml_path, encoding="utf-8") as stream:
        class_names = yaml.safe_load(stream)["names"]
    with open(val_log_file, encoding="utf-8") as f:
        lines = f.readlines()
        eval_lines = []
        for line in lines:
            if any(symbol in line for symbol in skip_symbols):
                continue
            entries = line.split(" ")
            entries = list(filter(lambda val: val != "", entries))
            entries = [e.strip("\n") for e in entries]
            eval_lines.extend(
                {
                    "class": entries[0],
                    "images": entries[1],
                    "targets": entries[2],
                    "precision": entries[3],
                    "recall": entries[4],
                    "map50": entries[5],
                    "map95": entries[6],
                }
                for e in entries
                if e in class_names or (e == "all" and "(AP)" not in entries and "(AR)" not in entries)
            )
    map_val = 0.0
    if len(eval_lines) > 1:
        print("There's more dicts")
        for lst in eval_lines:
            if lst["class"] == "all":
                map_val = lst["map50"]
    else:
        print("There's only one dict res")
        map_val = [res["map50"] for res in eval_lines][0]

    with open(eval_log_file, "a", encoding="utf-8") as f:
        f.write(f"{self.ds_names[list_ind]}: {map_val}\n")

@staticmethod
def fix_yaml(path):
    """Fix the train and validation paths in a given YAML file."""
    with open(path, encoding="utf-8") as file:
        yaml_data = yaml.safe_load(file)
    yaml_data["train"] = "train/images"
    yaml_data["val"] = "valid/images"
    with open(path, "w", encoding="utf-8") as file:
        yaml.safe_dump(yaml_data, file)

def parse_dataset(self, ds_link_txt="datasets_links.txt"):
    """
    Parse dataset links and download datasets.

    Args:
        ds_link_txt (str): Path to the file containing dataset links.

    Returns:
        ds_names (List[str]): List of dataset names.
        ds_cfg_list (List[Path]): List of paths to dataset configuration files.

    Examples:
        >>> benchmark = RF100Benchmark()
        >>> benchmark.set_key("api_key")
        >>> benchmark.parse_dataset("datasets_links.txt")
    """
    (shutil.rmtree("rf-100"), os.mkdir("rf-100")) if os.path.exists("rf-100") else os.mkdir("rf-100")
    os.chdir("rf-100")
    os.mkdir("ultralytics-benchmarks")
    safe_download("https://github.com/ultralytics/assets/releases/download/v0.0.0/datasets_links.txt")

    with open(ds_link_txt, encoding="utf-8") as file:
        for line in file:
            try:
                _, url, workspace, project, version = re.split("/+", line.strip())
                self.ds_names.append(project)
                proj_version = f"{project}-{version}"
                if not Path(proj_version).exists():
                    self.rf.workspace(workspace).project(project).version(version).download("yolov8")
                else:
                    print("Dataset already downloaded.")
                self.ds_cfg_list.append(Path.cwd() / proj_version / "data.yaml")
            except Exception:
                continue

    return self.ds_names, self.ds_cfg_list

def set_key(self, api_key):
    """
    Set Roboflow API key for processing.

    Args:
        api_key (str): The API key.

    Examples:
        Set the Roboflow API key for accessing datasets:
        >>> benchmark = RF100Benchmark()
        >>> benchmark.set_key("your_roboflow_api_key")
    """
    check_requirements("roboflow")
    from roboflow import Roboflow

    self.rf = Roboflow(api_key=api_key)

@staticmethod
def generate_results_dict(model_name, t_onnx, t_engine, model_info):
    """
    Generate a dictionary of profiling results.

    Args:
        model_name (str): Name of the model.
        t_onnx (tuple): ONNX model inference time statistics (mean, std).
        t_engine (tuple): TensorRT engine inference time statistics (mean, std).
        model_info (tuple): Model information (layers, params, gradients, flops).

    Returns:
        (dict): Dictionary containing profiling results.
    """
    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 generate_table_row(self, model_name, t_onnx, t_engine, model_info):
    """
    Generate a table row string with model performance metrics.

    Args:
        model_name (str): Name of the model.
        t_onnx (tuple): ONNX model inference time statistics (mean, std).
        t_engine (tuple): TensorRT engine inference time statistics (mean, std).
        model_info (tuple): Model information (layers, params, gradients, flops).

    Returns:
        (str): Formatted table row string with model metrics.
    """
    layers, params, gradients, flops = model_info
    return (
        f"| {model_name:18s} | {self.imgsz} | - | {t_onnx[0]:.1f}±{t_onnx[1]:.1f} ms | {t_engine[0]:.1f}±"
        f"{t_engine[1]:.1f} ms | {params / 1e6:.1f} | {flops:.1f} |"
    )

def get_files(self):
    """
    Return a list of paths for all relevant model files given by the user.

    Returns:
        (List[Path]): List of Path objects for the model files.
    """
    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)]

@staticmethod
def get_onnx_model_info(onnx_file: str):
    """Extracts metadata from an ONNX model file including parameters, GFLOPs, and input shape."""
    return 0.0, 0.0, 0.0, 0.0  # return (num_layers, num_params, num_gradients, num_flops)

@staticmethod
def iterative_sigma_clipping(data, sigma=2, max_iters=3):
    """
    Apply iterative sigma clipping to data to remove outliers.

    Args:
        data (numpy.ndarray): Input data array.
        sigma (float): Number of standard deviations to use for clipping.
        max_iters (int): Maximum number of iterations for the clipping process.

    Returns:
        (numpy.ndarray): Clipped data array with outliers removed.
    """
    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

@staticmethod
def print_table(table_rows):
    """
    Print a formatted table of model profiling results.

    Args:
        table_rows (List[str]): List of formatted table row strings.
    """
    gpu = torch.cuda.get_device_name(0) if torch.cuda.is_available() else "GPU"
    headers = [
        "Model",
        "size<br><sup>(pixels)",
        "mAP<sup>val<br>50-95",
        f"Speed<br><sup>CPU ({get_cpu_info()}) ONNX<br>(ms)",
        f"Speed<br><sup>{gpu} TensorRT<br>(ms)",
        "params<br><sup>(M)",
        "FLOPs<br><sup>(B)",
    ]
    header = "|" + "|".join(f" {h} " for h in headers) + "|"
    separator = "|" + "|".join("-" * (len(h) + 2) for h in headers) + "|"

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

def profile(self):
    """
    Profile YOLO models for speed and accuracy across various formats including ONNX and TensorRT.

    Returns:
        (List[Dict]): List of dictionaries containing profiling results for each model.

    Examples:
        Profile models and print results
        >>> from ultralytics.utils.benchmarks import ProfileModels
        >>> profiler = ProfileModels(["yolo11n.yaml", "yolov8s.yaml"])
        >>> results = profiler.profile()
    """
    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",
                imgsz=self.imgsz,
                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 profile_onnx_model(self, onnx_file: str, eps: float = 1e-3):
    """
    Profile an ONNX model, measuring average inference time and standard deviation across multiple runs.

    Args:
        onnx_file (str): Path to the ONNX model file.
        eps (float): Small epsilon value to prevent division by zero.

    Returns:
        mean_time (float): Mean inference time in milliseconds.
        std_time (float): Standard deviation of inference time in milliseconds.
    """
    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
    dynamic = not all(isinstance(dim, int) and dim >= 0 for dim in input_tensor.shape)  # dynamic input shape
    input_shape = (1, 3, self.imgsz, self.imgsz) if dynamic else input_tensor.shape

    # 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_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 profile_tensorrt_model(self, engine_file: str, eps: float = 1e-3):
    """
    Profile YOLO model performance with TensorRT, measuring average run time and standard deviation.

    Args:
        engine_file (str): Path to the TensorRT engine file.
        eps (float): Small epsilon value to prevent division by zero.

    Returns:
        mean_time (float): Mean inference time in milliseconds.
        std_time (float): Standard deviation of inference time in milliseconds.
    """
    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.zeros((self.imgsz, self.imgsz, 3), dtype=np.uint8)  # use uint8 for Classify

    # 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)