# Ultralytics YOLO π, AGPL-3.0 license
"""
Benchmark a YOLO model formats for speed and accuracy.
Usage:
from ultralytics.utils.benchmarks import ProfileModels, benchmark
ProfileModels(['yolov8n.yaml', 'yolov8s.yaml']).profile()
benchmark(model='yolov8n.pt', imgsz=160)
Format | `format=argument` | Model
--- | --- | ---
PyTorch | - | yolov8n.pt
TorchScript | `torchscript` | yolov8n.torchscript
ONNX | `onnx` | yolov8n.onnx
OpenVINO | `openvino` | yolov8n_openvino_model/
TensorRT | `engine` | yolov8n.engine
CoreML | `coreml` | yolov8n.mlpackage
TensorFlow SavedModel | `saved_model` | yolov8n_saved_model/
TensorFlow GraphDef | `pb` | yolov8n.pb
TensorFlow Lite | `tflite` | yolov8n.tflite
TensorFlow Edge TPU | `edgetpu` | yolov8n_edgetpu.tflite
TensorFlow.js | `tfjs` | yolov8n_web_model/
PaddlePaddle | `paddle` | yolov8n_paddle_model/
NCNN | `ncnn` | yolov8n_ncnn_model/
"""
import glob
import os
import platform
import re
import shutil
import time
from pathlib import Path
import numpy as np
import torch.cuda
import yaml
from ultralytics import YOLO, YOLOWorld
from ultralytics.cfg import TASK2DATA, TASK2METRIC
from ultralytics.engine.exporter import export_formats
from ultralytics.utils import ARM64, ASSETS, IS_JETSON, IS_RASPBERRYPI, LINUX, LOGGER, MACOS, TQDM, WEIGHTS_DIR
from ultralytics.utils.checks import IS_PYTHON_3_12, check_requirements, check_yolo
from ultralytics.utils.downloads import safe_download
from ultralytics.utils.files import file_size
from ultralytics.utils.torch_utils import get_cpu_info, select_device
def benchmark(
model=WEIGHTS_DIR / "yolo11n.pt",
data=None,
imgsz=160,
half=False,
int8=False,
device="cpu",
verbose=False,
eps=1e-3,
):
"""
Benchmark a YOLO model across different formats for speed and accuracy.
Args:
model (str | Path): Path to the model file or directory.
data (str | None): Dataset to evaluate on, inherited from TASK2DATA if not passed.
imgsz (int): Image size for the benchmark.
half (bool): Use half-precision for the model if True.
int8 (bool): Use int8-precision for the model if True.
device (str): Device to run the benchmark on, either 'cpu' or 'cuda'.
verbose (bool | float): If True or a float, assert benchmarks pass with given metric.
eps (float): Epsilon value for divide by zero prevention.
Returns:
(pandas.DataFrame): A pandas DataFrame with benchmark results for each format, including file size, metric,
and inference time.
Examples:
Benchmark a YOLO model with default settings:
>>> from ultralytics.utils.benchmarks import benchmark
>>> benchmark(model="yolo11n.pt", imgsz=640)
"""
import pandas as pd # scope for faster 'import ultralytics'
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)
is_end2end = getattr(model.model.model[-1], "end2end", False)
y = []
t0 = time.time()
for i, (name, format, suffix, cpu, gpu) in enumerate(zip(*export_formats().values())):
emoji, filename = "β", None # export defaults
try:
# Checks
if i == 7: # TF GraphDef
assert model.task != "obb", "TensorFlow GraphDef not supported for OBB task"
elif i == 9: # Edge TPU
assert LINUX and not ARM64, "Edge TPU export only supported on non-aarch64 Linux"
elif i in {5, 10}: # CoreML and TF.js
assert MACOS or LINUX, "CoreML and TF.js export only supported on macOS and Linux"
assert not IS_RASPBERRYPI, "CoreML and TF.js export not supported on Raspberry Pi"
assert not IS_JETSON, "CoreML and TF.js export not supported on NVIDIA Jetson"
if i in {5}: # CoreML
assert not IS_PYTHON_3_12, "CoreML not supported on Python 3.12"
if i in {6, 7, 8}: # TF SavedModel, TF GraphDef, and TFLite
assert not isinstance(model, YOLOWorld), "YOLOWorldv2 TensorFlow exports not supported by onnx2tf yet"
if i in {9, 10}: # TF EdgeTPU and TF.js
assert not isinstance(model, YOLOWorld), "YOLOWorldv2 TensorFlow exports not supported by onnx2tf yet"
if i in {11}: # Paddle
assert not isinstance(model, YOLOWorld), "YOLOWorldv2 Paddle exports not supported yet"
assert not is_end2end, "End-to-end models not supported by PaddlePaddle yet"
assert LINUX or MACOS, "Windows Paddle exports not supported yet"
if i in {12}: # NCNN
assert not isinstance(model, YOLOWorld), "YOLOWorldv2 NCNN exports not supported yet"
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 = "β" # 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
if i in {12}:
assert not is_end2end, "End-to-end torch.topk operation is not supported for NCNN prediction yet"
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"]
fps = round(1000 / (speed + eps), 2) # frames per second
y.append([name, "β
", round(file_size(filename), 1), round(metric, 4), round(speed, 2), fps])
except Exception as e:
if verbose:
assert type(e) is AssertionError, f"Benchmark failure for {name}: {e}"
LOGGER.warning(f"ERROR βοΈ Benchmark failure for {name}: {e}")
y.append([name, emoji, round(file_size(filename), 1), None, None, None]) # mAP, t_inference
# Print results
check_yolo(device=device) # print system info
df = pd.DataFrame(y, columns=["Format", "Statusβ", "Size (MB)", key, "Inference time (ms/im)", "FPS"])
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
class RF100Benchmark:
"""Benchmark YOLO model performance across various formats for speed and accuracy."""
def __init__(self):
"""Initialize the RF100Benchmark class for benchmarking YOLO model performance across various formats."""
self.ds_names = []
self.ds_cfg_list = []
self.rf = None
self.val_metrics = ["class", "images", "targets", "precision", "recall", "map50", "map95"]
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)
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.
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) 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
@staticmethod
def fix_yaml(path):
"""
Fixes the train and validation paths in a given YAML file.
Args:
path (str): Path to the YAML file to be fixed.
Examples:
>>> RF100Benchmark.fix_yaml("path/to/data.yaml")
"""
with open(path) as file:
yaml_data = yaml.safe_load(file)
yaml_data["train"] = "train/images"
yaml_data["val"] = "valid/images"
with open(path, "w") as file:
yaml.safe_dump(yaml_data, file)
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 = ["π", "β οΈ", "π‘", "β"]
with open(yaml_path) 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") as f:
f.write(f"{self.ds_names[list_ind]}: {map_val}\n")
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[str]): Paths of the models to profile.
num_timed_runs (int): Number of timed runs for the profiling.
num_warmup_runs (int): Number of warmup runs before profiling.
min_time (float): Minimum number of seconds to profile for.
imgsz (int): Image size used in the models.
half (bool): Flag to indicate whether to use FP16 half-precision for TensorRT profiling.
trt (bool): Flag to indicate whether to profile using TensorRT.
device (torch.device): Device used for profiling.
Methods:
profile: Profiles the models and prints the result.
Examples:
Profile models and print results
>>> from ultralytics.utils.benchmarks import ProfileModels
>>> profiler = ProfileModels(["yolov8n.yaml", "yolov8s.yaml"], imgsz=640)
>>> profiler.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[str]): List of paths of the models to be profiled.
num_timed_runs (int): Number of timed runs for the profiling.
num_warmup_runs (int): Number of warmup runs before the actual profiling starts.
min_time (float): Minimum time in seconds for profiling a model.
imgsz (int): Size of the image used during profiling.
half (bool): Flag to indicate whether to use FP16 half-precision for TensorRT profiling.
trt (bool): Flag to indicate whether to profile using TensorRT.
device (torch.device | None): Device used for profiling. If None, it is determined automatically.
Notes:
FP16 'half' argument option removed for ONNX as slower on CPU than FP32.
Examples:
Initialize and profile models
>>> from ultralytics.utils.benchmarks import ProfileModels
>>> profiler = ProfileModels(["yolov8n.yaml", "yolov8s.yaml"], imgsz=640)
>>> profiler.profile()
"""
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):
"""Profiles YOLO models for speed and accuracy across various formats including ONNX and TensorRT."""
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 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):
"""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):
"""Applies iterative sigma clipping to data to remove outliers based on specified sigma and iteration count."""
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 YOLO model performance with TensorRT, measuring average run time and standard deviation."""
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, measuring average inference time and standard deviation across multiple runs."""
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 generate_table_row(self, model_name, t_onnx, t_engine, model_info):
"""Generates a table row string with model performance metrics including inference times and model details."""
layers, params, gradients, flops = model_info
return (
f"| {model_name:18s} | {self.imgsz} | - | {t_onnx[0]:.1f}Β±{t_onnx[1]:.1f} ms | {t_engine[0]:.1f}Β±"
f"{t_engine[1]:.1f} ms | {params / 1e6:.1f} | {flops:.1f} |"
)
@staticmethod
def generate_results_dict(model_name, t_onnx, t_engine, model_info):
"""Generates a dictionary of profiling results including model 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),
}
@staticmethod
def print_table(table_rows):
"""Prints a formatted table of model profiling results, including speed and accuracy metrics."""
gpu = torch.cuda.get_device_name(0) if torch.cuda.is_available() else "GPU"
headers = [
"Model",
"size
(pixels)",
"mAPval
50-95",
f"Speed
CPU ({get_cpu_info()}) ONNX
(ms)",
f"Speed
{gpu} TensorRT
(ms)",
"params
(M)",
"FLOPs
(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)