§
    ­¸jú¾  ã                  ó’  — d 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	Z	ddl	m
Z
mZ erddlmZmZ ddlmZmZ ddlmZ dd	lmZm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! ddl"m#Z# ddl$m%Z% ddl&m'Z' ddl(m)Z)m*Z* ddl+m,Z, ddl-m.Z.m/Z/  e,¦   «         Z0d1d„Z1d2d„Z2d3d„Z3 G d „ d!ej4        ¦  «        Z5d4d%„Z6d5d'„Z7defd6d.„Z8efd7d0„Z9dS )8z$LW-DETR model and criterion classes.é    )ÚannotationsN)ÚTYPE_CHECKINGÚAnyÚcast)ÚTensorÚnn)ÚModelConfigÚTrainConfig)ÚMODEL_DEFAULTSÚModelDefaults)ÚBuilderArgs)ÚJoinerÚbuild_backbone)ÚSetCriterionÚ	dice_lossÚdice_loss_jitÚposition_supervised_lossÚsigmoid_ce_lossÚsigmoid_ce_loss_jitÚsigmoid_focal_lossÚsigmoid_varifocal_loss)ÚSegmentationHead)Úbuild_matcher)ÚMLP)ÚPostProcess)ÚTransformerÚbuild_transformer)Ú
get_logger)ÚNestedTensorÚnested_tensor_from_tensor_listÚlinearú	nn.LinearÚnum_classesÚintÚreturnc                óÖ  — | j         j        d         }t          t          j        ||z  ¦  «        ¦  «        }| j                              ¦   «                              |d¦  «        d|…         }| j        �4| j                             ¦   «                              |¦  «        d|…         nd}t          j	        | j
        ||du¬¦  «        }t          j        ¦   «         5  |j                              |¦  «         |�!|j        �|j                             |¦  «         ddd¦  «         n# 1 swxY w Y   | j         j        |j         _        | j        �|j        �| j        j        |j        _        |S )u¹  Return a new :class:`~torch.nn.Linear` resized to *num_classes* outputs.

    Tiles the existing weight rows when *num_classes* is larger than the current output size, or truncates them when
    smaller.  The returned module has ``out_features == num_classes`` so that ``nn.Linear`` metadata stays consistent
    with the actual weight shape â€” a requirement for correct ONNX export and ``torch.jit.trace`` serialisation.

    Args:
        linear: Source linear layer whose weights are used as the starting point.
        num_classes: Target number of output features.

    Returns:
        A new :class:`~torch.nn.Linear` with ``in_features`` unchanged and ``out_features == num_classes``.
    r   é   N)Úbias)ÚweightÚshaper$   ÚmathÚceilÚdetachÚrepeatr(   r   ÚLinearÚin_featuresÚtorchÚno_gradÚcopy_Úrequires_grad)r!   r#   ÚbaseÚnum_repeatsÚ
new_weightÚnew_biasÚ
new_linears          úd/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/rfdetr/models/lwdetr.pyÚ_resize_linearr;   :   s‚  € ð Œ=Ô˜qÔ!€DÝ•d”i ¨dÑ 2Ñ3Ô3Ñ4Ô4€KØ”×%Ò%Ñ'Ô'×.Ò.¨{¸AÑ>Ô>¸|À¸|ÔL€JØIOÌÐI`ˆvŒ{×!Ò!Ñ#Ô#×*Ò*¨;Ñ7Ô7¸¸¸ÔEÐEÐfj€HÝ”˜6Ô-¨{ÀÐQUÐAUÐVÑVÔV€Jå	Œ‰Œð ,ð ,ØÔ×Ò 
Ñ+Ô+Ð+ØÐ J¤OÐ$?ØŒO×!Ò! (Ñ+Ô+Ð+ð,ð ,ð ,ñ ,ô ,ð ,ð ,ð ,ð ,ð ,ð ,øøøð ,ð ,ð ,ð ,ð '-¤mÔ&A€JÔÔ#Ø„{Ð :¤?Ð#>Ø(.¬Ô(Aˆ
ŒÔ%ØÐs   Ã>D$Ä$D(Ä+D(Ú	parameterúnn.ParameterÚnum_rowsc                óÌ  — | j         d         }||k    r| S |dk    r9|                      ¦   «                              |g| j         dd…         ¢R ¦  «        }ndt          t	          j        ||z  ¦  «        ¦  «        gdg|                      ¦   «         dz
  z  ¢} |                      ¦   «         j        |Ž d|…         }t          j	        | 
                    ¦   «         | j        ¬¦  «        S )zQReturn a parameter with the first dimension resized by tiling or truncating rows.r   r'   N)r4   )r*   r-   Ú	new_zerosr$   r+   r,   Údimr.   r   Ú	ParameterÚcloner4   )r<   r>   Úcurrent_rowsÚnew_dataÚrepeatss        r:   Ú_resize_parameter_rowsrG   X   sà   € à”? 1Ô%€LØ�xÒÐØÐà�qÒÐØ×#Ò#Ñ%Ô%×/Ò/°Ð0P¸I¼OÈAÈBÈBÔ<OÐ0PÐ0PÑQÔQˆˆå•t”y ¨LÑ!8Ñ9Ô9Ñ:Ô:Ð[¸q¸cÀYÇ]Â]Á_Ä_ÐWXÑEXÑ>YÐ[ˆØ,�9×#Ò#Ñ%Ô%Ô,¨gÐ6°y¸°yÔAˆÝŒ<˜ŸšÑ(Ô(¸	Ô8OÐPÑPÔPÐPó    Úmoduleú	nn.ModuleÚNonec                óž  — t          | dd¦  «        }|�t          |¦  «        dk    rdS |d         }t          |t          j        ¦  «        r|j        dk    rdS t          j        ¦   «         5  |j        dd…          	                    ¦   «          |j
        �!|j
        dd…          	                    ¦   «          ddd¦  «         dS # 1 swxY w Y   dS )zFReset keypoint precision-Cholesky output rows to unit Gaussian values.ÚlayersNr   éÿÿÿÿé   é   é   )ÚgetattrÚlenÚ
isinstancer   r/   Úout_featuresr1   r2   r)   Úzero_r(   )rI   rM   Úfinal_layers      r:   Ú$_reset_keypoint_gaussian_output_rowsrX   f   s  € å�V˜X tÑ,Ô,€FØ€~�˜V™œ¨Ò)Ð)Øˆà˜”*€KÝ�k¥2¤9Ñ-Ô-ð °Ô1IÈQÒ1NÐ1NØˆå	Œ‰Œð *ð *ØÔ˜1˜Q˜3Ô×%Ò%Ñ'Ô'Ð'ØÔÐ'ØÔ˜Q˜q˜SÔ!×'Ò'Ñ)Ô)Ð)ð*ð *ð *ñ *ô *ð *ð *ð *ð *ð *ð *ð *øøøð *ð *ð *ð *ð *ð *s   Á+A
CÃCÃ	Cc                  ó  ‡ — e Zd ZdZ	 	 	 	 	 	 	 	 dFdGˆ fd„ZdHd„ZedId„¦   «         ZedId„¦   «         ZdJd „Z	edKd#„¦   «         Z
dLd$„ZdMd%„ZdMd&„ZdNd*„ZdOd,„ZdPdQd2„ZdRd5„Zej        j        	 dPdSd=„¦   «         ZdTd?„ZdUdC„ZdVdE„Zˆ xZS )WÚLWDETRz@This is the Group DETR v3 module that performs object detection.Fr'   NÚbackboner   Útransformerr   Úsegmentation_headúSegmentationHead | Noner#   r$   Únum_queriesÚaux_lossÚboolÚ
group_detrÚ	two_stageÚlite_refpoint_refineÚbbox_reparamÚuse_grouppose_keypointsÚnum_keypoints_per_classúlist[int] | NoneÚ grouppose_keypoint_dim_downscaler%   rK   c           	     ó  •‡ — t          ¦   «                              ¦   «          |‰ _        |‰ _        |j        }t          j        ||¦  «        ‰ _        t          ||dd¦  «        ‰ _	        |‰ _
        d}t          j        ||z  |¦  «        ‰ _        t          j        ||z  |¦  «        ‰ _        t
          j                             ‰ j        j        j        d¦  «         |‰ _        |‰ _        |‰ _        |	‰ _        ‰ j        s‰ j	        ‰ j        j        _	        nd‰ j        j        _	        |
‰ _        |‰ _        |pg ‰ _        |‰ _        ‰ j        rCt7          ‰ j        ¦  «        |k    r+t9          dt7          ‰ j        ¦  «        › d|› d|› d�¦  «        ‚d	‰ _        ‰  ‰ j        r¦t          |‰ j        z  |‰ j        z  d
d¦  «        ‰ _        t?          t
          j        ‰ j        j         d         ¦  «        }t
          j                             |j        j        d¦  «         t
          j                             |j!        j        d¦  «         nd‰ _        ‰  ‰  "                    d‰  #                    ‰ j        ¦  «        ¦  «         d}tI          j%        d|z
  |z  ¦  «         }tM          j'        |¦  «        |z  ‰ j        j!        _        t?          t
          j        ‰ j	        j         d         ¦  «        }t
          j                             |j        j        d¦  «         t
          j                             |j!        j        d¦  «         |‰ _(        ‰ j(        r³t          j)        ˆ fd„tU          |¦  «        D ¦   «         ¦  «        ‰ j        _+        t          j)        ˆ fd„tU          |¦  «        D ¦   «         ¦  «        ‰ j        _,        ‰ j        r>‰ j        �7t          j)        ˆ fd„tU          |¦  «        D ¦   «         ¦  «        ‰ j        _-        d	‰ _.        dS )aÂ  Initializes the model.

        Parameters:
            backbone: torch module of the backbone to be used. See backbone.py
            transformer: torch module of the transformer architecture. See transformer.py
            num_classes: number of object classes
            num_queries: number of object queries, ie detection slot. This is the maximal number of objects
                         Conditional DETR can detect in a single image. For COCO, we recommend 100 queries.
            aux_loss: True if auxiliary decoding losses (loss at each decoder layer) are to be used.
            group_detr: Number of groups to speed detr training. Default is 1.
            lite_refpoint_refine: TODO
        rP   é   r   Nznum_keypoints_per_class has z) entries but the detection head only has z= classes. Class-logit boosts for keypoint classes with id >= zV would be silently truncated. Increase num_classes or shorten num_keypoints_per_class.Fé   rN   Ú_kp_active_maskg{®Gáz„?r'   c                óB   •— g | ]}t          j        ‰j        ¦  «        ‘ŒS © )ÚcopyÚdeepcopyÚ
bbox_embed©Ú.0Ú_Úselfs     €r:   ú
<listcomp>z#LWDETR.__init__.<locals>.<listcomp>Ú   s%   ø€ ÐKÐKÐK°A•”˜tœÑ/Ô/ÐKÐKÐKrH   c                óB   •— g | ]}t          j        ‰j        ¦  «        ‘ŒS ro   )rp   rq   Úclass_embedrs   s     €r:   rw   z#LWDETR.__init__.<locals>.<listcomp>Ý   s&   ø€ ÐLÐLÐL°Q•”˜tÔ/Ñ0Ô0ÐLÐLÐLrH   c                óB   •— g | ]}t          j        ‰j        ¦  «        ‘ŒS ro   )rp   rq   Úkeypoint_embedrs   s     €r:   rw   z#LWDETR.__init__.<locals>.<listcomp>á   s&   ø€ ÐSÐSÐS¸A•T”] 4Ô#6Ñ7Ô7ÐSÐSÐSrH   )/ÚsuperÚ__init__r_   r\   Úd_modelr   r/   ry   r   rr   r]   Ú	EmbeddingÚrefpoint_embedÚ
query_featÚinitÚ	constant_r)   Údatar[   r`   rb   rd   Údecoderre   rf   rg   ri   rS   Ú
ValueErrorÚ_kp_zero_pad_warnedr{   r   rM   r(   Úregister_bufferÚ_create_kp_active_maskr+   Úlogr1   Úonesrc   Ú
ModuleListÚrangeÚenc_out_bbox_embedÚenc_out_class_embedÚenc_out_keypoint_embedÚ_export)rv   r[   r\   r]   r#   r_   r`   rb   rc   rd   re   rf   rg   ri   Ú
hidden_dimÚ	query_dimÚkeypoint_out_layerÚ
prior_probÚ
bias_valueÚbbox_out_layerÚ	__class__s   `                   €r:   r}   zLWDETR.__init__y   sô  øø€ õ8 	‰Œ×ÒÑÔÐØ&ˆÔØ&ˆÔØ Ô(ˆ
Ýœ9 Z°Ñ=Ô=ˆÔÝ˜j¨*°a¸Ñ;Ô;ˆŒØ!2ˆÔØˆ	Ý œl¨;¸Ñ+CÀYÑOÔOˆÔÝœ, {°ZÑ'?ÀÑLÔLˆŒÝ
Œ×Ò˜$Ô-Ô4Ô9¸1Ñ=Ô=Ð=à ˆŒØ ˆŒØ$ˆŒð %9ˆÔ!ØÔ(ð 	7Ø26´/ˆDÔÔ$Ô/Ð/à26ˆDÔÔ$Ô/à(ˆÔð (?ˆÔ$Ø'>Ð'DÀ"ˆÔ$Ø0PˆÔ-ØÔ'ð 	­C°Ô0LÑ,MÔ,MÐP[Ò,[Ð,[Ýðh­s°4Ô3OÑ/PÔ/Pð hð hØ'ðhð hØfqðhð hð hñô ð ð $)ˆÔ ØÐ'ØÔ'ð 	'Ý"%Ø˜dÔCÑCØ˜dÔCÑCØØñ	#ô #ˆDÔõ "&¥b¤i°Ô1DÔ1KÈBÔ1OÑ!PÔ!PÐÝŒG×ÒÐ0Ô7Ô<¸aÑ@Ô@Ð@ÝŒG×ÒÐ0Ô5Ô:¸AÑ>Ô>Ð>Ð>à"&ˆDÔàÐ$Ø×ÒÐ.°×0KÒ0KÈDÔLhÑ0iÔ0iÑjÔjÐjð ˆ
Ý”h  J¡°*Ñ<Ñ=Ô=Ð=ˆ
Ý%*¤Z°Ñ%<Ô%<¸zÑ%IˆÔÔÔ"õ �bœi¨¬Ô)?ÀÔ)CÑDÔDˆÝ
Œ×Ò˜.Ô/Ô4°aÑ8Ô8Ð8Ý
Œ×Ò˜.Ô-Ô2°AÑ6Ô6Ð6ð #ˆŒØŒ>ð 
	Ý24´-ØKÐKÐKÐK½¸zÑ9JÔ9JÐKÑKÔKñ3ô 3ˆDÔÔ/õ 46´=ØLÐLÐLÐL½%À
Ñ:KÔ:KÐLÑLÔLñ4ô 4ˆDÔÔ0ð Ô+ð °Ô0CÐ0OÝ:<¼-ØSÐSÐSÐSÅÀzÑARÔARÐSÑSÔSñ;ô ;�Ô Ô7ð ˆŒˆˆrH   c                óÒ   ‡— t          | j        ‰¦  «        | _        | j        rD| j        j        €J ‚t          j        ˆfd„| j        j        D ¦   «         ¦  «        | j        _        dS dS )a¯  Resize the detection classification head to *num_classes* outputs.

        Replaces ``self.class_embed`` (and each ``enc_out_class_embed`` when the model uses two-stage detection) with a
        new :class:`torch.nn.Linear` whose ``out_features`` equals *num_classes*.  When *num_classes* is larger than the
        current head the existing weights are tiled; when smaller they are truncated.  Replacing the module (rather than
        mutating ``.data``) keeps ``nn.Linear.out_features`` consistent with the actual weight shape, which is required
        for correct ONNX export.

        Args:
            num_classes: Target number of output classes (including background).
        Nc                ó`   •— g | ]*}t          t          t          j        |¦  «        ‰¦  «        ‘Œ+S ro   )r;   r   r   r/   )rt   Úmr#   s     €r:   rw   z6LWDETR.reinitialize_detection_head.<locals>.<listcomp>÷   s/   ø€ ÐoÐoÐoÀQ•¥¥R¤Y°Ñ 2Ô 2°KÑ@Ô@ÐoÐoÐorH   )r;   ry   rc   r\   r�   r   rŒ   )rv   r#   s    `r:   Úreinitialize_detection_headz"LWDETR.reinitialize_detection_headæ   sz   ø€ õ *¨$Ô*:¸KÑHÔHˆÔàŒ>ð 	ØÔ#Ô7ÐCÐCÐCÝ35´=ØoÐoÐoÐoÈ$ÔJZÔJnÐoÑoÔoñ4ô 4ˆDÔÔ0Ð0Ð0ð	ð 	rH   ú	list[int]r   c                ó  — | s!t          j        ddt           j        ¬¦  «        S t          | ¦  «        }t          j        t	          | ¦  «        |t           j        ¬¦  «        }t          | ¦  «        D ]\  }}d||d|…f<   Œ|S )zECreate a compact class-by-keypoint active mask for a keypoint schema.r   ©ÚdtypeTN)r1   Úzerosra   ÚmaxrS   Ú	enumerate)rg   Úmax_kpÚ	kp_activeÚ	class_idxÚnum_keypointss        r:   r‰   zLWDETR._create_kp_active_maskú   s�   € ð 'ð 	7Ý”;˜q !­5¬:Ð6Ñ6Ô6Ð6åÐ,Ñ-Ô-ˆÝ”K¥Ð$;Ñ <Ô <¸fÍEÌJÐWÑWÔWˆ	Ý(1Ð2IÑ(JÔ(Jð 	8ð 	8Ñ$ˆI�}Ø37ˆI�i  - Ð/Ñ0Ð0ØÐrH   c                óÆ  — | s!t          j        ddt           j        ¬¦  «        S t          | ¦  «        }t          j        d|z   d|z   t           j        ¬¦  «        }t	          | ¦  «        D ]w\  }}|dk    rŒdt          | d|…         ¦  «        z   }||z   }t	          | ¦  «        D ]<\  }}|dk    s||k    rŒdt          | d|…         ¦  «        z   }	|	|z   }
d|||…|	|
…f<   Œ=Œx|S )zGCreate an attention mask that blocks cross-class keypoint interactions.r'   rŸ   r   NT)r1   r¡   ra   Úsumr£   )rg   Útotal_keypointsÚmaskÚclass_idx_iÚnum_keypoints_iÚstart_iÚend_iÚclass_idx_jÚnum_keypoints_jÚstart_jÚend_js              r:   Ú_create_keypoint_class_maskz"LWDETR._create_keypoint_class_mask  s'  € ð 'ð 	7Ý”;˜q !­5¬:Ð6Ñ6Ô6Ð6åÐ5Ñ6Ô6ˆÝŒ{˜1˜Ñ.°°OÑ0CÍ5Ì:ÐVÑVÔVˆÝ,5Ð6MÑ,NÔ,Nð 
	:ð 
	:Ñ(ˆK˜Ø !Ò#Ð#ØØ�#Ð5°l°{°lÔCÑDÔDÑDˆGØ˜oÑ-ˆEÝ09Ð:QÑ0RÔ0Rð :ð :Ñ,�˜_Ø" aÒ'Ð'¨;¸+Ò+EÐ+EØØ�cÐ"9¸,¸;¸,Ô"GÑHÔHÑH�Ø /Ñ1�Ø59��W˜U�] G¨E MÐ1Ñ2Ð2ð:ð ˆrH   c                óp   — d„ | j                              d¬¦  «                             ¦   «         D ¦   «         S )zJReturn the current keypoint schema inferred from the active-keypoint mask.c                ó,   — g | ]}t          |¦  «        ‘ŒS ro   ©r$   ©rt   r§   s     r:   rw   z6LWDETR.get_num_keypoints_per_class.<locals>.<listcomp>"  s    € ÐaÐaÐa }•�MÑ"Ô"ÐaÐaÐarH   r'   ©rA   )rm   r©   Útolist)rv   s    r:   Úget_num_keypoints_per_classz"LWDETR.get_num_keypoints_per_class   s7   € àaÐa¸Ô8L×8PÒ8PÐUVÐ8PÑ8WÔ8W×8^Ò8^Ñ8`Ô8`ÐaÑaÔaÐarH   Ú
state_dictúdict[str, Tensor]c                óÔ   — |                       d¦  «        }t          |t          ¦  «        r|j        dk    rdS d„ |                     d¬¦  «                             ¦   «         D ¦   «         S )z<Infer the keypoint schema stored in a checkpoint state dict.rm   é   Nc                ó,   — g | ]}t          |¦  «        ‘ŒS ro   r·   r¸   s     r:   rw   zFLWDETR.get_num_keypoints_per_class_from_checkpoint.<locals>.<listcomp>*  s    € ÐXÐXÐX }•�MÑ"Ô"ÐXÐXÐXrH   r'   r¹   )ÚgetrT   r   Úndimr©   rº   )r¼   Úactive_masks     r:   Ú+get_num_keypoints_per_class_from_checkpointz2LWDETR.get_num_keypoints_per_class_from_checkpoint$  sh   € ð !—n’nÐ%6Ñ7Ô7ˆÝ˜+¥vÑ.Ô.ð 	°+Ô2BÀaÒ2GÐ2GØ�4ØXÐX¸¿ºÈA¸Ñ8NÔ8N×8UÒ8UÑ8WÔ8WÐXÑXÔXÐXrH   c                óŠ  — | j         r|sdS t          |¦  «        }t          |¦  «        }|| _        |                      |¦  «                             | j        j        ¦  «        | _        t          | j	        d¦  «        r|| j	        _        t          | j	        dd¦  «        }|�Åt          |d¦  «        r||_        t          |dd¦  «        }t          |t          j        ¦  «        rt          ||¦  «        |_        t          |d¦  «        r|                     ¦   «          nIt          |d¦  «        r9|j        }|                      |¦  «                             |j        ¦  «        |_        dD ]Z}t          | j	        |d¦  «        }t          |dd¦  «        }	|�/t          |	t          j        ¦  «        rt          |	|¦  «        |_        Œ[dS )	zMResize schema-dependent GroupPose state to match ``num_keypoints_per_class``.Nrg   r…   Úkeypoint_pos_embedr´   Úkeypoint_class_mask)Úkeypoint_query_initializerÚkeypoint_query_initializer_encÚqueries)rf   Úlistr©   rg   r‰   Útorm   ÚdeviceÚhasattrr\   rR   rT   r   rB   rG   rÆ   r´   rÇ   rÊ   )
rv   rg   Úschemarª   r…   rÆ   Úcurrent_maskÚinitializer_nameÚinitializerrÊ   s
             r:   Úreinitialize_keypoint_headz!LWDETR.reinitialize_keypoint_head,  sÓ  € àÔ+ð 	Ð3Jð 	ØˆFåÐ-Ñ.Ô.ˆÝ˜f™+œ+ˆØ'-ˆÔ$Ø#×:Ò:¸6ÑBÔB×EÒEÀdÔFZÔFaÑbÔbˆÔå�4Ô#Ð%>Ñ?Ô?ð 	>Ø7=ˆDÔÔ4å˜$Ô*¨I°tÑ<Ô<ˆØÐÝ�wÐ 9Ñ:Ô:ð 9Ø28�Ô/Ý!(¨Ð2FÈÑ!MÔ!MÐÝÐ,­b¬lÑ;Ô;ð iÝ-CÐDVÐXgÑ-hÔ-h�Ô*Ý�wÐ =Ñ>Ô>ð oØ×3Ò3Ñ5Ô5Ð5Ð5Ý˜Ð"7Ñ8Ô8ð oØ&Ô:�Ø.2×.NÒ.NÈvÑ.VÔ.V×.YÒ.YÐZfÔZmÑ.nÔ.n�Ô+à `ð 	Wð 	WÐÝ! $Ô"2Ð4DÀdÑKÔKˆKÝ˜k¨9°dÑ;Ô;ˆGØÐ&­:°g½r¼|Ñ+LÔ+LÐ&Ý&<¸WÀoÑ&VÔ&V�Ô#øð		Wð 	WrH   c                óÚ   — | j         r| j        €dS t          | j        ¦  «         t          | j        dd¦  «        }t          |t          j        ¦  «        r|D ]}t          |¦  «         ŒdS dS )a¸  Reset keypoint Gaussian precision outputs to unit values.

        Keypoint channels 4, 5, and 6 encode the lower-triangular precision
        Cholesky parameters ``log_l11``, ``l21``, and ``log_l22``. Zeroing the
        final prediction rows gives ``L = identity`` at the start of finetuning
        while preserving learned keypoint location, visibility, findability, and
        class-logit channels loaded from the checkpoint.
        Nr�   )rf   r{   rX   rR   r\   rT   r   rŒ   )rv   Úenc_keypoint_embedr{   s      r:   Ú"reset_keypoint_gaussian_parametersz)LWDETR.reset_keypoint_gaussian_parametersL  s”   € ð Ô+ð 	¨tÔ/BÐ/JØˆFå,¨TÔ-@ÑAÔAÐAÝ$ TÔ%5Ð7OÐQUÑVÔVÐÝÐ(­"¬-Ñ8Ô8ð 	EØ"4ð Eð E�Ý4°^ÑDÔDÐDÐDð	Eð 	EðEð ErH   c                ó  — d| _         | j        | _        | j        | _        |                      ¦   «         D ]T\  }}t          |d¦  «        r?t          |j        ¦  «        r+t          |d¦  «        r|j         s|                     ¦   «          ŒUd S )NTÚexportr‘   )r‘   ÚforwardÚ_forward_originÚforward_exportÚnamed_modulesrÎ   ÚcallablerØ   )rv   Únamer›   s      r:   rØ   zLWDETR.export^  s“   € ØˆŒØ#œ|ˆÔØÔ*ˆŒØ×)Ò)Ñ+Ô+ð 	ð 	‰GˆD�!Ý�q˜(Ñ#Ô#ð ­°´Ñ(:Ô(:ð ½wÀqÈ)Ñ?TÔ?Tð Ð]^Ô]fð Ø—’‘
”
�
øð	ð 	rH   Úkeypoints_compactÚbatch_size_expectedÚnum_queries_expectedc                óš  — | j         r| j        s|S |                     ¦   «         dk    r|S |j        \  }}}}||k    s||k    rt	          d|› d|› d|› d|› d�	¦  «        ‚t          | j        ¦  «        }t          | j        ¦  «        |z  }	t          | j        ¦  «        }
||	k    r|S ||
k    r!t	          d|› d|
› d|	› d	| j        › d
�	¦  «        ‚t          j	        |||	||j
        |j        ¬¦  «        }d}t          | j        ¦  «        D ]C\  }}||z  }t          |¦  «        D ])}|dd…dd…|dd…f         |dd…dd…||z   dd…f<   |dz  }Œ*ŒD|S )zDConvert compact GroupPose keypoints to class-padded keypoint layout.rP   z8_format_keypoint_output received tensor with batch_size=z, num_queries=z but expected batch_size=zW. Shape mismatch silently bypassed in earlier versions; raise to surface upstream bugs.zE_format_keypoint_output received tensor with total_compact_keypoints=z) but schema expects either compact total=z or padded total=z for num_keypoints_per_class=ú.©r    rÍ   r   Nr'   )rf   rg   rA   r*   r†   r¢   rS   r©   r1   r¡   r    rÍ   r£   r�   )rv   rß   rà   rá   Ú
batch_sizer_   Útotal_compact_keypointsÚkeypoint_dimÚmax_num_keypointsÚtotal_padded_keypointsÚtotal_actual_keypointsÚpaddedÚcompact_idxr¦   Úkeypoint_countÚclass_offsetÚkeypoint_idxs                    r:   Ú_format_keypoint_outputzLWDETR._format_keypoint_outputf  sL  € ð Ô+ð 	%°4Ô3Oð 	%Ø$Ð$à× Ò Ñ"Ô" aÒ'Ð'Ø$Ð$àIZÔI`ÑFˆ
�KÐ!8¸,ØÐ,Ò,Ð,°Ð?SÒ0SÐ0SÝðhÈ:ð hð hÐepð hð hØ+>ðhð hØNbðhð hð hñô ð õ   Ô <Ñ=Ô=ÐÝ!$ TÔ%AÑ!BÔ!BÐEVÑ!VÐÝ!$ TÔ%AÑ!BÔ!BÐà"Ð&<Ò<Ð<à$Ð$Ø"Ð&<Ò<Ð<ÝðxÐXoð xð xØ;Qðxð xà#9ðxð xàX\ÔXtðxð xð xñô ð õ ”ØØØ"ØØ#Ô)Ø$Ô+ð
ñ 
ô 
ˆð ˆÝ)2°4Ô3OÑ)PÔ)Pð 	!ð 	!Ñ%ˆI�~Ø$Ð'8Ñ8ˆLÝ % nÑ 5Ô 5ð !ð !�Ø?PÐQRÐQRÐQRÐTUÐTUÐTUÐWbÐdeÐdeÐdeÐQeÔ?f��q�q�q˜!˜!˜!˜\¨LÑ8¸!¸!¸!Ð;Ñ<Ø˜qÑ ��ð!ð ˆrH   Úkeypoint_predictionsc           	     óB  — | j         s=t          j        g |j        dd…         ¢| j        j        ‘R |j        |j        ¬¦  «        S t          | j         ¦  «        }t          | j         ¦  «        } |d         j
        g |j        dd…         ¢|‘|‘R Ž }|| j                             |j        ¦  «        z  }|                     d¬¦  «        }| j        j        }|dz
  }|j        d         |k     r—|j        d         |k     rE| j        s>t                               d|j        d         ||j        d         |dz
  ¦  «         d	| _        t          j        | |j        g |j        dd…         ¢||j        d         z
  ‘R Ž gd¬¦  «        }n|j        d         |k    r|d
d|…f         }|S )zIAggregate keypoint class-logit contributions into detection-class logits.Néþÿÿÿrä   ).rQ   rN   r¹   r'   zòKeypoint class-logit boost has %d classes but detection head has %d foreground classes; zero-padding boost for classes %d..%d. Detection classes with no keypoint schema will receive zero boost. This warning is emitted once per model instance.T.)rg   r1   r¡   r*   ry   rU   r    rÍ   rS   r¢   Úviewrm   rÌ   r©   r‡   ÚloggerÚwarningÚcatr@   )rv   rñ   Únum_keypoint_classesrè   Úclass_contribÚclass_boostÚdetection_num_classesÚforeground_num_classess           r:   Ú _aggregate_keypoint_class_logitsz'LWDETR._aggregate_keypoint_class_logits™  s  € àÔ+ð 	Ý”;ØQÐ&Ô,¨S¨b¨SÔ1ÐQ°4Ô3CÔ3PÐQÐQØ*Ô0Ø+Ô2ðñ ô ð õ  # 4Ô#?Ñ@Ô@ÐÝ Ô <Ñ=Ô=ÐØ9Ð,¨VÔ4Ô9ð 
Ø!Ô'¨¨¨Ô,ð
Ø.Bð
ØDUð
ð 
ð 
ˆð &¨Ô(<×(?Ò(?ÀÔ@SÑ(TÔ(TÑTˆØ#×'Ò'¨BÐ'Ñ/Ô/ˆà $Ô 0Ô =ÐØ!6¸Ñ!:ÐØÔ˜RÔ Ð#8Ò8Ð8ØÔ  Ô$Ð'=Ò=Ð=ð
 Ô/ð 
4Ý—N’Nðdð $Ô)¨"Ô-Ø.Ø#Ô)¨"Ô-Ø.°Ñ2ñô ð ð 04�DÔ,Ýœ)àØ)�KÔ)Ðq¨;Ô+<¸S¸b¸SÔ+AÐqÐCXÐ[fÔ[lÐmoÔ[pÑCpÐqÐqÐqðð ðñ ô ˆKˆKð Ô˜rÔ"Ð%:Ò:Ð:ð & cÐ+AÐ,AÐ+AÐ&AÔBˆKØÐrH   Úsamplesr   Útargetsúlist[dict[str, Tensor]] | Noneúdict[str, Any]c           
     óô  — t          |t          t          f¦  «        rt          |¦  «        }|                      |¦  «        \  }}}g }g }|D ]G}|                     ¦   «         \  }	}
|                     |	¦  «         |                     |
¦  «         |
€J ‚ŒH| j        r| j        j	        }| j
        j	        }n2| j        j	        d| j        …         }| j
        j	        d| j        …         }| j        �| j        r| j        j        n| j        j        }d}|�3g }|D ].}|                     ¦   «         \  }}|                     |¦  «         Œ/|                      ||||||¬¦  «        }| j        r|dd…         \  }}}}}}n|dd…         \  }}}}d}d}|d         }i }|��Ï| j        r|                      |¦  «        }|ddd…f         |ddd…f         z  |ddd…f         z   }|ddd…f                              ¦   «         |ddd…f         z  }t)          j        ||gd¬¦  «        }n*|                      |¦  «        |z                        ¦   «         }|                      |¦  «        }d} | j        �rO| j        ��G|€t3          d	¦  «        ‚|                      |¦  «        }!t)          j        |!d
d
d
¬¦  «        }!|ddd…f                              d¦  «        }"|ddd…f                              d¦  «        }#|!ddd…f         |"z  |#z   }$|!ddd…f         }%t)          j        |$|%gd¬¦  «        }&g }'t;          |&j        d         ¦  «        D ]J}(|&|(         })|'                     |                      |)|)j        d         |)j        d         ¦  «        ¦  «         ŒKt)          j         |'d¬¦  «        } ||  !                    | ¦  «        z   }| j        �O ||d         j"        |tG          tH          tJ          tJ          f         |j"        j        dd…         ¦  «        ¦  «        }*|d         |d         dœ}| j        �|*d         |d<   | �| d         |d<   | j&        r$|  '                    ||| j        �|*nd| ¦  «        |d<   | j(        �r‹| j        j)        €J ‚|�|}+n‡| j        r| j*        nd},| +                    |,d¬¦  «        }-g }.t;          |,¦  «        D ]8}/ | j        j)        |/         |-|/         ¦  «        }0|.                     |0¦  «         Œ9t)          j        |.d¬¦  «        }+d}1| j        rG|�E|                      ||j        d         |j        d         ¦  «        }1|+|  !                    |1¦  «        z   }+| j        �X ||d         j"        |gtG          tH          tJ          tJ          f         |j"        j        dd…         ¦  «        d¬¦  «        d         }2|�(|+|dœ|d<   | j        �|2|d         d<   |1�|1|d         d<   n|+|dœ}| j        �|2|d<   |1�|1|d<   |S )a8  The forward expects a NestedTensor, which consists of:

           - samples.tensor: batched images, of shape [batch_size x 3 x H x W]
           - samples.mask: a binary mask of shape [batch_size x H x W], containing 1 on padded pixels

        It returns a dict with the following elements:
           - "pred_logits": the classification logits (including no-object) for all queries.
                            Shape= [batch_size x num_queries x num_classes]
           - "pred_boxes": The normalized boxes coordinates for all queries, represented as
                           (center_x, center_y, width, height). These values are normalized in [0, 1], relative to the
                           size of each individual image (disregarding possible padding). See PostProcess for
                           information on how to retrieve the unnormalized bounding box.
           - "aux_outputs": Optional, only returned when auxiliary losses are activated. It is a list of
                            dictionaries containing the two above keys for each decoder layer.
        N©Úcross_attn_srcsrO   rP   rN   .r¿   r¹   úKuse_grouppose_keypoints=True requires keypoint_hs from transformer outputs.ç        )ÚnanÚposinfÚneginfró   r   r'   )Úpred_logitsÚ
pred_boxesÚ
pred_masksÚpred_keypointsÚaux_outputsT©Úskip_blocksÚenc_outputs),rT   rË   r   r    r[   Ú	decomposeÚappendÚtrainingr€   r)   r�   r_   r]   Úsparse_forwardrÙ   r\   rf   re   rr   Úexpr1   ÚconcatÚsigmoidry   r{   r†   Ú
nan_to_numÚ	unsqueezer÷   r�   r*   rð   Ústackrý   Útensorsr   Útupler$   r`   Ú_set_aux_lossrc   r�   rb   Úchunk)3rv   rþ   rÿ   ÚfeaturesÚpossÚcross_attn_featuresÚsrcsÚmasksÚfeatÚsrcr«   Úrefpoint_embed_weightÚquery_feat_weightÚseg_head_fwdr  ÚfeatureÚ	cross_srcru   Útransformer_outputsÚhsÚref_unsigmoidÚhs_encÚref_encÚkeypoint_hsÚenc_kp_predictionsÚcls_tsÚoutÚoutputs_coord_deltaÚoutputs_coord_cxcyÚoutputs_coord_whÚoutputs_coordÚoutputs_classÚoutputs_keypointsÚoutputs_keypoints_deltaÚref_whÚref_xyÚkeypoints_xyÚkeypoints_otherÚoutputs_keypoints_compactÚlayer_outputs_keypointsÚ	layer_idxÚcompact_predsÚoutputs_masksÚcls_encrb   Úhs_enc_listÚcls_enc_listÚg_idxÚcls_enc_gidxÚkeypoints_encÚ	masks_encs3                                                      r:   rÙ   zLWDETR.forwardË  sF  € õ  �g¥¥f˜~Ñ.Ô.ð 	>Ý4°WÑ=Ô=ˆGØ.2¯mªm¸GÑ.DÔ.DÑ+ˆ�$Ð+àˆØˆØð 	$ð 	$ˆDØŸšÑ(Ô(‰IˆC�Ø�KŠK˜ÑÔÐØ�LŠL˜ÑÔÐØÐ#Ð#Ð#Ð#àŒ=ð 	KØ$(Ô$7Ô$>Ð!Ø $¤Ô 6ÐÐð %)Ô$7Ô$>Ð?QÀÔAQÐ?QÔ$RÐ!Ø $¤Ô 6Ð7I¸Ô9IÐ7IÔ JÐàÔ!Ð-ØDHÄMÐu˜4Ô1Ô@Ð@ÐW[ÔWmÔWuˆLàˆØÐ*Ø ˆOØ.ð 2ð 2�Ø&×0Ò0Ñ2Ô2‘�	˜1Ø×&Ò& yÑ1Ô1Ð1Ð1à"×.Ò.ØØØØ!ØØ+ð /ñ 
ô 
Ðð Ô'ð 	&ØReÐfhÐghÐfhÔRiÑOˆB�˜v w°Ð=OÐ=Oà1DÀRÀaÀRÔ1HÑ.ˆB�˜v wØˆKØ!%Ðð % RÔ(ˆà ˆØ‰>ØÔ ð PØ&*§o¢o°bÑ&9Ô&9Ð#Ø%8¸¸b¸q¸b¸Ô%AÀMÐRUÐWXÐWYÐWYÐRYÔDZÑ%ZÐ]jÐknÐprÐqrÐprÐkrÔ]sÑ%sÐ"Ø#6°s¸A¸B¸B°wÔ#?×#CÒ#CÑ#EÔ#EÈÐVYÐ[\Ð[]Ð[]ÐV]ÔH^Ñ#^Ð Ý %¤Ð.@ÐBRÐ-SÐY[Ð \Ñ \Ô \��à!%§¢°Ñ!4Ô!4°}Ñ!D× MÒ MÑ OÔ O�à ×,Ò,¨RÑ0Ô0ˆMØ $ÐàÔ+ñ i°Ô0CÑ0OØÐ&Ý$Ð%rÑsÔsÐsà*.×*=Ò*=¸kÑ*JÔ*JÐ'õ +0Ô*:Ð;RÐX[ÐdgÐpsÐ*tÑ*tÔ*tÐ'Ø& s¨A¨B¨B wÔ/×9Ò9¸"Ñ=Ô=�Ø& s¨B¨Q¨B wÔ/×9Ò9¸"Ñ=Ô=�Ø6°s¸B¸Q¸B°wÔ?À&ÑHÈ6ÑQ�Ø"9¸#¸q¸r¸r¸'Ô"B�Ý,1¬I°|À_Ð6UÐ[]Ð,^Ñ,^Ô,^Ð)à*,Ð'Ý!&Ð'@Ô'FÀqÔ'IÑ!JÔ!Jð ð �IØ$=¸iÔ$H�MØ+×2Ò2Ø×4Ò4Ø)Ø)Ô/°Ô2Ø)Ô/°Ô2ñô ñô ð ð õ %*¤KÐ0GÈQÐ$OÑ$OÔ$OÐ!Ø -°×0UÒ0UÐVgÑ0hÔ0hÑ h�àÔ%Ð1Ø , ¨X°a¬[Ô-@À"ÅdÍ5ÕQTÕVYÐQYÌ?Ð\cÔ\kÔ\qÐrtÐruÐruÔ\vÑFwÔFwÑ xÔ x�à"/°Ô"3À=ÐQSÔCTÐUÐUˆCØÔ%Ð1Ø$1°"Ô$5��LÑ!Ø Ð,Ø(9¸"Ô(=�Ð$Ñ%ØŒ}ð Ø%)×%7Ò%7Ø!Ø!Ø%)Ô%;Ð%G�M�MÈTØ%ñ	&ô &��MÑ"ð Œ>ñ .	:ØÔ#Ô7ÐCÐCÐCØÐ!ð !��à04´ÐD˜Tœ_˜_À1�
Ø$Ÿlšl¨:¸1˜lÑ=Ô=�Ø!�Ý" :Ñ.Ô.ð 6ð 6�EØ#N 4Ô#3Ô#GÈÔ#NÈ{Ð[`ÔOaÑ#bÔ#b�LØ ×'Ò'¨Ñ5Ô5Ð5Ð5åœ) L°aÐ8Ñ8Ô8�Ø ˆMØÔ+ð YÐ0BÐ0NØ $× <Ò <Ø&Ø&Ô,¨QÔ/Ø&Ô,¨QÔ/ñ!ô !�ð
 " D×$IÒ$IÈ-Ñ$XÔ$XÑX�àÔ%Ð1Ø(˜LØ˜Q”KÔ'àðõ ��s¥C˜xœ¨'¬/Ô*?ÀÀÀÔ*DÑEÔEØ $ðñ ô ð ô�	ð ˆ~Ø5<ÈGÐ%TÐ%T��MÑ"ØÔ)Ð5Ø7@�C˜Ô& |Ñ4Ø Ð,Ø;H�C˜Ô&Ð'7Ñ8øà&-¸WÐEÐE�ØÔ)Ð5Ø(1�C˜Ñ%Ø Ð,Ø,9�CÐ(Ñ)àˆ
rH   r  útuple[Tensor, ...]c                ó®  — |                       |¦  «        \  }}}}| j        j        d| j        …         }| j        j        d| j        …         }|                      |d||||¬¦  «        }| j        r|dd…         \  }	}
}}}}}n|dd…         \  }	}
}}d}d}d}d}|	��e| j        r|                      |	¦  «        }|ddd…f         |
ddd…f         z  |
ddd…f         z   }|ddd…f          	                    ¦   «         |
ddd…f         z  }t          j        ||gd¬¦  «        }n*|                      |	¦  «        |
z                        ¦   «         }|                      |	¦  «        }| j        �r_| j        ��W|€t          d	¦  «        ‚|                      |¦  «        }|ddd…f         }g d
g|
                     ¦   «         d
z
  z  ¢|j        d         ‘R }|
ddd…f                              |¦  «        }|
ddd…f                              |¦  «        }|                     d¦  «        |z  |z                        |j        ¦  «        }|ddd…f         }t          j        ||gd¬¦  «        }|                     ¦   «         dk    r|d         }|                      ||j        d         |j        d
         ¦  «        }||                      |¦  «        z   }| j        �1|                      |d         |	g|j        dd…         ¦  «        d         }nÄ| j        s
J d¦   «         ‚| j        j        €J ‚ | j        j        d         |¦  «        }|}| j        rG|�E|                      ||j        d         |j        d
         ¦  «        }||                      |¦  «        z   }| j        �3|                      |d         |g|j        dd…         d¬¦  «        d         }|�|||fS |�|||fS ||fS )a7  Run the export graph on a plain image batch.

        Args:
            tensors: Normalized images of shape ``(B, 3, H, W)``.

        Returns:
            ``(boxes, logits)`` for detection, ``(boxes, logits, masks)`` for segmentation, or
            ``(boxes, logits, keypoints)`` for keypoint models.
        Nr  rQ   rP   .r¿   rN   r¹   r  r'   ró   é   r   z,if not using decoder, two_stage must be TrueTr  )r[   r€   r)   r_   r�   r\   rf   re   rr   r  r1   r  r  ry   r{   r†   rA   r*   r.   Úflattenrô   r÷   rð   rý   r]   rc   r�   )rv   r  r#  ru   r!  r  r'  r(  r,  r-  r.  r/  r0  r1  r2  rD  r:  r5  r6  r7  r8  r9  r;  Údelta_xyrF   r<  r=  r>  r?  s                                r:   rÛ   zLWDETR.forward_exporty  sÃ  € ð *.¯ª°wÑ)?Ô)?Ñ&ˆˆa��à $Ô 3Ô :Ð;M¸TÔ=MÐ;MÔ NÐØ œOÔ2Ð3E°TÔ5EÐ3EÔFÐà"×.Ò.ØØØØ!ØØ+ð /ñ 
ô 
Ðð Ô'ð 	&ØUhÐikÐjkÐikÔUlÑRˆB�˜v w°Ð=OÐQRÐQRà1DÀRÀaÀRÔ1HÑ.ˆB�˜v wØˆKØ!%ÐàˆØ Ðà‰>ØÔ ð PØ&*§o¢o°bÑ&9Ô&9Ð#Ø%8¸¸b¸q¸b¸Ô%AÀMÐRUÐWXÐWYÐWYÐRYÔDZÑ%ZÐ]jÐknÐprÐqrÐprÐkrÔ]sÑ%sÐ"Ø#6°s¸A¸B¸B°wÔ#?×#CÒ#CÑ#EÔ#EÈÐVYÐ[\Ð[]Ð[]ÐV]ÔH^Ñ#^Ð Ý %¤Ð.@ÐBRÐ-SÐY[Ð \Ñ \Ô \��à!%§¢°Ñ!4Ô!4°}Ñ!D× MÒ MÑ OÔ O�Ø ×,Ò,¨RÑ0Ô0ˆMØÔ+ñ i°Ô0CÑ0OØÐ&Ý$Ð%rÑsÔsÐsØ*.×*=Ò*=¸kÑ*JÔ*JÐ'ð
 3°3¸¸¸°7Ô;�ØR˜a˜S M×$5Ò$5Ñ$7Ô$7¸!Ñ$;Ñ<ÐR¸x¼~ÈbÔ?QÐRÐR�Ø& s¨A¨B¨B wÔ/×6Ò6°wÑ?Ô?�Ø& s¨B¨Q¨B wÔ/×6Ò6°wÑ?Ô?�Ø (× 0Ò 0°Ñ 4Ô 4°vÑ =ÀÑ F×LÒLÈXÌ^Ñ\Ô\�Ø"9¸#¸q¸r¸r¸'Ô"B�Ý$)¤I¨|¸_Ð.MÐSUÐ$VÑ$VÔ$VÐ!Ø$×(Ò(Ñ*Ô*¨aÒ/Ð/Ø(9¸"Ô(=Ð%Ø$(×$@Ò$@Ø%Ø%Ô+¨AÔ.Ø%Ô+¨AÔ.ñ%ô %Ð!ð
 !.°×0UÒ0UÐVgÑ0hÔ0hÑ h�ØÔ%Ð1Ø $× 6Ò 6Ø˜”Gàðð ”M " # #Ô&ñ!ô !ð ô!�øð ”>ÐQÐQÐ#QÑQÔQ�>ØÔ#Ô7ÐCÐCÐCØC˜DÔ,Ô@ÀÔCÀFÑKÔKˆMØ#ˆMØÔ+ð iÐ0BÐ0NØ$(×$@Ò$@Ø&Ø&Ô,¨QÔ/Ø&Ô,¨QÔ/ñ%ô %Ð!ð
 !.°×0UÒ0UÐVgÑ0hÔ0hÑ h�ØÔ%Ð1Ø $× 6Ò 6Ø˜”Gàðð ”M " # #Ô&Ø $ð !7ñ !ô !ð ô!�ð Ð$Ø  -°Ð>Ð>ØÐ(Ø  -Ð1BÐBÐBà  -Ð/Ð/rH   r9  r8  rD  ú-list[Tensor] | list[dict[str, Tensor]] | Noner:  úTensor | Noneúlist[dict[str, Any]]c                ó.  ‡— ddgŠ|d d…         |d d…         g}|�2‰                      d¦  «         |                      |d d…         ¦  «         |�2‰                      d¦  «         |                      |d d…         ¦  «         ˆfd„t          |Ž D ¦   «         S )Nr
  r  rN   r  r  c                óD   •— g | ]}d „ t          ‰|¦  «        D ¦   «         ‘ŒS )c                ó   — i | ]\  }}||“Œ	S ro   ro   )rt   rÞ   Úvalues      r:   ú
<dictcomp>z3LWDETR._set_aux_loss.<locals>.<listcomp>.<dictcomp>ò  s   € ÐIÐIÐI¡  u��uÐIÐIÐIrH   )Úzip)rt   Úlayer_valuesÚnamess     €r:   rw   z(LWDETR._set_aux_loss.<locals>.<listcomp>ò  s2   ø€ ÐkÐkÐkÈlÐIÐIµ°E¸<Ñ0HÔ0HÐIÑIÔIÐkÐkÐkrH   )r  rY  )rv   r9  r8  rD  r:  Úvaluesr[  s         @r:   r  zLWDETR._set_aux_lossß  s´   ø€ ð  Ð-ˆØ*¨3¨B¨3Ô/°¸sÀ¸sÔ1CÐDˆØÐ$Ø�LŠL˜Ñ&Ô&Ð&Ø�MŠM˜-¨¨¨Ô,Ñ-Ô-Ð-ØÐ(Ø�LŠLÐ)Ñ*Ô*Ð*Ø�MŠMÐ+¨C¨R¨CÔ0Ñ1Ô1Ð1ØkÐkÐkÐkÕ^aÐciÐ^jÐkÑkÔkÐkrH   únn.ModuleList | Nonec                óê  — | j         d         j        }t          |d¦  «        rt          t          j        |j        ¦  «        S t          |d¦  «        r9t          |j        d¦  «        r$t          t          j        |j        j        ¦  «        S t          |d¦  «        rXt          |j        d¦  «        rCt          |j        j        d¦  «        r)t          t          j        |j        j        j        ¦  «        S dS )a_  Resolve the list of transformer blocks/layers from backbone[0].encoder.

        Supports multiple backbone architectures:
        - encoder.blocks (standard ViT)
        - encoder.trunk.blocks (aimv2)
        - encoder.encoder.encoder.layer (HuggingFace DinoV2)

        Returns:
            List of transformer layers, or None if not found.
        r   ÚblocksÚtrunkÚencoderÚlayerN)	r[   ra  rÎ   r   r   rŒ   r_  r`  rb  )rv   Úencs     r:   Ú_get_backbone_encoder_layersz#LWDETR._get_backbone_encoder_layersô  sÕ   € ð Œm˜AÔÔ&ˆÝ�3˜Ñ!Ô!ð 	3Ý�œ s¤zÑ2Ô2Ð2Ý�3˜Ñ Ô ð 	9¥W¨S¬Y¸Ñ%AÔ%Að 	9Ý�œ s¤yÔ'7Ñ8Ô8Ð8Ý�3˜	Ñ"Ô"ð 	B¥w¨s¬{¸IÑ'FÔ'Fð 	BÍ7ÐSVÔS^ÔSfÐhoÑKpÔKpð 	BÝ�œ s¤{Ô':Ô'@ÑAÔAÐAØˆtrH   Údrop_path_rateÚfloatÚvit_encoder_num_layersc                óf  — |                       ¦   «         }|€dS t          |t          |¦  «        ¦  «        }d„ t          j        d||¦  «        D ¦   «         }t          |¦  «        D ]K}t          ||         d¦  «        r3t          ||         j        d¦  «        r||         ||         j        _        ŒLdS )a¯  Update drop_path rates for backbone encoder layers with linear schedule.

        Applies a linear schedule where the first layer has drop_path_rate=0 and the last layer has
        drop_path_rate=drop_path_rate. Intermediate layers are interpolated linearly.

        Args:
            drop_path_rate: Maximum drop path rate (applied to last layer).
            vit_encoder_num_layers: Number of encoder layers to update.
        Nc                ó6   — g | ]}|                      ¦   «         ‘ŒS ro   )Úitem)rt   Úxs     r:   rw   z+LWDETR.update_drop_path.<locals>.<listcomp>  s    € ÐKÐKÐK �A—F’F‘H”HÐKÐKÐKrH   r   Ú	drop_pathÚ	drop_prob)	rd  ÚminrS   r1   Úlinspacer�   rÎ   rl  rm  )rv   re  rg  rM   ÚnÚdp_ratesÚis          r:   Úupdate_drop_pathzLWDETR.update_drop_path  s¸   € ð ×2Ò2Ñ4Ô4ˆØˆ>ØˆFÝÐ&­¨F©¬Ñ4Ô4ˆØKÐK¥e¤n°Q¸ÈÑ&JÔ&JÐKÑKÔKˆÝ�q‘”ð 	<ð 	<ˆAÝ�v˜a”y +Ñ.Ô.ð <µ7¸6À!¼9Ô;NÐP[Ñ3\Ô3\ð <Ø08¸´��q”	Ô#Ô-øð	<ð 	<rH   Ú	drop_ratec                ó€   — | j                              ¦   «         D ]#}t          |t          j        ¦  «        r||_        Œ$d S ©N)r\   ÚmodulesrT   r   ÚDropoutÚp)rv   rt  rI   s      r:   Úupdate_dropoutzLWDETR.update_dropout  sF   € ØÔ&×.Ò.Ñ0Ô0ð 	%ð 	%ˆFÝ˜&¥"¤*Ñ-Ô-ð %Ø$�”øð	%ð 	%rH   )Fr'   FFFFNr'   )r[   r   r\   r   r]   r^   r#   r$   r_   r$   r`   ra   rb   r$   rc   ra   rd   ra   re   ra   rf   ra   rg   rh   ri   r$   r%   rK   )r#   r$   r%   rK   )rg   r�   r%   r   )r%   r�   )r¼   r½   r%   rh   )rg   rh   r%   rK   )r%   rK   )rß   r   rà   r$   rá   r$   r%   r   )rñ   r   r%   r   rv  )rþ   r   rÿ   r   r%   r  )r  r   r%   rL  )
r9  r   r8  r   rD  rQ  r:  rR  r%   rS  )r%   r]  )re  rf  rg  r$   r%   rK   )rt  rf  r%   rK   )Ú__name__Ú
__module__Ú__qualname__Ú__doc__r}   rœ   Ústaticmethodr‰   r´   r»   rÄ   rÓ   rÖ   rØ   rð   rý   rÙ   rÛ   r1   ÚjitÚunusedr  rd  rs  rz  Ú__classcell__)r˜   s   @r:   rZ   rZ   v   s  ø€ € € € € ØJÐJð ØØØ%*Ø"Ø(-Ø48Ø01ðkð kð kð kð kð kð kðZð ð ð ð( ð	ð 	ð 	ñ „\ð	ð ðð ð ñ „\ðð2bð bð bð bð ðYð Yð Yñ „\ðYðWð Wð Wð Wð@Eð Eð Eð Eð$ð ð ð ð1ð 1ð 1ð 1ðf0ð 0ð 0ð 0ðdlð lð lð lð lð\d0ð d0ð d0ð d0ðL „YÔð ,0ðlð lð lð lñ Ôðlð(ð ð ð ð(<ð <ð <ð <ð&%ð %ð %ð %ð %ð %ð %ð %rH   rZ   Úargsú'BuilderArgs'úLWDETR | tuple[Any, None, None]c                óì  — | j         dz   }t          j        | j        ¦  «         t          d#i d| j        “d| j        “d| j        “d| j        “d| j        “d| j	        “d| j
        “d	| j        “d
| j        “d| j	        “d| j        “d| j        “d| j        “dt!          | d¦  «        r| j        nt!          | d¦  «        r| j        | j        fnd“d| j        “d| j        “d| j        “d| j        “d| j        d u “d| j        “d| j        “d| j        “d| j        “Ž}| j        r|d         j        d d fS | j        r|d d fS t=          | j        ¦  «        | _        tA          | ¦  «        }| j!        r!tE          | j	        | j#        | j$        ¬¦  «        nd }tK          ||||| j&        | j'        | j(        | j)        | j*        | j+        tY          | dd¦  «        tY          | d g ¦  «        tY          | d!d¦  «        ¬"¦  «        }|S )$Nr'   ra  rg  Úpretrained_encoderÚwindow_block_indexesrl  Úout_channelsÚout_feature_indexesÚprojector_scaleÚuse_cls_tokenr’   Úposition_embeddingÚfreeze_encoderÚ
layer_normÚtarget_shaper*   Ú
resolution)é€  r’  Úrms_normÚbackbone_loraÚforce_no_pretrainÚgradient_checkpointingÚload_dinov2_weightsÚ
patch_sizeÚnum_windowsÚpositional_encoding_sizeÚdual_projectorr   )Údownsample_ratiorf   Frg   ri   )
r#   r_   r`   rb   rc   rd   re   rf   rg   ri   ro   )-r#   r1   rÍ   r   ra  rg  r‡  rˆ  rl  r’   rŠ  r‹  rŒ  r�  rŽ  r�  rÎ   r*   r‘  r“  r”  r•  r–  Úpretrain_weightsr˜  r™  rš  r›  Úencoder_onlyÚbackbone_onlyrS   Únum_feature_levelsr   r]   r   Ú
dec_layersÚmask_downsample_ratiorZ   r_   r`   rb   rc   rd   re   rR   )rƒ  r#   r[   r\   r]   Úmodels         r:   Úbuild_modelr¤  !  sÞ  € ð Ô" QÑ&€KÝ	„L�”ÑÔÐåð ð ð Ø”�ðà#Ô:Ð:ðð  Ô2Ð2ðð "Ô6Ð6ð	ð
 ”.�.ðð ”_�_ðð !Ô4Ð4ðð Ô,Ð,ðð Ô(Ð(ðð ”?�?ðð  Ô2Ð2ðð Ô*Ð*ðð ”?�?ðõ  �t˜WÑ%Ô%ðeˆDŒJˆJå8?ÀÀlÑ8SÔ8SÐc�4”? D¤OÐ4Ð4ÐYcøð#ð& ”�ð'ð( Ô(Ð(ð)ð* Ô0Ð0ð+ð,  $Ô:Ð:ð-ð. !Ô1°TÐ9Ð9ð/ð0 ”?�?ð1ð2 Ô$Ð$ð3ð4 "&Ô!>Ð!>ð5ð6 Ô*Ð*ð7€Hð: Ôð /Ø˜Œ{Ô" D¨$Ð.Ð.ØÔð $Ø˜˜tÐ#Ð#å! $Ô"6Ñ7Ô7€DÔÝ# DÑ)Ô)€Kð Ô!ð	ÕØŒOØŒOØ!Ô7ð	
ñ 	
ô 	
ð 	
ð ð õ ØØØØØÔ$Ø”Ø”?Ø”.Ø!Ô6ØÔ&å '¨Ð.GÈÑ OÔ OÝ '¨Ð.GÈÑ LÔ LÝ)0°Ð7YÐ[\Ñ)]Ô)]ðñ ô €Eð  €LrH   ú tuple[SetCriterion, PostProcess]c                óZ  ‡
— t          j        | j        ¦  «        }t          | ¦  «        }| j        | j        dœ}| j        |d<   | j        r| j        |d<   | j        |d<   t          | dd¦  «        }|rPt          | dd¦  «        |d	<   t          | d
d¦  «        |d<   t          | dd¦  «        |d<   t          | dd¦  «        |d<   | j
        rœi }t          | j        dz
  ¦  «        D ]5Š
|                     ˆ
fd„|                     ¦   «         D ¦   «         ¦  «         Œ6| j        r1|                     d„ |                     ¦   «         D ¦   «         ¦  «         |                     |¦  «         g d¢}| j        r|                     d¦  «         |r|                     d¦  «         t          | dd¦  «        }| j        rQt#          | j        dz   ||| j        || j        || j        | j        | j        | j        t          | dg ¦  «        ¬¦  «        }nJt#          | j        dz   ||| j        || j        || j        | j        | j        t          | dg ¦  «        ¬¦  «        }|                     |¦  «         t5          | j        t          | dg ¦  «        t          | dd¦  «        t          | dd¦  «         ¬¦  «        }	||	fS )N)Úloss_ceÚ	loss_bboxÚ	loss_giouÚloss_mask_ceÚloss_mask_dicerf   FÚkeypoint_l1_loss_coefr  Úloss_keypoints_l1Úkeypoint_findable_loss_coefÚloss_keypoints_findableÚkeypoint_visible_loss_coefÚloss_keypoints_visibleÚkeypoint_nll_loss_coefÚloss_keypoints_nllr'   c                ó(   •— i | ]\  }}|d ‰› �z   |“ŒS )ru   ro   )rt   ÚkÚvrr  s      €r:   rX  z6build_criterion_and_postprocessors.<locals>.<dictcomp>‚  s)   ø€ Ð#SÐ#SÐ#S±t°q¸! A¨¨A¨¨¡K°Ð#SÐ#SÐ#SrH   c                ó    — i | ]\  }}|d z   |“ŒS )Ú_encro   )rt   rµ  r¶  s      r:   rX  z6build_criterion_and_postprocessors.<locals>.<dictcomp>„  s"   € Ð#RÐ#RÐ#R±d°a¸ A¨¡J°Ð#RÐ#RÐ#RrH   )ÚlabelsÚboxesÚcardinalityr$  Ú	keypointsÚsum_group_lossesrg   )ÚmatcherÚweight_dictÚfocal_alphaÚlossesrb   r½  Úuse_varifocal_lossÚuse_position_supervised_lossÚia_bce_lossÚmask_point_sample_ratiorg   )
r¾  r¿  rÀ  rÁ  rb   r½  rÂ  rÃ  rÄ  rg   Úpostprocess_trace_alphagš™™™™™É?Úeval_masks_head_resolution)Ú
num_selectrg   Útrace_alphaÚupsample_masks_to_image_size)r1   rÍ   r   Úcls_loss_coefÚbbox_loss_coefÚgiou_loss_coefr]   Úmask_ce_loss_coefÚmask_dice_loss_coefrR   r`   r�   r¡  ÚupdateÚitemsrc   r  r   r#   rÀ  rb   rÂ  rÃ  rÄ  rÅ  rÌ   r   rÈ  )rƒ  rÍ   r¾  r¿  Úhas_keypointsÚaux_weight_dictrÁ  r½  Ú	criterionÚpostprocessrr  s             @r:   Ú"build_criterion_and_postprocessorsrÖ  o  s$  ø€ ÝŒ\˜$œ+Ñ&Ô&€FÝ˜DÑ!Ô!€GØ"Ô0¸tÔ?RÐSÐS€KØ#Ô2€K�ÑØÔð AØ&*Ô&<ˆ�NÑ#Ø(,Ô(@ˆÐ$Ñ%å˜DÐ";¸UÑCÔC€MØð YÝ+2°4Ð9PÐRUÑ+VÔ+VˆÐ'Ñ(Ý18¸Ð?\Ð^aÑ1bÔ1bˆÐ-Ñ.Ý07¸Ð>ZÐ\_Ñ0`Ô0`ˆÐ,Ñ-Ý,3°DÐ:RÐTWÑ,XÔ,XˆÐ(Ñ)à„}ð ,ØˆÝ�t”¨Ñ*Ñ+Ô+ð 	Uð 	UˆAØ×"Ò"Ð#SÐ#SÐ#SÐ#S¸{×?PÒ?PÑ?RÔ?RÐ#SÑ#SÔ#SÑTÔTÐTÐTØŒ>ð 	TØ×"Ò"Ð#RÐ#R¸k×>OÒ>OÑ>QÔ>QÐ#RÑ#RÔ#RÑSÔSÐSØ×Ò˜?Ñ+Ô+Ð+à/Ð/Ð/€FØÔð Ø�Š�gÑÔÐØð #Ø�Š�kÑ"Ô"Ð"å˜tÐ%7¸Ñ?Ô?ÐØÔð 
Ý ØÔ˜qÑ ØØ#ØÔ(ØØ”Ø-Ø#Ô6Ø)-Ô)JØÔ(Ø$(Ô$@Ý$+¨DÐ2KÈRÑ$PÔ$Pð
ñ 
ô 
ˆ	ˆ	õ !ØÔ˜qÑ ØØ#ØÔ(ØØ”Ø-Ø#Ô6Ø)-Ô)JØÔ(Ý$+¨DÐ2KÈRÑ$PÔ$Pð
ñ 
ô 
ˆ	ð ‡L‚L�ÑÔÐÝØ”?Ý '¨Ð.GÈÑ LÔ Lå˜DÐ";¸SÑAÔAå)0°Ð7SÐUZÑ)[Ô)[Ð%[ðñ ô €Kð �kÐ!Ð!rH   Úmodel_configú'ModelConfig'Útrain_configú'TrainConfig' | NoneÚdefaultsr   c                óì   — ddl m} |j        s|j        rt	          d¦  «        ‚|€ddlm}  |dd¬¦  «        } || ||¦  «        }t          |¦  «        }t          |t          ¦  «        s
J d¦   «         ‚|S )	uQ  Build an LWDETR model directly from a ModelConfig.

    A config-native alternative to ``build_model(build_namespace(mc, tc))``. Constructs the namespace internally from
    ``model_config``, an optional ``train_config``, and ``defaults``, then delegates to :func:`build_model`.

    Note:
        The internal ``SimpleNamespace`` bridge is transitional â€” it will be eliminated once all builder functions
        accept config objects directly. Callers should not rely on the namespace shape or pass it externally.

    Args:
        model_config: Architecture configuration.
        train_config: Training hyperparameter configuration. If ``None``,
            a minimal dummy ``TrainConfig(dataset_dir=".", output_dir=".")`` is constructed, matching the previous
            default behavior.
        defaults: Hardcoded architectural constants. Defaults to ``MODEL_DEFAULTS``.

    Returns:
        Fully initialised LWDETR model.

    Raises:
        ValueError: If ``defaults`` request ``encoder_only`` or ``backbone_only``,
            which would make the return type differ from ``LWDETR``.
    r   ©Ú_namespace_from_configsz`build_model_from_config() requires defaults.encoder_only=False and defaults.backbone_only=False.N)r
   rã   )Údataset_dirÚ
output_dirz]build_model() returned a non-LWDETR result even though encoder_only/backbone_only were False.)
Úrfdetr._namespacerÞ  rž  rŸ  r†   Úrfdetr.configr
   r¤  rT   rZ   )r×  rÙ  rÛ  rÞ  r
   Únsr£  s          r:   Úbuild_model_from_configrä  ¸  s»   € ð8 :Ð9Ð9Ð9Ð9Ð9àÔð 
 Ô 6ð 
ÝØnñ
ô 
ð 	
ð ÐØ-Ð-Ð-Ð-Ð-Ð-à"�{¨s¸sÐCÑCÔCˆà	 Ð	  ¨|¸XÑ	FÔ	F€BÝ˜‰OŒO€EÝ�e�VÑ$Ô$ð ð Øgñô Ð$ð €LrH   ú'TrainConfig'c                óF   — ddl m}  || ||¦  «        }t          |¦  «        S )a¼  Build criterion and postprocessor directly from config objects.

    A config-native alternative to ``build_criterion_and_postprocessors(build_namespace(mc, tc))``.

    Args:
        model_config: Architecture configuration.
        train_config: Training hyperparameter configuration.
        defaults: Hardcoded architectural constants. Defaults to ``MODEL_DEFAULTS``.

    Returns:
        A 2-tuple of ``(SetCriterion, PostProcess)``.
    r   rÝ  )rá  rÞ  rÖ  )r×  rÙ  rÛ  rÞ  rã  s        r:   Úbuild_criterion_from_configrç  è  s8   € ð" :Ð9Ð9Ð9Ð9Ð9à	 Ð	  ¨|¸XÑ	FÔ	F€BÝ-¨bÑ1Ô1Ð1rH   )r!   r"   r#   r$   r%   r"   )r<   r=   r>   r$   r%   r=   )rI   rJ   r%   rK   )rƒ  r„  r%   r…  )rƒ  r„  r%   r¥  )r×  rØ  rÙ  rÚ  rÛ  r   r%   rZ   )r×  rØ  rÙ  rå  rÛ  r   r%   r¥  ):r~  Ú
__future__r   rp   r+   Útypingr   r   r   r1   r   r   râ  r	   r
   Úrfdetr.models._defaultsr   r   Úrfdetr.models._typesr   Úrfdetr.models.backboner   r   Úrfdetr.models.criterionr   r   r   r   r   r   r   r   Ú rfdetr.models.heads.segmentationr   Úrfdetr.models.matcherr   Úrfdetr.models.mathr   Úrfdetr.models.postprocessr   Úrfdetr.models.transformerr   r   Úrfdetr.utilities.loggerr   Úrfdetr.utilities.tensorsr   r    rõ   r;   rG   rX   ÚModulerZ   r¤  rÖ  rä  rç  ro   rH   r:   ú<module>rö     s¾  ðð$ +Ð *à "Ð "Ð "Ð "Ð "Ð "à €€€Ø €€€Ø +Ð +Ð +Ð +Ð +Ð +Ð +Ð +Ð +Ð +à €€€Ø Ð Ð Ð Ð Ð Ð Ð àð 7Ø6Ð6Ð6Ð6Ð6Ð6Ð6Ð6à AÐ AÐ AÐ AÐ AÐ AÐ AÐ AØ ,Ð ,Ð ,Ð ,Ð ,Ð ,Ø 9Ð 9Ð 9Ð 9Ð 9Ð 9Ð 9Ð 9ð	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 =Ø /Ð /Ð /Ð /Ð /Ð /Ø "Ð "Ð "Ð "Ð "Ð "Ø 1Ð 1Ð 1Ð 1Ð 1Ð 1Ø DÐ DÐ DÐ DÐ DÐ DÐ DÐ DØ .Ð .Ð .Ð .Ð .Ð .Ø QÐ QÐ QÐ QÐ QÐ QÐ QÐ Qà	ˆ‰Œ€ðð ð ð ð<Qð Qð Qð Qð*ð *ð *ð *ð h
%ð h
%ð h
%ð h
%ð h
%ˆRŒYñ h
%ô h
%ð h
%ðVKð Kð Kð Kð\F"ð F"ð F"ð F"ðV *.Ø,ð-ð -ð -ð -ð -ðf -ð2ð 2ð 2ð 2ð 2ð 2ð 2rH   