§
    l¬¸jY  ã                  óT  — d dl m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Zd dl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 d dlmZ ed„ ¦   «         Z G d	„ d
ej        j        ¦  «        Zd1d„Z  e¦   «         	 	 	 	 d2d3d„¦   «         Z!	 	 	 	 	 d4d5d)„Z"	 	 	 	 	 	 	 	 	 	 d6d7d0„Z#dS )8é    )ÚannotationsN)Ú	lru_cache)ÚPath)ÚASSETSÚ	IS_JETSONÚLOGGERÚTORCH_VERSIONÚThreadingLockedÚimreadÚis_dgxÚ	is_jetson)Úcheck_requirementsÚcheck_tensorrtÚcheck_version)Ú	TORCH_2_4c                 óH   — ddl } |                      | j        j        ¦  «        S )zVReturn the shared TensorRT logger, kept alive for every builder and inference runtime.r   N)ÚtensorrtÚLoggerÚINFO)Útrts    úo/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/ultralytics/utils/export/engine.pyÚget_tensorrt_loggerr      s&   € ð ÐÐÐà�:Š:�c”j”oÑ&Ô&Ð&ó    c                  ó,   ‡ — e Zd ZdZdˆ fd„Zdd„Zˆ xZS )Ú_NormalizeCoordszVWrap a model with input-relative box and pose coordinates for per-tensor quantization.Úmodelútorch.nn.ModuleÚhÚintÚwÚtaskÚstrÚncÚ	kpt_shapeútuple | Nonec                óœ   •— t          ¦   «                              ¦   «          || _        || _        || _        || _        || _        || _        dS )z:Initialize with the wrapped model and prediction metadata.N)ÚsuperÚ__init__r   r   r    r!   r#   r$   )Úselfr   r   r    r!   r#   r$   Ú	__class__s          €r   r(   z_NormalizeCoords.__init__   sG   ø€ å‰Œ×ÒÑÔÐØˆŒ
ØˆŒØˆŒØˆŒ	ØˆŒØ"ˆŒˆˆr   Úxútorch.Tensorc                óô  — |                       |¦  «        }t          |t          t          f¦  «        r|d         n|}t	          j        | j        | j        | j        | j        g|j        |j	        ¬¦  «         
                    ddd¦  «        }|dd…dd…f         |z  g}| j        dk    �r%| j        �r|                     |dd…dd| j        z   …f         ¦  «         |j        \  }}}|dd…d| j        z   d…f          
                    || j        d         | j        d         |¦  «        }	t	          j        | j        | j        g|j        |j	        ¬¦  «         
                    dddd¦  «        }
t	          j        |	dd…dd…dd…f         |
z  |	dd…dd…dd…f         gd¬¦  «        }	|                     |	                     |d	|¦  «        ¦  «         nk|                     |dd…dd| j        z   …f         ¦  «         |j        d         d| j        z   k    r)|                     |dd…d| j        z   d…f         ¦  «         t	          j        |d¬¦  «        }t          |t          t          f¦  «        r|g|dd…         ¢R n|S )
zJRun the wrapped model and normalize its coordinate channels by input size.r   )ÚdtypeÚdeviceé   é   NÚposeé   )Údiméÿÿÿÿ)r   Ú
isinstanceÚtupleÚlistÚtorchÚtensorr    r   r.   r/   Úviewr!   r$   Úappendr#   ÚshapeÚcatÚreshape)r)   r+   ÚyÚdetÚbox_whÚpartsÚbÚ_ÚaÚkptsÚkpt_whs              r   Úforwardz_NormalizeCoords.forward)   sr  € à�JŠJ�q‰MŒMˆÝ  ¥U­D MÑ2Ô2Ð9ˆa�Œdˆd¸ˆÝ”˜tœv t¤v¨t¬v°t´vÐ>ÀcÄiÐX[ÔXbÐcÑcÔc×hÒhÐijÐlmÐopÑqÔqˆØ�Q�Q�Q˜˜˜�U”˜fÑ$Ð%ˆØŒ9˜ÒÑ 4¤>ÑØ�LŠL˜˜Q˜Q˜Q  A¨¬¡K Ð/Ô0Ñ1Ô1Ð1Ø”i‰GˆAˆq�!Ø�q�q�q˜!˜dœg™+˜-˜-Ð'Ô(×-Ò-¨a°´ÀÔ1BÀDÄNÐSTÔDUÐWXÑYÔYˆDÝ”\ 4¤6¨4¬6Ð"2¸#¼)ÈCÌJÐWÑWÔW×\Ò\Ð]^Ð`aÐcdÐfgÑhÔhˆFÝ”9˜d 1 1 1 a a a¨¨!¨ 8œn¨vÑ5°t¸A¸A¸A¸q¸q¸qÀ!À"À"¸H´~ÐFÈAÐNÑNÔNˆDØ�LŠL˜Ÿš a¨¨QÑ/Ô/Ñ0Ô0Ð0Ð0à�LŠL˜˜Q˜Q˜Q  A¨¬¡K Ð/Ô0Ñ1Ô1Ð1ØŒy˜Œ|˜a $¤'™kÒ)Ð)Ø—’˜S    A¨¬¡K M MÐ!1Ô2Ñ3Ô3Ð3ÝŒi˜ 1Ð%Ñ%Ô%ˆÝ *¨1­uµd¨mÑ <Ô <ÐE�ˆ}�a˜˜˜”eˆ}ˆ}ˆ}À#ÐEr   )r   r   r   r   r    r   r!   r"   r#   r   r$   r%   )r+   r,   )Ú__name__Ú
__module__Ú__qualname__Ú__doc__r(   rI   Ú__classcell__)r*   s   @r   r   r      s_   ø€ € € € € Ø`Ð`ð#ð #ð #ð #ð #ð #ðFð Fð Fð Fð Fð Fð Fð Fr   r   Úonnxútypes.ModuleTypeÚreturnr   c                óð   — d                      t          j        d¦  «        dd…         ¦  «        }ddddddddddd	œ
                     |d
¦  «        }t	          || j                             ¦   «         ¦  «        S )z@Return max ONNX opset for this torch version with ONNX fallback.ú.Nr3   é   é   é   é   é   )
z1.8z1.9z1.10z1.11z1.12z1.13z2.0z2.1z2.2z2.3é   )Újoinr	   ÚsplitÚgetÚminÚdefsÚonnx_opset_version)rO   ÚversionÚopsets      r   Úbest_onnx_opsetrb   >   s…   € à�hŠh•}Ô*¨3Ñ/Ô/°°°Ô3Ñ4Ô4€GàØØØØØØØØØðð ÷ 
‚cˆ'�2ÑÔð 
õ ˆu�d”i×2Ò2Ñ4Ô4Ñ5Ô5Ð5r   rV   r   r   Úimú'torch.Tensor | tuple[torch.Tensor, ...]Úoutput_fileú
Path | strra   Úinput_namesúlist[str] | NoneÚoutput_namesÚdynamicúdict | Noner"   c                óŠ   — |€dg}|€dg}t           rddini }t          j        j        | ||f||||dœ|¤Ž t	          |¦  «        S )a«  Export a PyTorch model to ONNX format.

    Args:
        model (torch.nn.Module): The PyTorch model to export.
        im (torch.Tensor | tuple[torch.Tensor, ...]): Example input tensor(s) for tracing.
        output_file (Path | str): Path to save the exported ONNX file.
        opset (int): ONNX opset version to use for export.
        input_names (list[str] | None): List of input tensor names. Defaults to ``["images"]``.
        output_names (list[str] | None): List of output tensor names. Defaults to ``["output0"]``.
        dynamic (dict | None): Dictionary specifying dynamic axes for inputs and outputs.

    Returns:
        (str): Path to the exported ONNX file.
    NÚimagesÚoutput0ÚdynamoF)Úopset_versionrg   ri   Údynamic_axes)r   r9   rO   Úexportr"   )r   rc   re   ra   rg   ri   rj   Úkwargss           r   Ú
torch2onnxrt   R   s„   € ð0 ÐØ�jˆØÐØ!�{ˆÝ"+Ð3ˆh˜ÐÐ°€FÝ	„JÔØØ
Øð	ð ØØ!Øð	ð 	ð ð	ð 	ð 	õ ˆ{ÑÔÐr   ©r0   é   é€  rw   FÚ Ú	onnx_fileÚquantizeúint | str | Noner=   útuple[int, int, int, int]ÚboolÚprefixc           
     ó–  — |dk    r|€t          d¦  «        ‚t          d¦  «         ddl}|                     | d¬¦  «        j        j        d         j        }|dk    �rIddlm} t          t          | ¦  «                             d	¦  «        ¦  «        }	g d}}
|D ];}|
                     |d
         ¦  «         ||
d         j        d         z  }|dk    r nŒ<t          j        |
¦  «        }~
~|                     t          j        ¦  «                             d¦  «        }t'          j        |› d|j        d         › d�¦  «         |r&d|› dd                     d„ |D ¦   «         ¦  «        › �ini } || fd||                     ¦   «                              ¦   «         iddgddg|	dœ|¤Ž |	S ddlm} t          t          | ¦  «                             d¦  «        ¦  «        }	t'          j        |› d�¦  «         t5          j        t9          t:          dz  ¦  «        |ddd…         ¦  «        dddd…f                              d dd¦  «        }t?          j        ||dd…         ¦  «        }t?          j         ||¦  «         !                    t>          j        d!¬"¦  «        d#z  }| "                    | #                    | d$d%||i¬&¦  «        |	¦  «         |	S )'a;  Bake reduced precision into an ONNX model for TensorRT 11 strongly-typed builds using NVIDIA ModelOpt.

    TensorRT 11 is strongly-typed only: it removed the FP16/INT8 builder flags and the ``IInt8Calibrator`` interface, so
    reduced precision must be expressed in the ONNX graph itself before building. FP16 is applied via ModelOpt AutoCast
    mixed-precision conversion and INT8 via explicit Q/DQ quantization with calibration.

    Args:
        onnx_file (str): Path to the FP32 ONNX file to convert.
        quantize (int | str | None): Precision scheme, 8 for INT8 Q/DQ nodes or 16 for FP16 precision.
        dataset (ultralytics.data.build.InfiniteDataLoader | None): Dataloader providing INT8 calibration images.
            Required when ``quantize=8``.
        shape (tuple[int, int, int, int]): Input shape (batch, channels, height, width) used for dynamic calibration.
        dynamic (bool): Whether the ONNX model uses dynamic input shapes.
        prefix (str): Prefix for log messages.

    Returns:
        (str): Path to the precision-converted ONNX file.
    é   Nz:INT8 ModelOpt quantization requires a calibration dataset.znvidia-modelopt[onnx]>=0.44r   F©Úload_external_data)rz   z
.int8.onnxÚimgr5   i   ç     ào@z- quantizing ONNX to INT8 with ModelOpt using z calibration images...Úcalibration_shapesú:r+   c              3  ó4   K  — | ]}t          |¦  «        V — Œd S )N)r"   )Ú.0Úds     r   ú	<genexpr>z)modelopt_quantize_onnx.<locals>.<genexpr>®   s(   è è € Ð@WÐ@WÈAÅÀQÁÄÐ@WÐ@WÐ@WÐ@WÐ@WÐ@Wr   Úint8ÚmaxÚcpuz.*\.2/Conv$z.*/dfl/)Úquantize_modeÚcalibration_dataÚcalibration_methodÚcalibration_epsÚnodes_to_excludeÚoutput_path)Úautocastz
.fp16.onnxzB converting ONNX to FP16 mixed precision with ModelOpt AutoCast...zbus.jpgr0   .r3   ÚC)Úorderéÿ   Úfp16T)Úlow_precision_typeÚkeep_io_typesr�   )$Ú
ValueErrorr   rO   ÚloadÚgraphÚinputÚnameÚmodelopt.onnx.quantizationrz   r"   r   Úwith_suffixr<   r=   r9   r>   ÚtoÚfloat32Údiv_r   ÚinforZ   r�   ÚnumpyÚmodelopt.onnxr”   Úcv2Úresizer   r   Ú	transposeÚnpÚbroadcast_toÚastypeÚsaveÚconvert_to_mixed_precision)ry   rz   Údatasetr=   rj   r~   rO   Ú
input_nameÚmodelopt_quantizeÚout_filerm   ÚnÚbatchÚcalibrs   r”   rc   s                    r   Úmodelopt_quantize_onnxr·   |   s!  € ð4 �1‚}€}˜˜ÝÐUÑVÔVÐVõ Ð4Ñ5Ô5Ð5Ø€K€K€Kà—’˜9¸�Ñ?Ô?ÔEÔKÈAÔNÔS€JØ�1‚}�}ØLÐLÐLÐLÐLÐLå•t˜I‘”×2Ò2°<Ñ@Ô@ÑAÔAˆð ˜�ˆØð 	ð 	ˆEØ�MŠM˜% œ,Ñ'Ô'Ð'Ø�˜”Ô! !Ô$Ñ$ˆAØ�CŠxˆxØ�ð å”	˜&Ñ!Ô!ˆØ�EØ—’�œÑ'Ô'×,Ò,¨UÑ3Ô3ˆÝŒ�vÐrÐrÈEÌKÐXYÌNÐrÐrÐrÑsÔsÐsØ^eÐmÐ&¨:Ð(YÐ(Y¸¿ºÐ@WÐ@WÐQVÐ@WÑ@WÔ@WÑ8WÔ8WÐ(YÐ(YÐZÐZÐkmˆØÐØð	
à Ø(¨%¯)ª)©+¬+×*;Ò*;Ñ*=Ô*=Ð>Ø$ð #˜Gð -¨jÐ9Ø ð	
ð 	
ð ð	
ð 	
ð 	
ð ˆà&Ð&Ð&Ð&Ð&Ð&å•4˜	‘?”?×.Ò.¨|Ñ<Ô<Ñ=Ô=€HÝ
„K�6Ð]Ð]Ð]Ñ^Ô^Ð^õ 
Œ•F�6 IÑ-Ñ.Ô.°°e°q¸°e´Ñ	=Ô	=¸cÀ4À4ÀRÀ4¸iÔ	H×	RÒ	RÐSTÐVWÐYZÑ	[Ô	[€BÝ	Œ�2�u˜Q˜R˜R”yÑ	!Ô	!€BÝ	Œ˜˜UÑ	#Ô	#×	*Ò	*­2¬:¸SÐ	*Ñ	AÔ	AÀCÑ	G€BØ‡I‚IØ×+Ò+ØØ%ØØ(¨"Ð-ð	 	,ñ 	
ô 	
ð 	ñô ð ð €Or   úPath | str | NoneÚ	workspaceú
int | NoneÚdlaÚmetadataÚverbosec           
     óH  ‡‡‡.‡/‡0— ddl }t          d¬¦  «        st          ¦   «         rt          d¦  «         	 ddlŠ0n"# t
          $ r t          ¦   «          ddlŠ0Y nw xY wt          ‰0j        dd¬¦  «         t          ‰0j        d	d
¬¦  «         t          j	        d|
› d‰0j        › d�¦  «         |p!t          | ¦  «                             d¦  «        }t          ¦   «         }|	r‰0j        j        n‰0j        j        |_        ‰0                     |¦  «        }|                     ¦   «         }t'          ‰pddz  ¦  «        }t'          ‰0j                             dd¦  «        d         ¦  «        }|dk    }|dk    }|dk    r8t+          |d¦  «        r!|                     ‰0j        j        |¦  «         n||_        |rdndt'          ‰0j        j        ¦  «        z  }|                     |¦  «        Š/t;          |dd¦  «        o|dk    }t;          |dd¦  «        o|dk    }t=          d„ |                     | d¬¦  «        j         j!        D ¦   «         ¦  «        }|o| }|r|€tE          d¦  «        ‚‰�«tF          stE          d¦  «        ‚t          ‰0j        d¦  «        rtE          d ¦  «        ‚t          j	        |
› d!‰› d�¦  «         |s|stE          d"¦  «        ‚‰0j$        j%        |_&        t'          ‰¦  «        |_'        | (                    ‰0j)        j*        ¦  «         |r|s|rtW          | |rdn|||||
¦  «        } ‰0 ,                    ‰/|¦  «        }| -                    | ¦  «        st]          d#| › �¦  «        ‚ˆ/fd$„t_          ‰/j0        ¦  «        D ¦   «         }ˆ/fd%„t_          ‰/j1        ¦  «        D ¦   «         }|D ]0}t          j	        |
› d&|j2        › d'|j3        › d(|j4        › �¦  «         Œ1|D ]0}t          j	        |
› d)|j2        › d'|j3        › d(|j4        › �¦  «         Œ1|rì| 5                    ¦   «         }d|d         d*d*f}g |dd+…         ¢ˆfd,„|d+d…         D ¦   «         ¢R }|D ]x}tm          d-„ to          |j3        |¦  «        D ¦   «         ¦  «        } tm          d.„ to          |j3        |¦  «        D ¦   «         ¦  «        }!| 8                    |j2        | ||!¬/¦  «         Œy| 9                    |¦  «         |r|s| :                    |¦  «         t          j	        |
› d0|rd1nd2|rd3nd4z   › d5|› �¦  «         |r2|s0| (                    ‰0j)        j;        ¦  «         ‰0j<        j=        |_>        |s|s|r!|s| (                    ‰0j)        j?        ¦  «         |�rö|�só G ˆˆ0fd6„d7‰0j@        ¦  «        }" |"|tƒ          t          | ¦  «                             d8¦  «        ¦  «        ¬9¦  «        |_B        ˆ/fd:„t_          ‰/jC        ¦  «        D ¦   «         }#ˆ.fd;„|#D ¦   «         }$|$rd<t‰          |$¦  «        › d=�nd=}%d}&t_          ‰/jC        ¦  «        D ]ä}'‰/ E                    |'¦  «        }(|(j2         F                    |%¦  «        sŒ2|(jG        ‰0jH        jI        k    r#d>|(j2         J                    ¦   «         v r‰0jK        })nA|(jG        ‰0jH        jL        k    r+|(j2         M                    d?¦  «        s	d@|(j2        v r‰0jN        })nŒ«|)|(_O        t_          |(j1        ¦  «        D ]}*|( P                    |*|)¦  «         Œ|&dz  }&Œå|&r\t+          ‰0j)        dA¦  «        r‰0j)        jQ        n‰0j)        jR        }| (                    |¦  «         t          j	        |
› dB|&› dC�¦  «         t+          |dD¦  «        r| S                    ‰/|¦  «        }+n.| T                    ‰/|¦  «        }+|+€dn|+ U                    ¦   «         }+|+€t]          dE¦  «        ‚t­          |dF¦  «        5 },|�st¯          jX        |¦  «        }-|, Y                    tµ          |-¦  «         [                    dGdHd¬I¦  «        ¦  «         |, Y                    |- \                    ¦   «         ¦  «         |, Y                    |+¦  «         ddd¦  «         n# 1 swxY w Y   tƒ          |¦  «        S )Ja•  Export a YOLO model to TensorRT engine format.

    Args:
        onnx_file (str): Path to the ONNX file to be converted.
        output_file (Path | str | None): Path to save the generated TensorRT engine file.
        workspace (int | None): Workspace size in GB for TensorRT.
        quantize (int | str | None): Precision scheme, 16 for FP16 or 8 for INT8.
        dynamic (bool, optional): Enable dynamic input shapes.
        shape (tuple[int, int, int, int], optional): Input shape (batch, channels, height, width).
        dla (int | None): DLA core to use (Jetson devices only).
        dataset (ultralytics.data.build.InfiniteDataLoader, optional): Dataset for INT8 calibration, unused when the
            ONNX graph already carries Q/DQ ranges.
        metadata (dict | None): Metadata to include in the engine file.
        verbose (bool, optional): Enable verbose logging.
        prefix (str, optional): Prefix for log messages.

    Returns:
        (str): Path to the exported engine file.

    Raises:
        ValueError: If DLA is enabled on non-Jetson devices or required precision is not set.
        RuntimeError: If the ONNX file cannot be parsed.

    Notes:
        TensorRT version compatibility is handled for workspace size and engine building. On TensorRT 7-10, INT8
        calibration uses an ``IInt8Calibrator`` over ``dataset`` and writes a calibration cache, while FP16/INT8 are
        enabled with builder flags. On TensorRT 11 these were removed in favor of strongly-typed networks, so reduced
        precision is baked into the ONNX with NVIDIA ModelOpt before building (FP16 AutoCast, INT8 explicit Q/DQ) by
        `modelopt_quantize_onnx`. The TensorRT 7-10 path keeps the head Sigmoid layers in FP32 to preserve
        confidence-score calibration (see #24668) and the head's output convolutions in FP16 for accuracy. Metadata is
        serialized and written to the engine file if provided.
    r   Né   )Újetpackz10.15z>=7.0.0T)Úhardz!=10.2.0z5https://github.com/ultralytics/ultralytics/pull/24367)Úmsgú
z starting export with TensorRT z...z.enginei   @rS   r0   é
   é   Úset_memory_pool_limitÚplatform_has_fast_fp16é   Úplatform_has_fast_int8r€   c              3  ó,   K  — | ]}|j         d k    V — ŒdS )ÚQuantizeLinearN)Úop_type)rˆ   r´   s     r   rŠ   zonnx2engine.<locals>.<genexpr>*  s*   è è € ÐoÐo°ˆaŒiÐ+Ò+ÐoÐoÐoÐoÐoÐor   Fr�   z4INT8 TensorRT export requires a calibration dataset.z.DLA is only available on NVIDIA Jetson devicesz>=11.0.0,<11.1.0zLDLA is not supported in TensorRT 11.0; export with TensorRT 10.x to use DLA.z enabling DLA on core zeDLA requires either quantize=16 (FP16) or quantize=8 (INT8). Please enable one of them and try again.zfailed to load ONNX file: c                ó:   •— g | ]}‰                      |¦  «        ‘ŒS © )Ú	get_input©rˆ   ÚiÚnetworks     €r   ú
<listcomp>zonnx2engine.<locals>.<listcomp>K  s'   ø€ ÐFÐFÐF qˆg×Ò Ñ"Ô"ÐFÐFÐFr   c                ó:   •— g | ]}‰                      |¦  «        ‘ŒS rÎ   )Ú
get_outputrÐ   s     €r   rÓ   zonnx2engine.<locals>.<listcomp>L  s'   ø€ ÐIÐIÐI¨ˆw×!Ò! !Ñ$Ô$ÐIÐIÐIr   z input "z" with shapeú z	 output "é    r3   c              3  ó\   •K  — | ]&}t          t          d ‰pd ¦  «        |z  ¦  «        V — Œ'dS )r3   N)r   rŒ   )rˆ   r‰   r¹   s     €r   rŠ   zonnx2engine.<locals>.<genexpr>U  s>   øè è € Ð"VÐ"VÀq¥3¥s¨1¨i¨n¸1Ñ'=Ô'=ÀÑ'AÑ#BÔ#BÐ"VÐ"VÐ"VÐ"VÐ"VÐ"Vr   c              3  ó0   K  — | ]\  }}|d k    r|n|V — ŒdS ©r5   NrÎ   )rˆ   r‰   Úlos      r   rŠ   zonnx2engine.<locals>.<genexpr>W  ó2   è è € ÐXÐX±U°Q¸  b¢ ˜A˜A¨bÐXÐXÐXÐXÐXÐXr   c              3  ó0   K  — | ]\  }}|d k    r|n|V — ŒdS rÚ   rÎ   )rˆ   r‰   Úhis      r   rŠ   zonnx2engine.<locals>.<genexpr>X  rÜ   r   )r]   ÚoptrŒ   z
 building ÚINT8ÚFPÚ16Ú32z engine as c                  óN   •— e Zd ZdZ	 ddˆ ˆfd„Zdd	„Zdd„Zdd„Zdd„Zdd„Z	dS )ú%onnx2engine.<locals>.EngineCalibratora  Custom INT8 calibrator for TensorRT engine optimization.

            This calibrator provides the necessary interface for TensorRT to perform INT8 quantization calibration using
            a dataset. It handles batch generation, caching, and calibration algorithm selection.

            Attributes:
                dataset: Dataset for calibration.
                data_iter: Iterator over the calibration dataset.
                algo (trt.CalibrationAlgoType): Calibration algorithm type.
                batch (int): Batch size for calibration.
                cache (Path): Path to save the calibration cache.

            Methods:
                get_algorithm: Get the calibration algorithm to use.
                get_batch_size: Get the batch size to use for calibration.
                get_batch: Get the next batch to use for calibration.
                read_calibration_cache: Use existing cache instead of calibrating again.
                write_calibration_cache: Write calibration cache to disk.
            rx   Úcacher"   rQ   ÚNonec                óð   •— ‰j                              | ¦  «         || _        t          |¦  «        | _        ‰�‰j        j        n‰j        j        | _        |j	        | _
        t          |¦  «        | _        dS )z;Initialize the INT8 calibrator with dataset and cache path.N)ÚIInt8Calibratorr(   r°   ÚiterÚ	data_iterÚCalibrationAlgoTypeÚENTROPY_CALIBRATION_2ÚMINMAX_CALIBRATIONÚalgoÚ
batch_sizerµ   r   ræ   )r)   r°   ræ   r»   r   s      €€r   r(   z.onnx2engine.<locals>.EngineCalibrator.__init__  sr   ø€ ð Ô#×,Ò,¨TÑ2Ô2Ð2Ø&�”Ý!% g¡¤�”ð �ð Ô+ÔAÐAàÔ0ÔCð ”	ð
 %Ô/�”
Ý! %™[œ[�”
�
�
r   útrt.CalibrationAlgoTypec                ó   — | j         S )z%Get the calibration algorithm to use.)rï   ©r)   s    r   Úget_algorithmz3onnx2engine.<locals>.EngineCalibrator.get_algorithm�  s
   € à”yÐ r   r   c                ó   — | j         pdS )z*Get the batch size to use for calibration.r0   )rµ   ró   s    r   Úget_batch_sizez4onnx2engine.<locals>.EngineCalibrator.get_batch_size”  s   € à”z� QÐ&r   úlist[int] | Nonec                óò   — 	 t          | j        ¦  «        d         dz  }|j        j        dk    r|                     d¦  «        n|}t          |                     ¦   «         ¦  «        gS # t          $ r Y dS w xY w)zOGet the next batch to use for calibration, as a list of device memory pointers.rƒ   r„   r�   ÚcudaN)Únextrë   r/   Útyper¢   r   Údata_ptrÚStopIteration)r)   ÚnamesÚim0ss      r   Ú	get_batchz/onnx2engine.<locals>.EngineCalibrator.get_batch˜  s}   € ð Ý ¤Ñ/Ô/°Ô6¸Ñ>�DØ.2¬kÔ.>À%Ò.GÐ.G˜4Ÿ7š7 6™?œ?˜?ÈT�DÝ §¢¡¤Ñ0Ô0Ð1Ð1øÝ$ð  ð  ð  à˜4˜4ð øøøs   ‚A%A( Á(
A6Á5A6úbytes | Nonec                óŽ   — | j                              ¦   «         r)| j         j        dk    r| j                              ¦   «         S dS dS )zSUse existing cache instead of calibrating again, otherwise, implicitly return None.ú.cacheN)ræ   ÚexistsÚsuffixÚ
read_bytesró   s    r   Úread_calibration_cachez<onnx2engine.<locals>.EngineCalibrator.read_calibration_cache¢  sK   € à”:×$Ò$Ñ&Ô&ð 3¨4¬:Ô+<ÀÒ+HÐ+HØœ:×0Ò0Ñ2Ô2Ð2ð3ð 3Ð+HÐ+Hr   Úbytesc                ó:   — | j                              |¦  «        }dS )z Write calibration cache to disk.N)ræ   Úwrite_bytes)r)   ræ   rE   s      r   Úwrite_calibration_cachez=onnx2engine.<locals>.EngineCalibrator.write_calibration_cache§  s   € à”J×*Ò*¨5Ñ1Ô1���r   N)rx   )ræ   r"   rQ   rç   )rQ   rñ   )rQ   r   )rQ   r÷   )rQ   r  )ræ   r  rQ   rç   )
rJ   rK   rL   rM   r(   rô   rö   r   r  r  )r»   r   s   €€r   ÚEngineCalibratorrå   j  s°   ø€ € € € € ðð ð.  ð)ð )ð )ð )ð )ð )ð )ð )ð"!ð !ð !ð !ð'ð 'ð 'ð 'ð ð  ð  ð  ð3ð 3ð 3ð 3ð
2ð 2ð 2ð 2ð 2ð 2r   r  r  )r°   ræ   c                óD   •— g | ]}‰                      |¦  «        j        ‘ŒS rÎ   )Ú	get_layerrŸ   rÐ   s     €r   rÓ   zonnx2engine.<locals>.<listcomp>·  s*   ø€ ÐNÐNÐN¨q�×"Ò" 1Ñ%Ô%Ô*ÐNÐNÐNr   c                ó‚   •— g | ];}t          j        d |¦  «        xŠ¯t          ‰                     d¦  «        ¦  «        ‘Œ<S )z/model\.(\d+)/r0   )ÚreÚmatchr   Úgroup)rˆ   r´   Úms     €r   rÓ   zonnx2engine.<locals>.<listcomp>¸  sA   ø€ Ð[Ð[Ð[ q½2¼8ÐDUÐWXÑ;YÔ;YÐ6Y°aÐ[•3�q—w’w˜q‘z”z‘?”?Ð[Ð[Ð[r   z/model.ú/Úsigmoidz.2/Convz/dfl/ÚOBEY_PRECISION_CONSTRAINTSz	 keeping z% head layers out of INT8 for accuracyÚbuild_serialized_networkz3TensorRT engine build failed, check logs for errorsÚwbr1   Úlittle)Ú	byteorderÚsigned)]rO   r   r   r   r   ÚImportErrorr   Ú__version__r   r¥   r   r¡   r   r   ÚVERBOSEr   Úmin_severityÚBuilderÚcreate_builder_configr   r[   ÚhasattrrÆ   ÚMemoryPoolTypeÚ	WORKSPACEÚmax_workspace_sizeÚNetworkDefinitionCreationFlagÚEXPLICIT_BATCHÚcreate_networkÚgetattrÚanyrœ   r�   Únoder›   r   Ú
DeviceTypeÚDLAÚdefault_device_typeÚDLA_coreÚset_flagÚBuilderFlagÚGPU_FALLBACKr·   Ú
OnnxParserÚparse_from_fileÚRuntimeErrorÚrangeÚ
num_inputsÚnum_outputsrŸ   r=   r.   Úcreate_optimization_profiler7   ÚzipÚ	set_shapeÚadd_optimization_profileÚset_calibration_profilerà   ÚProfilingVerbosityÚDETAILEDÚprofiling_verbosityÚFP16ré   r"   Úint8_calibratorÚ
num_layersrŒ   r  Ú
startswithrû   Ú	LayerTypeÚ
ACTIVATIONÚlowerr£   ÚCONVOLUTIONÚendswithÚfloat16Ú	precisionÚset_output_typer  ÚSTRICT_TYPESr  Úbuild_engineÚ	serializeÚopenÚjsonÚdumpsÚwriteÚlenÚto_bytesÚencode)1ry   re   r¹   rz   rj   r=   r»   r°   r¼   r½   r~   rO   ÚloggerÚbuilderÚconfigÚworkspace_bytesÚ	trt_majorÚis_trt10Úis_trt11ÚflagÚuse_fp16Úuse_int8ÚqdqÚ	calibrateÚparserÚinputsÚoutputsÚinpÚoutÚprofileÚ	min_shapeÚ	max_shapeÚinp_minÚinp_maxr  rþ   ÚindicesÚheadÚcountrÑ   Úlayerr.   ÚjÚengineÚtÚmetar  rÒ   r   s1     `   `                                       @@@r   Úonnx2engineru  Õ   sA
  øøøøø€ ðZ €K€K€Kõ ˜ÐÑÔð  �v™xœxð  Ý�wÑÔÐðØÐÐÐÐøÝð ð ð ÝÑÔÐØÐÐÐÐÐðøøøõ �#”/ 9°4Ð8Ñ8Ô8Ð8Ý�#”/ :Ð3jÐkÑkÔkÐkå
„KÐP�VÐPÐP¸C¼OÐPÐPÐPÑQÔQÐQØÐG¥ i¡¤×!<Ò!<¸YÑ!GÔ!G€Kå Ñ"Ô"€FØ07ÐL˜#œ*Ô,Ð,¸S¼Z¼_€FÔð �kŠk˜&Ñ!Ô!€GØ×*Ò*Ñ,Ô,€FÝ˜9˜>¨¨gÑ6Ñ7Ô7€OÝ�C”O×)Ò)¨#¨qÑ1Ô1°!Ô4Ñ5Ô5€IØ˜BŠ€Hà˜BŠ€HØ˜ÒÐÝ�6Ð2Ñ3Ô3ð 	8Ø×(Ò(¨Ô);Ô)EÀÑWÔWÐWÐWà(7ˆFÔ%àÐZˆ1ˆ1˜q¥C¨Ô(IÔ(XÑ$YÔ$YÑY€DØ×$Ò$ TÑ*Ô*€Gå�wÐ 8¸$Ñ?Ô?ÐRÀHÐPRÂN€HÝ�wÐ 8¸$Ñ?Ô?ÐQÀHÐPQÂM€HÝ
ÐoÐo°T·Y²Y¸yÐ]b°YÑ5cÔ5cÔ5iÔ5nÐoÑoÔoÑ
oÔ
o€CØÐ$ ˜W€IØð Q�W�_ÝÐOÑPÔPÐPð €Ýð 	OÝÐMÑNÔNÐNÝ˜œÐ*<Ñ=Ô=ð 	mõ ÐkÑlÔlÐlÝŒ�vÐ=Ð=°SÐ=Ð=Ð=Ñ>Ô>Ð>Øð 	 ð 	ÝØwñô ð ð &)¤^Ô%7ˆÔ"Ý˜c™(œ(ˆŒØ�Š˜œÔ4Ñ5Ô5Ð5ð ð p�Xð p ð pÝ*¨9¸CÐ6M°b°bÀXÈwÐX]Ð_fÐhnÑoÔoˆ	ð �^Š^˜G VÑ,Ô,€FØ×!Ò! )Ñ,Ô,ð EÝÐC¸	ÐCÐCÑDÔDÐDð GÐFÐFÐF­E°'Ô2DÑ,EÔ,EÐFÑFÔF€FØIÐIÐIÐI­e°GÔ4GÑ.HÔ.HÐIÑIÔI€GØð Vð VˆÝŒ�vÐTÐT s¤xÐTÐT¸S¼YÐTÐTÈÌÐTÐTÑUÔUÐUÐUØð Wð WˆÝŒ�vÐUÐU¨¬ÐUÐU¸c¼iÐUÐUÈ#Ì)ÐUÐUÑVÔVÐVÐVàð 
4Ø×5Ò5Ñ7Ô7ˆØ˜˜aœ " bÐ)ˆ	ØW�e˜B˜Q˜B”iÐWÐ"VÐ"VÐ"VÐ"VÈEÐRSÐRTÐRTÌIÐ"VÑ"VÔ"VÐWÐWˆ	Øð 	Mð 	MˆCÝÐXÐX½cÀ#Ä)ÈYÑ>WÔ>WÐXÑXÔXÑXÔXˆGÝÐXÐX½cÀ#Ä)ÈYÑ>WÔ>WÐXÑXÔXÑXÔXˆGØ×Ò˜cœh¨G¸ÀGÐÑLÔLÐLÐLØ×'Ò'¨Ñ0Ô0Ð0Øð 	4˜Xð 	4Ø×*Ò*¨7Ñ3Ô3Ð3å
„KØÐrÐr xÐX˜V˜V°TÀXÐ=W¸T¸TÐSWÑ5XÐrÐrÐepÐrÐrñô ð ð ð E˜ð EØ�Š˜œÔ,Ñ-Ô-Ð-Ø%(Ô%;Ô%DˆÔ"Øð .�Hð . ð .¨Xð .Ø�Š˜œÔ,Ñ-Ô-Ð-ð ñ hZ˜ñ hZð?	2ð ?	2ð ?	2ð ?	2ð ?	2ð ?	2ð ?	2ð ?	2˜sÔ2ñ ?	2ô ?	2ð ?	2ðD "2Ð!1ØÝ•d˜9‘o”o×1Ò1°(Ñ;Ô;Ñ<Ô<ð"
ñ "
ô "
ˆÔð OÐNÐNÐNµE¸'Ô:LÑ4MÔ4MÐNÑNÔNˆØ[Ð[Ð[Ð[¨EÐ[Ñ[Ô[ˆØ,3Ð<Ð(�˜W™œÐ(Ð(Ð(Ð(¸ˆØˆÝ�wÔ)Ñ*Ô*ð 	ð 	ˆAØ×%Ò% aÑ(Ô(ˆEØ”:×(Ò(¨Ñ.Ô.ð ØØŒz˜Sœ]Ô5Ò5Ð5¸)ÀuÄz×GWÒGWÑGYÔGYÐ:YÐ:YØœ��Ø”˜sœ}Ô8Ò8Ð8¸e¼j×>QÒ>QÐR[Ñ>\Ô>\Ð8Ð`gÐkpÔkuÐ`uÐ`uØœ��àØ#ˆEŒOÝ˜5Ô,Ñ-Ô-ð 0ð 0�Ø×%Ò% a¨Ñ/Ô/Ð/Ð/Ø�Q‰JˆEˆEØð 	Zõ ˜3œ?Ð,HÑIÔIð2�”Ô:Ð:à”_Ô1ð ð
 �OŠO˜DÑ!Ô!Ð!ÝŒK˜6ÐXÐX¨EÐXÐXÐXÑYÔYÐYõ ˆwÐ2Ñ3Ô3ð @Ø×1Ò1°'¸6ÑBÔBˆˆà×%Ò% g¨vÑ6Ô6ˆØ˜��¨V×-=Ò-=Ñ-?Ô-?ˆØ€~ÝÐPÑQÔQÐQÝ	ˆk˜4Ñ	 Ô	 ð  AØÐÝ”:˜hÑ'Ô'ˆDØ�GŠG•C˜‘I”I×&Ò& q°HÀTÐ&ÑJÔJÑKÔKÐKØ�GŠG�D—K’K‘M”MÑ"Ô"Ð"Ø	�Š�‰Œˆðð ð ñ ô ð ð ð ð ð ð øøøð ð ð ð õ ˆ{ÑÔÐs"   ¸= ½AÁAß3Bb
â
bâb)rO   rP   rQ   r   )rV   NNN)r   r   rc   rd   re   rf   ra   r   rg   rh   ri   rh   rj   rk   rQ   r"   )NNru   Frx   )ry   r"   rz   r{   r=   r|   rj   r}   r~   r"   rQ   r"   )
NNNFru   NNNFrx   )ry   r"   re   r¸   r¹   rº   rz   r{   rj   r}   r=   r|   r»   rº   r¼   rk   r½   r}   r~   r"   rQ   r"   )$Ú
__future__r   rQ  r  ÚtypesÚ	functoolsr   Úpathlibr   r¨   r¦   r«   r9   Úultralytics.utilsr   r   r   r	   r
   r   r   r   Úultralytics.utils.checksr   r   r   Úultralytics.utils.torch_utilsr   r   ÚnnÚModuler   rb   rt   r·   ru  rÎ   r   r   ú<module>r     sÿ  ðð #Ð "Ð "Ð "Ð "Ð "à €€€Ø 	€	€	€	Ø €€€Ø Ð Ð Ð Ð Ð Ø Ð Ð Ð Ð Ð à 
€
€
€
Ø Ð Ð Ð Ø €€€à rÐ rÐ rÐ rÐ rÐ rÐ rÐ rÐ rÐ rÐ rÐ rÐ rÐ rÐ rÐ rÐ rÐ rÐ rÐ rØ VÐ VÐ VÐ VÐ VÐ VÐ VÐ VÐ VÐ VØ 3Ð 3Ð 3Ð 3Ð 3Ð 3ð ð'ð 'ñ „ð'ðFð Fð Fð Fð F�u”x”ñ Fô Fð FðD6ð 6ð 6ð 6ð( €ÑÔð
 Ø$(Ø%)Øð&ð &ð &ð &ñ Ôð&ðV "&ØØ'7ØØðVð Vð Vð Vð Vðv &*Ø Ø!%ØØ'7ØØØ ØØðKð Kð Kð Kð Kð Kð Kr   