§
    l¬¸j«  ã                  ó¬   — d Z ddlmZ ddlZddlZddlmZ ddlm	Z	 ddl
mZ ddlmZ  ed„ d	D ¦   «         ¦  «        Z G d
„ d¦  «        Zddd„Zdd„ZdS )uÊ  Shared ReID encoder used by BoT-SORT, Deep OC-SORT, and TrackTrack.

* `.pt` YOLO checkpoints â€” loaded via `YOLO()`; embeddings are pulled from the second-to-last
layer through the predictor's `embed=[...]` argument (works with classification and ReID
backbones).
* Compatible exported models (`.torchscript`, `.onnx`, `.engine`, OpenVINO model directories, â€¦) â€”
loaded via `AutoBackend`; the model is expected to output the embedding tensor directly.
é    )ÚannotationsN)ÚAutoBackend)Ú	xywh2xyxy©Úsave_one_box)Úsmart_inference_modec              #  ó"   K  — | ]
}d |› d�V — ŒdS )Úyolo26z
-reid.onnxN© )Ú.0Úks     úo/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/ultralytics/trackers/utils/reid.pyú	<genexpr>r      s.   è è € Ð@Ð@°1Ð. Ð.Ð.Ð.Ð@Ð@Ð@Ð@Ð@Ð@ó    Únsmlxc                  ó�   — e Zd ZdZ ed¦  «        ddd„¦   «         Zedd„¦   «         Zdd„Z e	j
        ¦   «         dd„¦   «         ZdS )ÚReIDzXReID encoder. Routes `.pt` to the YOLO predictor path; everything else to `AutoBackend`.Féà   NÚmodelÚstrÚimgszÚintÚdeviceústr | torch.device | NoneÚfp16Úboolc                óP  — || _         d| _        |�t          j        |¦  «        n3t          j        t          j                             ¦   «         rdnd¦  «        | _        t          |¦  «                             d¦  «        | _        | j        r]ddl	m
}  ||¦  «        | _        |                      t          | j        j        j        ¦  «        dz
  g| j        dd¬	¦  «         d| _        dS dd
lm}  |t          |¦  «        ¦  «        j        t"          v rddlm}  |t          |¦  «        ¦  «        }t)          t          |¦  «        | j        |d¬¦  «        | _        | j        j        | _        t+          | j        dd¦  «        }|�|                     ¦   «         d         j        nd}	t          |	¦  «        dk    rjt1          |	d         t2          ¦  «        r|	d         dk    r|	d         | _        t1          |	d         t2          ¦  «        r|	d         dk    r|	d         | _         dS dS dS dS )aL  Initialize encoder for re-identification.

        Args:
            model (str): Path to a ReID model. `.pt` runs through the YOLO predictor (embed-layer extraction); other
                extensions go through `AutoBackend`.
            imgsz (int): Square input size used for crop preprocessing on the AutoBackend path. Overridden by the
                model's own static input size when one is detected.
            device (str | torch.device | None): Inference device; defaults to CUDA if available.
            fp16 (bool): Use half precision when the backend supports it.
        NÚcudaÚcpuz.ptr   )ÚYOLOé   F)Úembedr   ÚverboseÚsave)ÚPath)Úattempt_download_asset)r   r   r#   Úsessionr   é   )r   Ú
batch_sizeÚtorchr   r   Úis_availabler   ÚendswithÚis_ptÚultralyticsr    r   Úlenr   Úpathlibr%   ÚnameÚREID_ASSETSÚultralytics.utils.downloadsr&   r   ÚgetattrÚ
get_inputsÚshapeÚ
isinstancer   )
Úselfr   r   r   r   r    r%   r&   r'   r6   s
             r   Ú__init__zReID.__init__   s  € ð ˆŒ
ØˆŒà$*Ð$6�EŒL˜Ñ Ô Ð ½E¼LÕSXÔS]×SjÒSjÑSlÔSlÐIwÈÈÐrwÑ<xÔ<xð 	Œõ ˜‘Z”Z×(Ò(¨Ñ/Ô/ˆŒ
àŒ:ð 	*Ø(Ð(Ð(Ð(Ð(Ð(à˜˜e™œˆDŒJà�JŠJ�c $¤*Ô"2Ô"8Ñ9Ô9¸AÑ=Ð>ÀtÄ{Ð\aÐhmˆJÑnÔnÐnØˆDŒIˆIˆIà$Ð$Ð$Ð$Ð$Ð$àˆt•C˜‘J”JÑÔÔ$­Ð3Ð3ØNÐNÐNÐNÐNÐNà.Ð.­s°5©z¬zÑ:Ô:�Ý$¥S¨¡Z¤Z¸¼È$ÐX]Ð^Ñ^Ô^ˆDŒJØœ
œˆDŒIõ ˜dœj¨)°TÑ:Ô:ˆGØ5<Ð5H�G×&Ò&Ñ(Ô(¨Ô+Ô1Ð1ÈbˆEÝ�5‰zŒz˜QŠˆÝ˜e Aœh­Ñ,Ô,ð /°°q´¸A²°Ø&+¨A¤h�D”OÝ˜e Aœh­Ñ,Ô,ð *°°q´¸A²°Ø!& q¤�D”J�J�Jð	 ˆð*ð *°°r   Úimgú
np.ndarrayÚdetsÚreturnúlist[np.ndarray]c           	     óv   ‡ — ˆ fd„t          t          j        |dd…dd…f         ¦  «        ¦  «        D ¦   «         S )a  Crop detection regions from image, converting xywh to xyxy first.

        Args:
            img (np.ndarray): BGR image.
            dets (np.ndarray): Detections in xywh format (first 4 columns used).

        Returns:
            (list[np.ndarray]): Cropped image patches.
        c                ó4   •— g | ]}t          |‰d ¬¦  «        ‘ŒS )F)r$   r   )r   Údetr:   s     €r   ú
<listcomp>z)ReID._crop_detections.<locals>.<listcomp>S   s(   ø€ ÐgÐgÐg°s•˜S #¨EÐ2Ñ2Ô2ÐgÐgÐgr   Nr(   )r   r*   Ú
from_numpy)r:   r<   s   ` r   Ú_crop_detectionszReID._crop_detectionsH   sL   ø€ ð hÐgÐgÐg½iÍÔHXÐY]Ð^_Ð^_Ð^_ÐacÐbcÐacÐ^cÔYdÑHeÔHeÑ>fÔ>fÐgÑgÔgÐgr   Úcropsútorch.Tensorc           
     ód  — t          j        t          |¦  «        d| j        | j        t           j        ¬¦  «        }t          |¦  «        D ]°\  }}t          j        t          j        |dddd…f         ¦  «        ¦  «         	                    ddd¦  «         
                    d¦  «                             ¦   «         d	z  }t           j        j                             || j        | j        fd
d¬¦  «        d         ||<   Œ±|                     | j        ¦  «        }| j        r|                     ¦   «         n|S )zTStack a list of valid image crops into a normalized BCHW float tensor at self.imgsz.é   ©Údtype.Néÿÿÿÿr!   r   é   g     ào@ÚbilinearF)ÚsizeÚmodeÚalign_corners)r*   Úemptyr/   r   Úfloat32Ú	enumeraterC   ÚnpÚascontiguousarrayÚpermuteÚ	unsqueezeÚfloatÚnnÚ
functionalÚinterpolateÚtor   r   Úhalf)r8   rE   ÚbatchÚiÚcÚts         r   Ú_crops_to_tensorzReID._crops_to_tensorU   s  € å”�C ™JœJ¨¨4¬:°t´zÍÌÐWÑWÔWˆÝ˜eÑ$Ô$ð 	ð 	‰DˆAˆqÝÔ ¥Ô!5°a¸¸T¸T¸r¸T¸	´lÑ!CÔ!CÑDÔD×LÒLÈQÐPQÐSTÑUÔU×_Ò_Ð`aÑbÔb×hÒhÑjÔjÐmrÑrˆAÝ”xÔ*×6Ò6Ø˜œ T¤ZÐ0°zÐQVð 7ñ ô àôˆE�!‰HˆHð —’˜œÑ%Ô%ˆØ#œyÐ3ˆu�zŠz‰|Œ|ˆ|¨eÐ3r   úlist[np.ndarray | None]c           
     óú  ‡— |                       ||¦  «        }d„ |D ¦   «         }d„ t          ||¦  «        D ¦   «         }|sdgt          |¦  «        z  S | j        rt| j                             |¦  «        }t          |¦  «        t          |¦  «        k    r,|d         j        d         t          |¦  «        k    r|d         }d„ |D ¦   «         Š�n|                      |¦  «        }| j        |j        d         }	}|�|	|k    r|                      |¦  «        }n¿g }
t          d|	|¦  «        D ]Ž}||||z   …         }|j        d         |k     rFt          j        | |dd…         j        ||j        d         z
  g|j        dd…         ¢R Ž gd¦  «        }|
                     |                      |¦  «        ¦  «         Œ�t          j        |
d¦  «        d|	…         }d„ |D ¦   «         Št          ‰¦  «        Šˆfd	„|D ¦   «         S )
z(Extract embeddings for detected objects.c                ó6   — g | ]}t          |j        ¦  «        ‘ŒS r   )r   rN   )r   r`   s     r   rB   z!ReID.__call__.<locals>.<listcomp>d   s    € Ð-Ð-Ð- !•�a”f‘”Ð-Ð-Ð-r   c                ó   — g | ]	\  }}|¯|‘Œ
S r   r   )r   ÚcropÚkeeps      r   rB   z!ReID.__call__.<locals>.<listcomp>e   s!   € ÐHÐHÐH¡
  dÀ4ÐH�tÐHÐHÐHr   Nr   c                óZ   — g | ](}|                      ¦   «                              ¦   «         ‘Œ)S r   ©r   Únumpy©r   Úfs     r   rB   z!ReID.__call__.<locals>.<listcomp>m   ó(   € Ð:Ð:Ð:¨q˜1Ÿ5š5™7œ7Ÿ=š=™?œ?Ð:Ð:Ð:r   rK   rL   c                óZ   — g | ](}|                      ¦   «                              ¦   «         ‘Œ)S r   rj   rl   s     r   rB   z!ReID.__call__.<locals>.<listcomp>{   rn   r   c                ó6   •— g | ]}|rt          ‰¦  «        nd ‘ŒS ©N)Únext)r   rh   Úvalid_featss     €r   rB   z!ReID.__call__.<locals>.<listcomp>~   s+   ø€ ÐFÐFÐF¸ TÐ3•�[Ñ!Ô!Ð!¨tÐFÐFÐFr   )rD   Úzipr/   r-   r   Ú	predictorr6   rb   r)   Úranger*   ÚcatÚexpandÚappendÚiter)r8   r:   r<   rE   ÚvalidÚvalid_cropsÚfeatsr^   ÚbsÚnÚoutsÚsÚchunkrs   s                @r   Ú__call__zReID.__call__`   s#  ø€ ð ×%Ò% c¨4Ñ0Ô0ˆØ-Ð- uÐ-Ñ-Ô-ˆØHÐH­c°%¸Ñ.?Ô.?ÐHÑHÔHˆØð 	'Ø�6�C ™JœJÑ&Ð&àŒ:ð 	;Ø”J×(Ò(¨Ñ5Ô5ˆEÝ�5‰zŒz�S Ñ-Ô-Ò-Ð-°%¸´(´.ÀÔ2CÅsÈ;ÑGWÔGWÒ2WÐ2WØ˜aœ�Ø:Ð:°EÐ:Ñ:Ô:ˆK‰Kà×)Ò)¨+Ñ6Ô6ˆEØ”O U¤[°¤^�ˆBØˆz˜Q "šW˜WØŸ
š
 5Ñ)Ô)��à�Ý˜q ! R™œð 3ð 3�AØ! ! a¨"¡f *Ô-�EØ”{ 1”~¨Ò*Ð*Ý %¤	¨5Ð2C°%¸¸¸´*Ô2CÀBÈÌÐUVÌÑDWÐ2jÐZ_ÔZeÐfgÐfhÐfhÔZiÐ2jÐ2jÐ2jÐ*kÐmnÑ oÔ o˜Ø—K’K §
¢
¨5Ñ 1Ô 1Ñ2Ô2Ð2Ð2Ýœ	 $¨Ñ*Ô*¨2¨A¨2Ô.�Ø:Ð:°EÐ:Ñ:Ô:ˆKå˜;Ñ'Ô'ˆØFÐFÐFÐFÀÐFÑFÔFÐFr   )r   NF)r   r   r   r   r   r   r   r   )r:   r;   r<   r;   r=   r>   )rE   r>   r=   rF   )r:   r;   r<   r;   r=   rc   )Ú__name__Ú
__module__Ú__qualname__Ú__doc__r   r9   ÚstaticmethodrD   rb   r*   Úno_gradrƒ   r   r   r   r   r      s°   € € € € € ØbÐbàÐ˜%Ñ Ô ð**ð **ð **ð **ñ !Ô ð**ðX ð
hð 
hð 
hñ „\ð
hð	4ð 	4ð 	4ð 	4ð €U„]�_„_ðGð Gð Gñ „_ðGð Gð Gr   r   Ú	with_reidr   r   ú
str | Noner   r   c                óB   — | sdS |dk    rd„ }|S t          ||¬¦  «        S )aR  Return a ReID encoder, the native-features pass-through, or None.

    Args:
        with_reid (bool): Whether ReID is enabled at all.
        model (str | None): `"auto"` returns a callable that converts pre-extracted backbone features to numpy arrays;
            any other value loads a `ReID` model from that path. Ignored when `with_reid` is False.
        device (str | torch.device | None): Inference device for the ReID model; defaults to CUDA if available.

    Returns:
        (Callable | None): A `(img, dets) -> list[np.ndarray | None]` encoder, or None when ReID is disabled.
    NÚautoc                óf   — t          | t          j        ¦  «        rd„ | D ¦   «         S d„ | D ¦   «         S )Nc                ó   — g | ]}|‘ŒS r   r   rl   s     r   rB   z8build_encoder.<locals>._auto_encoder.<locals>.<listcomp>“   s   € Ð)Ð)Ð)˜a˜Ð)Ð)Ð)r   c                óZ   — g | ](}|                      ¦   «                              ¦   «         ‘Œ)S r   rj   rl   s     r   rB   z8build_encoder.<locals>._auto_encoder.<locals>.<listcomp>”   s(   € Ð3Ð3Ð3¨�A—E’E‘G”G—M’M‘O”OÐ3Ð3Ð3r   )r7   rT   Úndarray)r}   Ú_detss     r   Ú_auto_encoderz$build_encoder.<locals>._auto_encoder‘   s>   € Ý˜%¥¤Ñ,Ô,ð *Ø)Ð) 5Ð)Ñ)Ô)Ð)Ø3Ð3¨UÐ3Ñ3Ô3Ð3r   )r   )r   )rŠ   r   r   r“   s       r   Úbuild_encoderr”   �   sD   € ð ð ØˆtØ�‚€ð	4ð 	4ð 	4ð
 ÐÝ�˜fÐ%Ñ%Ô%Ð%r   Úfeatr;   Úsmoothúnp.ndarray | NoneÚalpharX   r=   ú+tuple[np.ndarray | None, np.ndarray | None]c                ó2  — t          j        | t           j        ¬¦  «        } t           j                             | ¦  «        }|dk     rd|fS | |z  } |€| |                      ¦   «         fS ||z  d|z
  | z  z   }| |t           j                             |¦  «        z  fS )a  L2-normalize `feat` and blend it into `smooth` via exponential moving average.

    Args:
        feat (np.ndarray): New (un-normalized) appearance feature.
        smooth (np.ndarray | None): Current smoothed feature, or None on the first update.
        alpha (float): EMA weight on the existing `smooth` (``1.0`` keeps it unchanged).

    Returns:
        curr (np.ndarray | None): The normalized float32 feature, or None when `feat` has zero norm.
        smooth (np.ndarray | None): The updated, normalized float32 feature.
    rI   gê-�™—q=NrL   )rT   ÚasarrayrR   ÚlinalgÚnormÚcopy)r•   r–   r˜   r�   s       r   Úsmooth_featurerŸ   š   s–   € õ Œ:�d¥"¤*Ð-Ñ-Ô-€DÝŒ9�>Š>˜$ÑÔ€DØˆe‚|€|Ø�Vˆ|ÐØ�$‰;€DØ€~Ø�T—Y’Y‘[”[Ð Ð Ø�V‰^˜q 5™y¨DÑ0Ñ0€FØ��"œ)Ÿ.š.¨Ñ0Ô0Ñ0Ð0Ð0r   rq   )rŠ   r   r   r‹   r   r   )r•   r;   r–   r—   r˜   rX   r=   r™   )r‡   Ú
__future__r   rk   rT   r*   Úultralytics.nn.autobackendr   Úultralytics.utils.opsr   Úultralytics.utils.plottingr   Úultralytics.utils.torch_utilsr   Ú	frozensetr2   r   r”   rŸ   r   r   r   ú<module>r¦      s  ððð ð #Ð "Ð "Ð "Ð "Ð "à Ð Ð Ð Ø €€€à 2Ð 2Ð 2Ð 2Ð 2Ð 2Ø +Ð +Ð +Ð +Ð +Ð +Ø 3Ð 3Ð 3Ð 3Ð 3Ð 3Ø >Ð >Ð >Ð >Ð >Ð >àˆiÐ@Ð@¸Ð@Ñ@Ô@Ñ@Ô@€ðfGð fGð fGð fGð fGñ fGô fGð fGðR&ð &ð &ð &ð &ð21ð 1ð 1ð 1ð 1ð 1r   