§
    l¬¸jÒq  ã                  ó6  — d Z ddlmZ ddlZddlZddlZddlZddlmZ ddl	m
Z
 ddlZddlZddlmZmZmZ ddlmZmZ ddlmZ dd	lmZ dd
lmZmZmZmZmZm Z m!Z!m"Z"m#Z#m$Z$ ddl%m&Z&m'Z'm(Z(m)Z) ddl*m+Z+ ddl,m-Z-m.Z. e#dz  dddddddfd„Z/ G d„ d¦  «        Z0dS )aì  
Benchmark YOLO model formats for speed and accuracy.

Usage:
    from ultralytics.utils.benchmarks import ProfileModels, benchmark
    ProfileModels(['yolo26n.yaml', 'yolov8s.yaml']).run()
    benchmark(model='yolo26n.pt', imgsz=160)

Format                  | `format=argument`         | Model
---                     | ---                       | ---
PyTorch                 | -                         | yolo26n.pt
TorchScript             | `torchscript`             | yolo26n.torchscript
ONNX                    | `onnx`                    | yolo26n.onnx
OpenVINO                | `openvino`                | yolo26n_openvino_model/
TensorRT                | `engine`                  | yolo26n.engine
CoreML                  | `coreml`                  | yolo26n.mlpackage
TensorFlow SavedModel   | `saved_model`             | yolo26n_saved_model/
TensorFlow GraphDef     | `pb`                      | yolo26n.pb
TensorFlow Edge TPU     | `edgetpu`                 | yolo26n_edgetpu.tflite
PaddlePaddle            | `paddle`                  | yolo26n_paddle_model/
MNN                     | `mnn`                     | yolo26n.mnn
NCNN                    | `ncnn`                    | yolo26n_ncnn_model/
IMX                     | `imx`                     | yolo26n_imx_model/
RKNN                    | `rknn`                    | yolo26n_rknn_model/
ExecuTorch              | `executorch`              | yolo26n_executorch_model/
Axelera AI              | `axelera`                 | yolo26n_axelera_model/
é    )ÚannotationsN)Údeepcopy©ÚPath)ÚRTDETRÚYOLOÚ	YOLOWorld)Ú	TASK2DATAÚTASK2METRIC)Úexport_formats)Ú	Segment26)
ÚARM64ÚASSETSÚ	IS_DOCKERÚ	IS_JETSONÚLINUXÚLOGGERÚMACOSÚTQDMÚWEIGHTS_DIRÚis_github_action_running)ÚIS_PYTHON_MINIMUM_3_13Úcheck_imgszÚ
check_yoloÚis_rockchip)Ú	file_size)Úget_cpu_infoÚselect_devicez
yolo26n.pté    ÚcpuFçü©ñÒMbP?Ú c                ó„  ‡'— ddl m}	  |	|¦  «        }t          |¦  «        }t          |t          ¦  «        r|d         |d         k    snn
J d¦   «         ‚ddl}
|
j                             d¦  «         |
j                             d¦  «         |
j         	                    d¦  «         |
j         
                    d¦  «         |
j                             d¦  «         |
j                             d¦  «         t          |d	¬
¦  «        }t          | t          t          f¦  «        rt!          | ¦  «        } |pt"          | j                 }t&          | j                 }g }t)          j        ¦   «         }|                     ¦   «         }|r6t-          t/          ¦   «         d         ¦  «        }||v sJ d|› d|› d�¦   «         ‚t1          t/          ¦   «                              ¦   «         Ž D �]ä\  }}}}}}}d\  }}	 |r||k    rŒt4          r|s|dv rŒ(|dk    r| j        dk    s
J d¦   «         ‚nz|dk    r7t6          rt8          r
J d¦   «         ‚t;          j        d¦  «        s
J d¦   «         ‚n=|dk    r7t>          st6          rt8          r
J d¦   «         ‚t>          rt4          r
J d¦   «         ‚|dv rt          | t@          ¦  «        r
J d¦   «         ‚|dk    rOt          | t@          ¦  «        r
J d¦   «         ‚t6          rtB          rt>          s
J d¦   «         ‚t4          r
J d¦   «         ‚|d k    r0t          | t@          ¦  «        r
J d!¦   «         ‚t4          r
J d"¦   «         ‚|d#k    rt          | t@          ¦  «        r
J d$¦   «         ‚|d%k    rRt          | t@          ¦  «        r
J d&¦   «         ‚| j        d'v s
J d(¦   «         ‚d)|  "                    ¦   «         v s
J d*¦   «         ‚|d+k    rHt          | t@          ¦  «        r
J d,¦   «         ‚t6          s
J d-¦   «         ‚tG          ¦   «         r
J d.¦   «         ‚|d/k    rt          | t@          ¦  «        r
J d0¦   «         ‚|d1k    rƒt          | t@          ¦  «        r
J d2¦   «         ‚t6          rt8          rtH          r
J d3¦   «         ‚| j        d4k    r:tK          d5„ | j&         '                    ¦   «         D ¦   «         ¦  «        r
J d6¦   «         ‚|d7k    r>t>          st6          rt8          r
J d8¦   «         ‚t>          rtQ          ¦   «         r
J d9¦   «         ‚d:|j)        v r|s
J d;¦   «         ‚d<|j)        v r|s
J d=¦   «         ‚|d>k    r%| j*        p| j+        p| j,        }t[          | ¦  «        }nqd?|v r|nd} | j.        dq|||||d	d@œ|¤Ž}t          | t^          ¦  «        rt_          |¦  «        nt!          || j        ¬A¦  «        }|t          |¦  «        v s
J dB¦   «         ‚dC}| j        dDk    s|dk    s
J dE¦   «         ‚|dk    s
J dF¦   «         ‚|dk    s!ta          j1        ¦   «         dGk    s
J dH¦   «         ‚|dIk    st>          s
J dJ¦   «         ‚|d1k    s
J dK¦   «         ‚| 2                    tf          dLz  |||| 4                    dM¦  «        d	¬N¦  «         | 5                    |d|d	d	||| 4                    dM¦  «        d	dO¬P¦
  «
        }|j6        |         |j7        dQ         }}tq          dR||z   z  dS¦  «        }| 9                    |dTtq          tu          |¦  «        d¦  «        tq          |dU¦  «        tq          |dS¦  «        |g¦  «         �ŒV# tv          $ rƒ}|r&tS          |¦  «        tx          u sJ dV|› dW|› �¦   «         ‚t{          j>        dV|› dW|› �¦  «         | 9                    ||tq          tu          |¦  «        d¦  «        dddg¦  «         Y d}~�ŒÞd}~ww xY wt          |¬X¦  «         |
 @                    |dYdZd[|d\d]gd^¬_¦  «        } |  A                    d`d¬a¦  «        } |  B                    |
 C                    ¦   «          D                    |
jE        ¦  «         F                    d>¦  «        ¦  «        }!| j,        }t)          j        ¦   «         |z
  }"db}#dc|› dd|› de|› df|"dg›dh|#› di|!› di�}$t{          jG        |$¦  «         t‘          djdkdldm¬n¦  «        5 }%|% I                    |$¦  «         ddd¦  «         n# 1 swxY w Y   |rYt          |t”          ¦  «        rD| |          K                    ¦   «         }&|Š't‡          ˆ'fdo„|&D ¦   «         ¦  «        sJ dp‰'› �¦   «         ‚|!S )raË  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.
        quantize (int | str | None): Requested precision: 16 (FP16), 8 (INT8), or None/32 (FP32). Exported rows apply it
            at export, where a format may reject an explicit 32 or fall back to the precision it requires; the native
            PyTorch row is not exported and only 16 affects it, selecting FP16 inference. Each format then runs
            inference at its own runtime precision.
        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.
        format (str): Export format for benchmarking. If not supplied all formats are benchmarked.
        **kwargs (Any): Export options; nms selects the same head for native and exported benchmarks.

    Returns:
        (polars.DataFrame): A Polars 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="yolo26n.pt", imgsz=640)
    r   )Ú_handle_deprecationé   Tz'benchmark() only supports square imgsz.NéÿÿÿÿÚASCII_BORDERS_ONLY_CONDENSEDF©ÚverboseÚArgumentzExpected format to be one of z, but got 'z'.)u   â�ŒN>   ÚpbÚedgetpuÚsaved_modelr+   Úobbz.TensorFlow GraphDef not supported for OBB taskr,   z3Edge TPU export only supported on non-aarch64 LinuxÚedgetpu_compilerz,Edge TPU benchmark requires edgetpu_compilerÚcoremlz;CoreML export only supported on macOS and non-aarch64 LinuxzSCoreML not benchmarked on macOS Python>=3.13 (coremltools/OpenVINO OpenMP deadlock)z;YOLOWorldv2 TensorFlow exports not supported by onnx2tf yetÚpaddlez,YOLOWorldv2 Paddle exports not supported yetz3Windows and Jetson Paddle exports not supported yetzTPaddlePaddle not benchmarked on Python>=3.13 (protobuf ABI conflict with TensorFlow)Úmnnz)YOLOWorldv2 MNN exports not supported yetzKMNN not benchmarked on Python>=3.13 (protobuf ABI conflict with TensorFlow)Úncnnz*YOLOWorldv2 NCNN exports not supported yetÚimxz%YOLOWorldv2 IMX exports not supported>   ÚposeÚdetectÚsegmentÚclassifyzbIMX export is only supported for detection, classification, pose estimation and segmentation tasksÚC2fz*IMX only supported for YOLOv8n and YOLO11nÚrknnz*YOLOWorldv2 RKNN exports not supported yetzRKNN only supported on Linuxz1RKNN Inference only supported on Rockchip devicesÚ
executorchz0YOLOWorldv2 ExecuTorch exports not supported yetÚaxeleraz)YOLOWorldv2 Axelera exports not supportedzGexport is only supported on Linux and is not supported on ARM64 Docker.r7   c              3  ó@   K  — | ]}t          |t          ¦  «        V — Œd S ©N)Ú
isinstancer   )Ú.0Úms     úl/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/ultralytics/utils/benchmarks.pyú	<genexpr>zbenchmark.<locals>.<genexpr>¸   s-   è è € Ð;tÐ;tÐYZ½JÀqÍ)Ñ<TÔ<TÐ;tÐ;tÐ;tÐ;tÐ;tÐ;tó    zDAxelera export does not currently support YOLO26 segmentation modelsÚlitertz6LiteRT benchmark only supported on Linux x86 and macOSzLLiteRT not benchmarked on macOS CI (ai-edge-litert/TF abseil mutex deadlock)r    zinference not supported on CPUÚcudazinference not supported on GPUú-Údata)ÚimgszÚformatÚquantizerH   Údevicer)   )Útaskzexport failedu   â�Žr5   z(GraphDef Pose inference is not supportedzinference not supportedÚDarwinzinference requires macOS>=10.13Úcoreaizinference requires macOSz,inference only supported on Axelera hardwarezbus.jpgÚnms)rI   rL   rK   rP   r)   r!   )
rH   ÚbatchrI   ÚrectÚplotsrL   rK   rP   r)   ÚconfÚ	inferenceéè  é   u   âœ…é   zBenchmark failure for z: )rL   ÚFormatu	   Statusâ�”z	Size (MB)zInference time (ms/im)ÚFPSÚrow)ÚschemaÚorientú )Úoffsetud   Benchmarks legend:  - âœ… Success  - â�Ž Export passed but validation failed  - â�Œï¸� Export failedz
Benchmarks complete for z on z
 at imgsz=z (z.2fzs)
ú
zbenchmarks.logÚaÚignorezutf-8)ÚerrorsÚencodingc              3  óL   •K  — | ]}t          j        |¦  «        °|‰k    V — Œd S r>   )ÚnpÚisnan)r@   ÚxÚfloors     €rB   rC   zbenchmark.<locals>.<genexpr>  s4   øè è € ÐAÐA µR´X¸a±[´[ÐA�1�u’9ÐAÐAÐAÐAÐAÐArD   z%Benchmark failure: metric(s) < floor © )LÚultralytics.cfgr$   r   r?   ÚlistÚpolarsÚConfigÚset_tbl_colsÚset_tbl_rowsÚset_tbl_width_charsÚset_tbl_hide_column_data_typesÚset_tbl_hide_dataframe_shapeÚset_tbl_formattingr   Ústrr   r   r
   rM   r   ÚtimeÚlowerÚ	frozensetr   ÚzipÚvaluesr   r   r   ÚshutilÚwhichr   r	   r   Ú__str__r   r   ÚanyÚmodelÚmodulesr   ÚtypeÚpt_pathÚ	ckpt_pathÚ
model_namer   Úexportr   ÚplatformÚsystemÚpredictr   ÚgetÚvalÚresults_dictÚspeedÚroundÚappendr   Ú	ExceptionÚAssertionErrorr   Úerrorr   Ú	DataFrameÚwith_row_indexÚwith_columnsÚallÚcastÚStringÚ	fill_nullÚinfoÚopenÚwriteÚfloatÚto_numpy)(r   rH   rI   rK   rL   r)   ÚepsrJ   Úkwargsr$   ÚplÚkeyÚyÚt0Ú
format_argÚformatsÚnameÚexport_formatÚsuffixr    ÚgpuÚ
valid_argsÚ_ÚemojiÚfilenameÚexported_modelÚexport_dataÚresultsÚmetricrŒ   ÚfpsÚeÚdfÚ
df_displayÚdtÚlegendÚsÚfÚmetricsri   s(                                          @rB   Ú	benchmarkr»   ?   sR  ø€ ðH 4Ð3Ð3Ð3Ð3Ð3à Ð  Ñ(Ô(€FÝ˜ÑÔ€EÝ#-¨eµTÑ#:Ô#:Ðoˆ5�Œ8�u˜Q”xÒÐÐÐÐÐFoÑÔÐDàÐÐÐà„I×Ò˜2ÑÔÐØ„I×Ò˜2ÑÔÐØ„I×!Ò! "Ñ%Ô%Ð%Ø„I×,Ò,¨TÑ2Ô2Ð2Ø„I×*Ò*¨4Ñ0Ô0Ð0Ø„I× Ò Ð!?Ñ@Ô@Ð@å˜6¨5Ð1Ñ1Ô1€FÝ�%�#�t˜Ñ%Ô%ð Ý�U‘”ˆØÐ(•9˜UœZÔ(€DÝ
�e”jÔ
!€Cà
€AÝ	Œ‰Œ€Bà—’‘”€JØð iÝ�NÑ,Ô,¨ZÔ8Ñ9Ô9ˆØ˜WÐ$Ð$Ð$Ð&hÀgÐ&hÐ&hÐZdÐ&hÐ&hÐ&hÑ$Ô$Ð$Ý@CÅ^ÑEUÔEU×E\ÒE\ÑE^ÔE^Ð@_ð ~Uñ ~UÑ<ˆˆm˜V S¨#¨z¸1Ø%‰ˆˆxð|	UØð ˜j¨MÒ9Ð9ØÝ%ð ¨jð ¸]ÐNnÐ=nÐ=nØð  Ò$Ð$Ø”z UÒ*Ð*Ð*Ð,\Ñ*Ô*Ð*Ð*Ø )Ò+Ð+ÝÐa¥UÐaÐaÐ,aÑaÔaÐ*Ý”|Ð$6Ñ7Ô7ÐgÐgÐ9gÑgÔgÐ7ÐgØ (Ò*Ð*ÝÐt¥Ðt­uÐtÐtÐ7tÑtÔtÐ5õ "ð Õ&<ð ð Øiñô Ð=ð Ð @Ð@Ð@Ý% e­YÑ7Ô7ÐvÐvÐ9vÑvÔvÐ7Ø Ò(Ð(Ý% e­YÑ7Ô7ÐgÐgÐ9gÑgÔgÐ7ÝÐp¥iÐpµEÐpÐpÐ;pÑpÔpÐ9õ 2ð ð Øjñô Ð1ð  Ò%Ð%Ý% e­YÑ7Ô7ÐdÐdÐ9dÑdÔdÐ7õ 2ð ð Øañô Ð1ð  Ò&Ð&Ý% e­YÑ7Ô7ÐeÐeÐ9eÑeÔeÐ7Ø Ò%Ð%Ý% e­YÑ7Ô7Ð`Ð`Ð9`Ñ`Ô`Ð7Ø”zÐ%NÐNÐNÐNØxñ OÔNÐNð  §¢¡¤Ð/Ð/Ð/Ð1]Ñ/Ô/Ð/Ø Ò&Ð&Ý% e­YÑ7Ô7ÐeÐeÐ9eÑeÔeÐ7ÝÐ<Ð<Ð<Ñ<Ô<�uÝ&™=œ=Ð]Ð]Ð*]Ñ]Ô]Ð(Ø Ò,Ð,Ý% e­YÑ7Ô7ÐkÐkÐ9kÑkÔkÐ7Ø 	Ò)Ð)Ý% e­YÑ7Ô7ÐdÐdÐ9dÑdÔdÐ7Ýð ¥eð µ	ð ð Ø]ñô Ð:ð "œJ¨)Ò3Ð3½Ð;tÐ;tÐ^cÔ^i×^qÒ^qÑ^sÔ^sÐ;tÑ;tÔ;tÑ8tÔ8tÐ3Ð3ØZñ 4Ô3Ðuð  Ò(Ð(ÝÐo¥Ðo­uÐoÐoÐ7oÑoÔoÐ5õ "ð Õ&>Ñ&@Ô&@ð ð Øbñô ÐAð ˜œÐ#Ð#ØÐ<Ð<Ð<Ñ<Ô<�sØ˜œÐ$Ð$ØÐ<Ð<Ð<Ñ<Ô<�sð  Ò#Ð#Ø œ=ÐO¨E¬OÐO¸uÔ?O�Ý!)¨%¡¤��à&,°
Ð&:Ð&:˜d˜dÀ�Ø'˜5œ<ð ØØ(Ø%Ø$Ø!Ø!ðð ð ðð �õ 6@ÀÅvÑ5NÔ5NÐ!s¥¨Ñ!1Ô!1Ð!1ÕTXÐYaÐhmÔhrÐTsÑTsÔTs�Ø¥ X¡¤Ð.Ð.Ð.°Ñ.Ô.Ð.ØˆEð ”: Ò'Ð'¨=¸DÒ+@Ð+@Ð+@ÐBlÑ+@Ô+@Ð@Ø  IÒ-Ð-Ð-Ð/HÑ-Ô-Ð-Ø  HÒ,Ð,µ´Ñ0AÔ0AÀXÒ0MÐ0MÐ0MÐOpÑ0MÔ0MÐMØ  HÒ,Ð,µÐ,Ð,Ð7QÑ,Ô,Ð5Ø  IÒ-Ð-Ð-Ð/]Ñ-Ô-Ð-Ø×"Ò"Ý˜Ñ"ØØØ!Ø—J’J˜uÑ%Ô%Øð #ñ ô ð ð %×(Ò(ØØØØØØØ!Ø—J’J˜uÑ%Ô%ØØð )ñ ô ˆGð $Ô0°Ô5°w´}À[Ô7Q�EˆFÝ˜ ¨¡Ñ,¨aÑ0Ô0ˆCØ�HŠH�d˜E¥5­°8Ñ)<Ô)<¸aÑ#@Ô#@Å%ÈÐPQÑBRÔBRÕTYÐZ_ÐabÑTcÔTcÐehÐiÑjÔjÐjÑjøÝð 	Uð 	Uð 	UØð WÝ˜A‘w”w¥.Ð0Ð0Ð0Ð2VÈ4Ð2VÐ2VÐSTÐ2VÐ2VÑ0Ô0Ð0ÝŒLÐ=°$Ð=Ð=¸!Ð=Ð=Ñ>Ô>Ð>Ø�HŠH�d˜E¥5­°8Ñ)<Ô)<¸aÑ#@Ô#@À$ÈÈdÐSÑTÔTÐTÐTÐTÐTÐTÑTøøøøð		Uøøøõ �fÐÑÔÐØ	�Š�a ¨;¸ÀSÐJbÐdiÐ jÐsxˆÑ	yÔ	y€BØ	×	Ò	˜3 qÐ	Ñ	)Ô	)€BØ—’ §¢¡¤§¢¨r¬yÑ!9Ô!9×!CÒ!CÀCÑ!HÔ!HÑIÔI€JàÔ€DÝ	Œ‰Œ�rÑ	€BØs€FØl TÐlÐl¨tÐlÐl¸uÐlÐlÈÐlÐlÐlÐTZÐlÐlÐ^hÐlÐlÐl€AÝ
„K��N„N€NÝ	Ð ¨H¸wÐ	GÑ	GÔ	Gð È1Ø	�Š�‰
Œ
ˆ
ðð ð ñ ô ð ð ð ð ð ð øøøð ð ð ð ð ð s•:˜g¥uÑ-Ô-ð sØ�S”'×"Ò"Ñ$Ô$ˆØˆÝÐAÐAÐAÐA gÐAÑAÔAÑAÔAÐrÐrÐCrÐkpÐCrÐCrÑrÔrÐAàÐs8   Ç0\5Ç9\5ÈT,\5Ü5
_Ü?A8^=Þ=_â8cãcã!cc                  óÊ   — e Zd ZdZ	 	 	 	 	 	 	 d4d5d„Zd„ Zd„ Zed6d„¦   «         Zed7d8d#„¦   «         Z	d9d:d'„Z
ed(„ ¦   «         Zd9d;d)„Zd<d0„Zed<d1„¦   «         Zed=d3„¦   «         ZdS )>ÚProfileModelsa3  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.
        quantize (int | str | None): Export precision for TensorRT profiling, e.g. 16 (FP16) or 8 (INT8).
        trt (bool): Flag to indicate whether to profile using TensorRT.
        device (torch.device): Device used for profiling.

    Methods:
        run: Profile YOLO models for speed and accuracy across various formats.
        get_files: Get all relevant model files.
        get_onnx_model_info: Extract metadata from an ONNX model.
        iterative_sigma_clipping: Apply sigma clipping to remove outliers.
        profile_tensorrt_model: Profile a TensorRT model.
        profile_onnx_model: Profile an ONNX model.
        generate_table_row: Generate a table row with model metrics.
        generate_results_dict: Generate a dictionary of profiling results.
        print_table: Print a formatted table of results.

    Examples:
        Profile models and print results
        >>> from ultralytics.utils.benchmarks import ProfileModels
        >>> profiler = ProfileModels(["yolo26n.yaml", "yolov8s.yaml"], imgsz=640)
        >>> profiler.run()
    éd   é
   é<   é€  é   TNÚpathsú	list[str]Únum_timed_runsÚintÚnum_warmup_runsÚmin_timerœ   rI   rK   úint | str | NoneÚtrtÚboolrL   útorch.device | str | Nonec	                ó”   — || _         || _        || _        || _        || _        || _        || _        t          |d¬¦  «        | _        dS )a�  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.
            quantize (int | str | None): Export precision for TensorRT profiling, e.g. 16 (FP16, default) or 8 (INT8).
            trt (bool): Flag to indicate whether to profile using TensorRT.
            device (torch.device | str | None): Device used for profiling. If None, it is determined automatically.

        Notes:
            quantize applies only to the TensorRT profiling export; ONNX profiling stays FP32 (FP16 is slower on CPU).
        Fr(   N)	rÃ   rÅ   rÇ   rÈ   rI   rK   rÊ   r   rL   )	ÚselfrÃ   rÅ   rÇ   rÈ   rI   rK   rÊ   rL   s	            rB   Ú__init__zProfileModels.__init__6  sO   € ð4 ˆŒ
Ø,ˆÔØ.ˆÔØ ˆŒØˆŒ
Ø ˆŒØˆŒÝ# F°EÐ:Ñ:Ô:ˆŒˆˆrD   c           	     óö  — |                       ¦   «         }|st          j        d¦  «         g S g }g }|D �]¯}|                     d¦  «        }|j        dv rÅt          t          |¦  «        ¦  «        }|                     d¬¦  «         |                     | j	        ¬¦  «        }| j
        rM| j        j        dk    r=|                     ¦   «         s)|                     d| j        | j	        | j        d¬	¦  «        }|                     d
| j	        | j        d¬¦  «        }n%|j        dk    r|                      |¦  «        }|}n�Œ|                      t          |¦  «        ¦  «        }	|                      t          |¦  «        ¦  «        }
|                     |                      |j        |
|	|¦  «        ¦  «         |                     |                      |j        |
|	|¦  «        ¦  «         �Œ±|                      |¦  «         |S )aÐ  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(["yolo26n.yaml", "yolo11s.yaml"])
            >>> results = profiler.run()
        z'No matching *.pt or *.onnx files found.z.engine¾   ú.ptú.ymlú.yamlFr(   ©rI   r    Úengine)rJ   rK   rI   rL   r)   Úonnx)rJ   rI   rL   r)   z.onnx)Ú	get_filesr   ÚwarningÚwith_suffixr¨   r   ru   Úfuser™   rI   rÊ   rL   r�   Úis_filer…   rK   Úget_onnx_model_infoÚprofile_tensorrt_modelÚprofile_onnx_modelrŽ   Úgenerate_table_rowÚstemÚgenerate_results_dictÚprint_table)rÎ   ÚfilesÚ
table_rowsÚoutputÚfileÚengine_filer   Ú
model_infoÚ	onnx_fileÚt_engineÚt_onnxs              rB   ÚrunzProfileModels.runY  sú  € ð —’Ñ Ô ˆàð 	ÝŒNÐDÑEÔEÐEØˆIàˆ
ØˆØð 	_ñ 	_ˆDØ×*Ò*¨9Ñ5Ô5ˆKØŒ{Ð6Ð6Ð6Ý�S ™YœY™œ�Ø—
’
 5�
Ñ)Ô)Ð)Ø"ŸZšZ¨d¬j˜ZÑ9Ô9�
Ø”8ð  ¤Ô 0°EÒ 9Ð 9À+×BUÒBUÑBWÔBWÐ 9Ø"'§,¢,Ø'Ø!%¤Ø"œjØ#œ{Ø %ð #/ñ #ô #�Kð "ŸLšLØ!Øœ*Øœ;Ø!ð	 )ñ ô �	�	ð ” Ò'Ð'Ø!×5Ò5°dÑ;Ô;�
Ø �	�	áà×2Ò2µ3°{Ñ3CÔ3CÑDÔDˆHØ×,Ò,­S°©^¬^Ñ<Ô<ˆFØ×Ò˜d×5Ò5°d´iÀÈÐS]Ñ^Ô^Ñ_Ô_Ð_Ø�MŠM˜$×4Ò4°T´YÀÈÐR\Ñ]Ô]Ñ^Ô^Ð^Ñ^à×Ò˜Ñ$Ô$Ð$ØˆrD   c                óè  ‡— g }| j         D ]«Št          ‰¦  «        Š‰                     ¦   «         r&g d¢}|                     ˆfd„|D ¦   «         ¦  «         ŒK‰j        dv r#|                     t          ‰¦  «        ¦  «         Œw|                     t          j        t          ‰¦  «        ¦  «        ¦  «         Œ¬t          j	        dt          |¦  «        › �¦  «         d„ t          |¦  «        D ¦   «         S )z¥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.
        )z*.ptz*.onnxz*.yamlc                ób   •— g | ]+}t          j         t          ‰|z  ¦  «        ¦  «        D ]}|‘ŒŒ,S rj   )Úglobru   )r@   Úextrç   Úpaths      €rB   ú
<listcomp>z+ProfileModels.get_files.<locals>.<listcomp>š  s=   ø€ Ð`Ð`Ð` sÅTÄYÍsÐSWÐZ]ÑS]ÉÌÑE_ÔE_Ð`Ð`¸T˜dÐ`Ð`Ð`Ð`rD   rÑ   zProfiling: c                ó,   — g | ]}t          |¦  «        ‘ŒS rj   r   )r@   rç   s     rB   ró   z+ProfileModels.get_files.<locals>.<listcomp>¡  s   € Ð5Ð5Ð5˜t•�T‘
”
Ð5Ð5Ð5rD   )rÃ   r   Úis_dirÚextendr¨   rŽ   ru   rð   r   r™   Úsorted)rÎ   rä   Ú
extensionsrò   s      @rB   rØ   zProfileModels.get_files�  sò   ø€ ð ˆØ”Jð 	3ð 	3ˆDÝ˜‘:”:ˆDØ�{Š{‰}Œ}ð 3Ø9Ð9Ð9�
Ø—’Ð`Ð`Ð`Ð`¨jÐ`Ñ`Ô`ÑaÔaÐaÐaØ”Ð 8Ð8Ð8Ø—’�S ™YœYÑ'Ô'Ð'Ð'à—’�TœY¥s¨4¡y¤yÑ1Ô1Ñ2Ô2Ð2Ð2åŒÐ1¥&¨¡-¤-Ð1Ð1Ñ2Ô2Ð2Ø5Ð5¥v¨e¡}¤}Ð5Ñ5Ô5Ð5rD   rê   ru   c                ó   — dS )z\Extract metadata from an ONNX model file including layers, parameters, gradients, and FLOPs.)ç        rú   rú   rú   rj   )rê   s    rB   rÝ   z!ProfileModels.get_onnx_model_info£  s
   € ð "Ð!rD   rW   é   rH   ú
np.ndarrayÚsigmaÚ	max_itersc                ó(  — t          j        | ¦  «        } t          |¦  «        D ]m}t          j        | ¦  «        t          j        | ¦  «        }}| | |||z  z
  k    | |||z  z   k    z           }t          |¦  «        t          | ¦  «        k    r n|} Œn| S )a{  Apply iterative sigma clipping to data to remove outliers.

        Args:
            data (np.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:
            (np.ndarray): Clipped data array with outliers removed.
        )rf   ÚarrayÚrangeÚmeanÚstdÚlen)rH   rý   rþ   r«   r  r  Úclipped_datas          rB   Úiterative_sigma_clippingz&ProfileModels.iterative_sigma_clipping¨  s—   € õ Œx˜‰~Œ~ˆÝ�yÑ!Ô!ð 	 ð 	 ˆAÝœ ™œ¥r¤v¨d¡|¤|�#ˆDà ¨°¸±Ñ);Ò!;ÀÈÈuÐWZÉ{ÑHZÒ@ZÑ [Ô\ˆLÝ�<Ñ Ô ¥C¨¡I¤IÒ-Ð-Ø�ØˆDˆDØˆrD   r!   rè   rž   c                óh  — | j         r!t          |¦  «                             ¦   «         sdS t          |¦  «        }t	          j        | j        | j        dft          j        ¬¦  «        }d}t          d¦  «        D ]U}t          j
        ¦   «         }t          | j        ¦  «        D ]} ||| j        d¬¦  «         Œt          j
        ¦   «         |z
  }ŒVt          t          | j        ||z   z  | j        z  ¦  «        | j        dz  ¦  «        }g }	t!          t          |¦  «        |¬¦  «        D ];} ||| j        d¬¦  «        }
|	                     |
d	         j        d
         ¦  «         Œ<|                      t	          j        |	¦  «        dd¬¦  «        }	t	          j        |	¦  «        t	          j        |	¦  «        fS )aw  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:
            (tuple[float, float]): Mean and standard deviation of inference time in milliseconds.
        )rú   rú   rû   )Údtyperú   F)rI   r)   é2   ©Údescr   rU   rW   ©rý   rþ   )rÊ   r   rÜ   r   rf   ÚzerosrI   Úuint8r  rv   Úperf_counterrÇ   Úmaxr�   rÈ   rÅ   r   rŽ   rŒ   r  r   r  r  )rÎ   rè   rž   r   Ú
input_dataÚelapsedr«   Ú
start_timeÚnum_runsÚ	run_timesr°   s              rB   rÞ   z$ProfileModels.profile_tensorrt_model¾  s³  € ð Œxð 	�t KÑ0Ô0×8Ò8Ñ:Ô:ð 	Ø�8õ �[Ñ!Ô!ˆÝ”X˜tœz¨4¬:°qÐ9ÅÄÐJÑJÔJˆ
ð ˆÝ�q‘”ð 	7ð 	7ˆAÝÔ*Ñ,Ô,ˆJÝ˜4Ô/Ñ0Ô0ð Cð C�Ø��j¨¬
¸EÐBÑBÔBÐBÐBÝÔ'Ñ)Ô)¨JÑ6ˆGˆGõ •u˜Tœ]¨g¸©mÑ<¸tÔ?SÑSÑTÔTÐVZÔViÐlnÑVnÑoÔoˆð ˆ	Ý•e˜H‘o”o¨KÐ8Ñ8Ô8ð 	<ð 	<ˆAØ�e˜J¨d¬jÀ%ÐHÑHÔHˆGØ×Ò˜W QœZÔ-¨kÔ:Ñ;Ô;Ð;Ð;à×1Ò1µ"´(¸9Ñ2EÔ2EÈQÐZ[Ð1Ñ\Ô\ˆ	ÝŒw�yÑ!Ô!¥2¤6¨)Ñ#4Ô#4Ð4Ð4rD   c                ó6   — t          d„ | D ¦   «         ¦  «         S )z<Check whether the tensor shape in the ONNX model is dynamic.c              3  óL   K  — | ]}t          |t          ¦  «        o|d k    V — Œ dS )r   N)r?   rÆ   )r@   Údims     rB   rC   z.ProfileModels.check_dynamic.<locals>.<genexpr>æ  s5   è è € ÐQÐQ¸S•z #¥sÑ+Ô+Ð8°°q²ÐQÐQÐQÐQÐQÐQrD   )r•   )Útensor_shapes    rB   Úcheck_dynamiczProfileModels.check_dynamicã  s$   € õ ÐQÐQÀLÐQÑQÔQÑQÔQÐQÐQrD   c                ób  — ddl }ddlm} |                     ¦   «         }|j        j        |_        d|_         ||t          j	        d¦  «        d|¬¦  «        }i }|j
                             ¦   «         D �]U}|j        }	|                      |j        ¦  «        r•t          |j        ¦  «        dk    rA|                      |j        d	d…         ¦  «        rt!          d
|j        › d|j        › �¦  «        ‚t          |j        ¦  «        dk    rd	d| j        | j        fnd	g|j        d	d…         ¢R }
n|j        }
d|	v rt&          j        }nVd|	v rt&          j        }nEd|	v rt&          j        }n4d|	v rt&          j        }n#d|	v rt&          j        }nt!          d|	› �¦  «        ‚t'          j        j        |
Ž                      |¦  «        ||j        <   �ŒWd}t9          d¦  «        D ]W}t;          j        ¦   «         }t9          | j        ¦  «        D ]}|                      |¦  «         Œt;          j        ¦   «         |z
  }ŒXtC          tE          | j#        ||z   z  | j        z  ¦  «        | j$        ¦  «        }g }tK          t9          |¦  «        |¬¦  «        D ]V}t;          j        ¦   «         }|                      |¦  «         | &                    t;          j        ¦   «         |z
  dz  ¦  «         ŒW|  '                    t'          j(        |¦  «        dd¬¦  «        }t'          j)        |¦  «        t'          j*        |¦  «        fS )at  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:
            (tuple[float, float]): Mean and standard deviation of inference time in milliseconds.
        r   N)ÚONNXBackendé   r    F)rL   Úfp16Úsession_optionsrX   r%   zUnsupported dynamic shape z of rû   Úfloat16rœ   ÚdoubleÚint64Úint32zUnsupported ONNX datatype rú   r
  rV   rW   é   r  )+ÚonnxruntimeÚultralytics.nn.backendsr  ÚSessionOptionsÚGraphOptimizationLevelÚORT_ENABLE_ALLÚgraph_optimization_levelÚintra_op_num_threadsÚtorchrL   ÚsessionÚ
get_inputsr�   r  Úshaper  Ú
ValueErrorr¦   rI   rf   r   Úfloat32Úfloat64r"  r#  ÚrandomÚrandÚastyper  rv   r  rÇ   Úforwardr  r�   rÈ   rÅ   r   rŽ   r  r   r  r  )rÎ   rê   rž   Úortr  Úsess_optionsÚbackendÚinput_data_dictÚinput_tensorÚ
input_typeÚinput_shapeÚinput_dtyper  r«   r  r  r  s                    rB   rß   z ProfileModels.profile_onnx_modelè  si  € ð 	"Ð!Ð!Ð!à7Ð7Ð7Ð7Ð7Ð7ð ×)Ò)Ñ+Ô+ˆØ03Ô0JÔ0YˆÔ-Ø,-ˆÔ)ð �+˜iµ´¸UÑ0CÔ0CÈ%ÐamÐnÑnÔnˆð ˆØ#œO×6Ò6Ñ8Ô8ð 	bñ 	bˆLØ%Ô*ˆJØ×!Ò! ,Ô"4Ñ5Ô5ð 1Ý�|Ô)Ñ*Ô*¨aÒ/Ð/°D×4FÒ4FÀ|ÔGYÐZ[ÐZ\ÐZ\ÔG]Ñ4^Ô4^Ð/Ý$Ð%mÀ,ÔBTÐ%mÐ%mÐZfÔZkÐ%mÐ%mÑnÔnÐnå69¸,Ô:LÑ6MÔ6MÐQRÒ6RÐ6R�Q˜˜4œ: t¤zÐ2Ð2ÐYZÐXtÐ]iÔ]oÐpqÐprÐprÔ]sÐXtÐXtð �ð +Ô0�ð ˜JÐ&Ð&Ý œj��Ø˜JÐ&Ð&Ý œj��Ø˜ZÐ'Ð'Ý œj��Ø˜JÐ&Ð&Ý œh��Ø˜JÐ&Ð&Ý œh��å Ð!J¸jÐ!JÐ!JÑKÔKÐKå13´´ÀÐ1M×1TÒ1TÐU`Ñ1aÔ1aˆO˜LÔ-Ñ.Ñ.ð ˆÝ�q‘”ð 	7ð 	7ˆAÝÔ*Ñ,Ô,ˆJÝ˜4Ô/Ñ0Ô0ð 1ð 1�Ø—’ Ñ0Ô0Ð0Ð0ÝÔ'Ñ)Ô)¨JÑ6ˆGˆGõ •u˜Tœ]¨g¸©mÑ<¸tÔ?SÑSÑTÔTÐVZÔViÑjÔjˆð ˆ	Ý•e˜H‘o”o¨IÐ6Ñ6Ô6ð 	Hð 	HˆAÝÔ*Ñ,Ô,ˆJØ�OŠO˜OÑ,Ô,Ð,Ø×Ò�dÔ/Ñ1Ô1°JÑ>À$ÑFÑGÔGÐGÐGà×1Ò1µ"´(¸9Ñ2EÔ2EÈQÐZ[Ð1Ñ\Ô\ˆ	ÝŒw�yÑ!Ô!¥2¤6¨)Ñ#4Ô#4Ð4Ð4rD   r„   rì   útuple[float, float]rë   ré   ú!tuple[float, float, float, float]c                ó”   — |\  }}}}d|d›d| j         › d|d         d›d|d         d›d	|d         d›d|d         d›d	|d
z  d›d|d›d�S )aÑ  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.
        z| Ú18sz | z | - | r   z.1fõ   Â±r%   z ms | g    €„.Az |rÕ   )	rÎ   r„   rì   rë   ré   Ú_layersÚparamsÚ
_gradientsÚflopss	            rB   rà   z ProfileModels.generate_table_row0  sÄ   € ð$ .8Ñ*ˆ�˜ UðI�Ðpð Ið I D¤Jð Ið I°v¸a´yÐpð Ið IÈÈqÌ	Ðpð Ið IÐ]eÐfgÔ]hÐpð Ið IØ˜Œ{ÐHðIð IØ&,¨s¡lÐHðIð IØ;@ÐHðIð Ið Ið	
rD   c                óŒ   — |\  }}}}| |t          |d¦  «        t          |d         d¦  «        t          |d         d¦  «        dœS )a¼  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.
        rû   r   )z
model/namezmodel/parameterszmodel/GFLOPszmodel/speed_ONNX(ms)zmodel/speed_TensorRT(ms))r�   )r„   rì   rë   ré   rD  rE  rF  rG  s           rB   râ   z#ProfileModels.generate_results_dictH  sT   € ð$ .8Ñ*ˆ�˜ Uà$Ø &Ý! %¨™OœOÝ$)¨&°¬)°QÑ$7Ô$7Ý(-¨h°q¬k¸1Ñ(=Ô(=ð
ð 
ð 	
rD   rå   c                óÚ  — t           j                             ¦   «         rt           j                             d¦  «        nd}ddddt	          ¦   «         › d�d|› d	�d
dg}dd                     d„ |D ¦   «         ¦  «        z   dz   }dd                     d„ |D ¦   «         ¦  «        z   dz   }t          j        d|› �¦  «         t          j        |¦  «         | D ]}t          j        |¦  «         ŒdS )z”Print a formatted table of model profiling results.

        Args:
            table_rows (list[str]): List of formatted table row strings.
        r   ÚGPUÚModelzsize<br><sup>(pixels)zmAP<sup>val<br>50-95zSpeed<br><sup>CPU (z) ONNX<br>(ms)zSpeed<br><sup>z TensorRT<br>(ms)zparams<br><sup>(M)zFLOPs<br><sup>(B)ú|c              3  ó"   K  — | ]
}d |› d �V — ŒdS )r^   Nrj   ©r@   Úhs     rB   rC   z,ProfileModels.print_table.<locals>.<genexpr>t  s*   è è € Ð:Ð:¨Q  A   Ð:Ð:Ð:Ð:Ð:Ð:rD   c              3  ó@   K  — | ]}d t          |¦  «        dz   z  V — ŒdS )rG   rW   N)r  rN  s     rB   rC   z,ProfileModels.print_table.<locals>.<genexpr>u  s1   è è € Ð"GÐ"G¸! 3­#¨a©&¬&°1©*Ñ#5Ð"GÐ"GÐ"GÐ"GÐ"GÐ"GrD   z

N)r,  rF   Úis_availableÚget_device_namer   Újoinr   r™   )rå   r©   ÚheadersÚheaderÚ	separatorr[   s         rB   rã   zProfileModels.print_tablec  s  € õ 05¬z×/FÒ/FÑ/HÔ/HÐS�eŒj×(Ò(¨Ñ+Ô+Ð+ÈeˆàØ#Ø"Ø@¥,¡.¤.Ð@Ð@Ð@Ø3˜SÐ3Ð3Ð3Ø Øð
ˆð �s—x’xÐ:Ð:°'Ð:Ñ:Ô:Ñ:Ô:Ñ:¸SÑ@ˆØ˜#Ÿ(š(Ð"GÐ"G¸wÐ"GÑ"GÔ"GÑGÔGÑGÈ#ÑMˆ	åŒ�O˜6�O�OÑ$Ô$Ð$ÝŒ�IÑÔÐØð 	ð 	ˆCÝŒK˜ÑÔÐÐð	ð 	rD   )r¾   r¿   rÀ   rÁ   rÂ   TN)rÃ   rÄ   rÅ   rÆ   rÇ   rÆ   rÈ   rœ   rI   rÆ   rK   rÉ   rÊ   rË   rL   rÌ   )rê   ru   )rW   rû   )rH   rü   rý   rœ   rþ   rÆ   )r!   )rè   ru   rž   rœ   )rê   ru   rž   rœ   )r„   ru   rì   r?  rë   r?  ré   r@  )rå   rÄ   )Ú__name__Ú
__module__Ú__qualname__Ú__doc__rÏ   rí   rØ   ÚstaticmethodrÝ   r  rÞ   r  rß   rà   râ   rã   rj   rD   rB   r½   r½     s_  € € € € € ðð ðF "Ø!ØØØ%'ØØ,0ð!;ð !;ð !;ð !;ð !;ðF4ð 4ð 4ðl6ð 6ð 6ð( ð"ð "ð "ñ „\ð"ð ðð ð ð ñ „\ðð*#5ð #5ð #5ð #5ð #5ðJ ðRð Rñ „\ðRðF5ð F5ð F5ð F5ð F5ðP
ð 
ð 
ð 
ð0 ð
ð 
ð 
ñ „\ð
ð4 ðð ð ñ „\ðð ð rD   r½   )1rZ  Ú
__future__r   rð   r†   r{   rv   Úcopyr   Úpathlibr   Únumpyrf   Ú
torch.cudar,  Úultralyticsr   r   r	   rk   r
   r   Úultralytics.engine.exporterr   Úultralytics.nn.modulesr   Úultralytics.utilsr   r   r   r   r   r   r   r   r   r   Úultralytics.utils.checksr   r   r   r   Úultralytics.utils.filesr   Úultralytics.utils.torch_utilsr   r   r»   r½   rj   rD   rB   ú<module>rh     s  ððð ð8 #Ð "Ð "Ð "Ð "Ð "à €€€Ø €€€Ø €€€Ø €€€Ø Ð Ð Ð Ð Ð Ø Ð Ð Ð Ð Ð à Ð Ð Ð Ø Ð Ð Ð à /Ð /Ð /Ð /Ð /Ð /Ð /Ð /Ð /Ð /Ø 2Ð 2Ð 2Ð 2Ð 2Ð 2Ð 2Ð 2Ø 6Ð 6Ð 6Ð 6Ð 6Ð 6Ø ,Ð ,Ð ,Ð ,Ð ,Ð ,ðð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð bÐ aÐ aÐ aÐ aÐ aÐ aÐ aÐ aÐ aÐ aÐ aØ -Ð -Ð -Ð -Ð -Ð -Ø EÐ EÐ EÐ EÐ EÐ EÐ EÐ Eð ˜Ñ
$Ø	Ø
ØØØØØðSð Sð Sð Sðleð eð eð eð eñ eô eð eð eð erD   