§
    ­¸j$t  ã                  óÖ   — U d Z ddlmZ ddlmZ ddlmZ ddlZddlm	c m
Z ddlmZm	Z	 ddlmZ  e¦   «         ZdZd	ed
<   d d„Z G d„ de	j        ¦  «        Zd!d„Zd"d„Zd!d„Zg d¢ZdS )#uã  Keypoint helper primitives for GroupPose-style decoding.

Keypoint predictions are dense ``[N, K, KEYPOINT_PRED_DIM]`` tensors where the trailing channel layout is fixed:

================  ===========  ==========================================================
Slot index        Name         Meaning
================  ===========  ==========================================================
``0``             ``x``        Normalized x coordinate (image-relative).
``1``             ``y``        Normalized y coordinate (image-relative).
``2``             ``findable`` Logit for "annotator could find this keypoint" (``v > 0``).
``3``             ``visible``  Logit for "fully visible" (``v == 2``).
``4`` â€“ ``6``     ``L_*``      Lower-triangular Cholesky parameters ``Lxx``, ``Lxy``, ``Lyy``.
``7``             ``class``    Per-keypoint class-logit contribution aggregated into detection-class logits.
================  ===========  ==========================================================
é    )Úannotations)ÚSequence)ÚcastN)ÚTensorÚnn)Ú
get_loggeré   ÚintÚKEYPOINT_PRED_DIMÚfeaturesr   ÚscaleÚshiftÚreturnc                ó   — |dz   | z  |z   S )u$  Apply AdaLN modulation to a feature tensor.

    Args:
        features: Input feature map to modulate.
        scale: Per-feature scale terms.
        shift: Per-feature shift terms.

    Returns:
        Modulated features with the same shape as ``features``.

    Example:
        Apply modulation to a batch of 2 query tokens, each with 4 channels:

        .. code-block:: python

            features = torch.zeros(2, 4)          # (batch, dim)
            scale    = torch.ones(2, 4) * 0.1     # small positive scale
            shift    = torch.ones(2, 4) * (-0.5)  # constant shift

            out = modulate(features, scale, shift)
            # scale=0.1 â†’ effective multiplier is 1.1; shift=-0.5 applied additively
            # out[0] â‰ˆ tensor([-0.5000, -0.5000, -0.5000, -0.5000])
    g      ð?© )r   r   r   s      úm/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/rfdetr/models/heads/keypoints.pyÚmodulater   '   s   € ð2 �C‰K˜8Ñ# eÑ+Ð+ó    c                  ó.   ‡ — e Zd ZdZddˆ fd
„Zdd„Zˆ xZS )ÚConditionalQueryInitializerzTInitialize keypoint query tokens with adaptive layer-normalization style modulation.NÚdimr
   Únum_queriesÚout_dimú
int | Noner   ÚNonec           	     ó†  •— t          ¦   «                              ¦   «          |p|}t          j        t	          j        ||¦  «        ¦  «        | _        t          j        |d¬¦  «        | _        t          j	        t          j
        ||¦  «        t          j        ¦   «         t          j
        ||dz  ¦  «        ¦  «        | _        t          d| j        d         ¦  «        }t          j                             |j        d¦  «         t          j                             |j        d¦  «         t          j
        ||¦  «        | _        dS )zäCreate the initializer.

        Args:
            dim: Query conditioning dimensionality.
            num_queries: Number of query tokens to instantiate.
            out_dim: Output embedding size. Defaults to ``dim``.
        F)Úelementwise_affineé   z	nn.Linearéÿÿÿÿr   N)ÚsuperÚ__init__r   Ú	ParameterÚtorchÚrandnÚqueriesÚ	LayerNormÚ
query_normÚ
SequentialÚLinearÚGELUÚadaLN_modulationr   ÚinitÚ	constant_ÚweightÚbiasÚout_proj)Úselfr   r   r   Úada_ln_projectionÚ	__class__s        €r   r!   z$ConditionalQueryInitializer.__init__F   sÿ   ø€ õ 	‰Œ×ÒÑÔÐØ�.˜Sˆå”|¥E¤K°¸WÑ$EÔ$EÑFÔFˆŒÝœ, wÀ5ÐIÑIÔIˆŒÝ "¤ÝŒI�c˜3ÑÔÝŒG‰IŒIÝŒI�c˜7 Q™;Ñ'Ô'ñ!
ô !
ˆÔõ
 ! ¨dÔ.CÀBÔ.GÑHÔHÐÝ
Œ×ÒÐ+Ô2°AÑ6Ô6Ð6Ý
Œ×ÒÐ+Ô0°!Ñ4Ô4Ð4åœ	 '¨7Ñ3Ô3ˆŒˆˆr   Úquery_featuresr   c                ó:  — |                       | j        ¦  «        }|                      |                     d¦  «        ¦  «        }|                     dd¬¦  «        \  }}}|                      t          |||¦  «        ¦  «        |z  | j        z   }t          d|¦  «        S )ax  Return modulated query embeddings.

        Args:
            query_features: Tensor of shape ``(B, dim)`` containing conditioning features.

        Returns:
            Tensor of shape ``(B, num_queries, out_dim)`` with initialized keypoint queries.

        Example:
            Initialize 17 keypoint queries conditioned on a batch of 2 detection embeddings:

            .. code-block:: python

                initializer = ConditionalQueryInitializer(dim=256, num_queries=17)
                query_features = torch.randn(2, 256)   # (B=2, dim=256)

                keypoint_queries = initializer(query_features)
                # keypoint_queries.shape == (2, 17, 256)
                # Each of the 17 query slots is independently modulated by the
                # per-detection conditioning vector before being passed to the
                # GroupPose keypoint decoder.
        éþÿÿÿr   r   ©r   r   )r'   r%   r+   Ú	unsqueezeÚchunkr0   r   r   )r1   r4   Únormed_query_featuresÚ
modulationr   r   ÚgateÚmodulated_query_featuress           r   Úforwardz#ConditionalQueryInitializer.forward^   s™   € ð0 !%§¢°´Ñ =Ô =ÐØ!×2Ò2°>×3KÒ3KÈBÑ3OÔ3OÑPÔPˆ
Ø'×-Ò-¨a°RÐ-Ñ8Ô8Ñˆˆu�dØ#'§=¢=µÐ:OÐQVÐX]Ñ1^Ô1^Ñ#_Ô#_ÐbfÑ#fÐimÔiuÑ#uÐ Ý�HÐ6Ñ7Ô7Ð7r   )N)r   r
   r   r
   r   r   r   r   )r4   r   r   r   )Ú__name__Ú
__module__Ú__qualname__Ú__doc__r!   r>   Ú__classcell__)r3   s   @r   r   r   C   s\   ø€ € € € € Ø^Ð^ð4ð 4ð 4ð 4ð 4ð 4ð 4ð08ð 8ð 8ð 8ð 8ð 8ð 8ð 8r   r   Úall_pred_keypointsÚtarget_keypointsÚtarget_classesÚtarget_areasÚnum_keypoints_per_classúSequence[int]ú%tuple[Tensor, Tensor, Tensor, Tensor]c           
     óž  — | j         \  }}}|dk    s
J d¦   «         ‚t          |¦  «        }|dk    rt          d¦  «        ‚|dk    r |                     ¦   «         |k    rˆt                               dt          |                     ¦   «         ¦  «        |¦  «         |                      d¦  «        dd…                              ¦   «         }	|  	                    |¦  «        |	z   }
|
|
|
|
fS ||z  dk    rt          d|› d	|› d
�¦  «        ‚||z  }|  
                    ||||¦  «        }|t          j        ||j        ¬¦  «        |f         }t          j        ||t          j        |j        ¬¦  «        }t!          |¦  «        D ]\  }}d||d|…f<   Œ||         }|                     t          j        ¦  «        }t          j        |j        ¦  «        j        }t          j        |¦  «        ||k    z  }t          j        |dd…dd…dd…f         ¦  «                             d¬¦  «        t          j        |dd…dd…dd…f         ¦  «                             d¬¦  «        z  }t          j        |dd…dd…df         ¦  «        |dd…dd…df         dk    z  }||z  |z  |                     d¦  «        z  }|                     d¦  «                             |j        ¬¦  «        }|                     d¬¦  «        }t          j        |dd…dd…df         ¦  «        }||z  }|                     d¦  «                             d¬¦  «                             |j        ¬¦  «        }t          j        |dd…dd…df         ¦  «        }||z  }|                     d¦  «                             d¬¦  «                             |j        ¬¦  «        }t          j        ||t          j        |¦  «        ¦  «        } |                      |¦  «                             ¦   «         }!t          j        |dd…dd…dd…f         ¦  «        }"t          j        |                     d¦  «        |dd…dd…dd…f         |"¦  «        }#t          j        |                     d¦  «        |dd…dd…dd…f         |"¦  «        }$t?          j         |#|$d¬¦  «                             d¦  «        }%|% !                    | d¦  «        |!                     d¦  «        z  }&|&                     d¦  «        |z  }'t          j        ||dd…dd…df         t          j        |dd…dd…df         ¦  «        ¦  «        }(t?          j"        |(|dd…dd…df         dk                         |j        ¦  «        d¬¦  «        })|) !                    | d¦  «                             d¦  «        |z  }*t          j        ||dd…dd…df         t          j        |dd…dd…df         ¦  «        ¦  «        }+t?          j"        |+|dd…dd…df         dk                         |j        ¦  «        d¬¦  «        },|, !                    | d¦  «                             d¦  «        |z  }-|dd…dd…df                              t          j        ¦  «        }.|dd…dd…df                              t          j        ¦  «        }/|dd…dd…df                              t          j        ¦  «        }0t          j        |.¦  «        t          j        |/¦  «        z  t          j        |0¦  «        z  }1|1                     ¦   «         s@t           #                    d|1                      ¦   «          $                    ¦   «         ¦  «         ||1z  }2t          j        |2                     d¦  «        |dd…dd…dd…f         |"¦  «        }3t          j        |2                     d¦  «        |dd…dd…dd…f         |"¦  «        }4|3|4z
                       t          j        ¦  «        }5|5dd…dd…df         }6|5dd…dd…df         }7t          j        |2|.t          j        |.¦  «        ¦  «        }8t          j        |2|/t          j        |/¦  «        ¦  «        }9t          j        |2|0t          j        |0¦  «        ¦  «        }:|8 %                    ¦   «         };|: %                    ¦   «         }<|;|6z  |9|7z  z   }=|<|7z  }>|=|=z  |>|>z  z   }?|2t          j        |=¦  «        z  t          j        |>¦  «        z  t          j        |?¦  «        z  }@t          j        |.¦  «        }At          j        |@|8|A¦  «        }8t          j        |@|9|A¦  «        }9t          j        |@|:|A¦  «        }:t          j        |@|6t          j        |6¦  «        ¦  «        }6t          j        |@|7t          j        |7¦  «        ¦  «        }7|8 %                    ¦   «         };|: %                    ¦   «         }<|;|6z  |9|7z  z   }=|<|7z  }>|=|=z  |>|>z  z   }?|@}2|2                     d¦  «                             |j        ¬¦  «        }B|B                     d¬¦  «        }Cd|?|                      |¦  «                             d¦  «        z  z  |8|:z   z
  }Dt          j&        |Ddt          j        |Dj        ¦  «        j        dz  t          j        |Dj        ¦  «        j'        ¬¦  «        }D|D !                    |2 d¦  «        }E|E                     d¦  «        |Cz  }F|Bdk    }Gt          j        |Gt          j        |F¦  «        |F¦  «        }F|'|*|-|FfS )u  Compute the keypoint loss vector per matched target.

    The tensor layout follows GroupPose-style keypoints where each target class
    defines how many valid keypoints it owns. The returned Gaussian NLL follows
    the r-flow formulation directly:
    ``0.5 * maha2 / area - (log_l11 + log_l22)``. It intentionally omits the
    normal Gaussian constant and does not clamp the precision Cholesky
    parameters, so valid losses may be negative.

    A non-finite value in any single prediction channel for a given keypoint (e.g. from an AMP
    overflow upstream) is excluded from both the forward value and the backward gradient of every
    loss term it would otherwise contribute to, without poisoning the other, still-finite keypoints
    in the same target. This guarantee holds only *inside* this function: model boundaries that
    combine the prediction with shared tensors sanitize the full prediction before those outer ops,
    while a non-finite value that later re-enters an unguarded outer op can still poison that op's
    local gradient, since this function cannot see or guard operations outside its call frame.
    ``compute_keypoint_matching_cost``'s ``_cdist_bce_with_logits`` (the sibling used by the matcher)
    does not yet have equivalent protection â€” a non-finite prediction there still poisons the whole
    matching-cost row for that query, tracked separately since it needs a different fix shape.

    Args:
        all_pred_keypoints: Predicted keypoints with shape ``(N, K_total, >=7)``.
        target_keypoints: Ground truth keypoints with shape ``(N, K_max, 3)``.
        target_classes: Class ids per target with shape ``(N,)``.
        target_areas: Target box areas with shape ``(N,)``.
        num_keypoints_per_class: Number of keypoints per class.

    Returns:
        Tuple of location, findable BCE, visible BCE, and raw Gaussian NLL losses.
        Each tensor has shape ``(n_targets,)``.

    Example:
        Compute losses for 2 matched targets, each with 17 keypoints (COCO layout):

        .. code-block:: python

            n_targets, K, pred_dim = 2, 17, 7
            all_pred_keypoints = torch.randn(n_targets, K, pred_dim)

            # Ground truth: 17 keypoints per target, each (x, y, visibility)
            # visibility: 0=not labeled, 1=labeled but occluded, 2=fully visible
            target_keypoints = torch.rand(n_targets, K, 3)
            target_keypoints[:, :, 2] = 2.0   # mark all keypoints fully visible

            target_classes = torch.zeros(n_targets, dtype=torch.long)  # single class (person)
            target_areas   = torch.tensor([0.05, 0.12])  # normalized box areas

            loc_loss, findable_loss, visible_loss, nll_loss = compute_l1_keypoint_loss(
                all_pred_keypoints,
                target_keypoints,
                target_classes,
                target_areas,
                num_keypoints_per_class=[17],
            )
            # Each output tensor has shape (2,) â€” one scalar loss per matched target.
            # loc_loss:      area-normalized mean L1 distance for visible keypoints
            # findable_loss: BCE for "annotator could locate this keypoint"
            # visible_loss:  BCE for "keypoint is fully visible (v==2)"
            # nll_loss:      Gaussian NLL incorporating Cholesky uncertainty parameters
    é   úIExpected all_pred_keypoints last dim >= 7 (x,y,2 logits + 3 chol params).r   zInum_keypoints_per_class must be non-empty when computing keypoint losses.zÂtarget_classes max index %d >= num_keypoints_per_class length %d; skipping keypoint loss for this batch to avoid crashing training. Check that your keypoint schema covers all annotation classes.r   Nztotal_padded_num_keypoints (z$) must be divisible by num_classes (ú))Údevice)ÚdtyperO   Té   r7   é   ©rP   ©Úminr   Únone)Ú	reductionç        é   é   é   zGNLL loss: %d keypoint(s) with non-finite uncertainty dropped from loss.ç      à?©ÚnanÚposinfÚneginf)(ÚshapeÚlenÚ
ValueErrorÚmaxÚloggerÚwarningr
   ÚreshapeÚsumÚ	new_zerosÚviewr#   ÚarangerO   ÚzerosÚboolÚ	enumerateÚtoÚfloat32ÚfinforP   ÚepsÚisfiniteÚallr8   ÚclampÚwhereÚ	ones_likeÚ	clamp_minÚsqrtÚ
zeros_likeÚFÚl1_lossÚmasked_fillÚ binary_cross_entropy_with_logitsÚdebugÚitemÚexpÚ
nan_to_numrU   )HrD   rE   rF   rG   rH   Ú	n_targetsÚtotal_padded_num_keypointsÚpred_dimÚnum_classesÚ
graph_zerorl   ÚkpadÚsplit_pred_keypointsÚselected_pred_keypointsÚactive_keypoints_maskÚ	class_idxÚnum_keypointsÚkeypoints_loss_maskÚareaÚarea_epsÚ
valid_areaÚvalid_xyÚvalid_visibilityÚlocation_loss_maskÚlocation_countÚvalid_countÚvalid_findableÚfindable_loss_maskÚfindable_countÚvalid_visible_predÚvisible_loss_maskÚvisible_countÚ	safe_areaÚsafe_area_sqrtÚzeros_xyÚsafe_pred_xyÚsafe_target_xyÚl1_per_keypointÚscaled_masked_l1Úlocation_lossÚsafe_pred_findableÚfindable_bceÚfindable_lossÚsafe_pred_visibleÚvisible_bceÚvisible_lossÚraw_log_l11Úraw_l21Úraw_log_l22Úfinite_uncertaintyÚgaussian_loss_maskÚsafe_pred_xy_nllÚsafe_target_xy_nllÚdxdyÚdxÚdyÚlog_l11Úl21Úlog_l22Úl11Úl22Úu0Úu1Úmaha2Úoverflow_safe_maskÚ	zeros_logÚgaussian_countÚgaussian_valid_countÚnll_rawÚnll_keypointsÚnll_lossÚno_validsH                                                                           r   Úcompute_l1_keypoint_lossrÅ   }   sÒ  € ðH 7IÔ6NÑ3€IÐ)¨8Ø�qŠ=ˆ=ˆ=Ðe‰=Œ=ˆ=ÝÐ-Ñ.Ô.€Kà�aÒÐÝÐdÑeÔeÐeà�1‚}€}˜×+Ò+Ñ-Ô-°Ò<Ð<Ý�ŠðMõ �×"Ò"Ñ$Ô$Ñ%Ô%Øñ	
ô 	
ð 	
ð( (×/Ò/°Ñ3Ô3°B°Q°BÔ7×;Ò;Ñ=Ô=ˆ
Ø"×,Ò,¨YÑ7Ô7¸*ÑDˆØ�e˜U EÐ)Ð)à! KÑ/°1Ò4Ð4Ýð,Ð+Eð ,ð ,Ø(ð,ð ,ð ,ñ
ô 
ð 	
ð &¨Ñ4€DØ-×2Ò2°9¸kÈ4ÐQYÑZÔZÐØ2ÝŒ�YÐ';Ô'BÐCÑCÔCÀ^ÐSôÐõ "œK¨°TÅÄÐThÔToÐpÑpÔpÐÝ$-Ð.EÑ$FÔ$Fð @ð @Ñ ˆ	�=Ø;?Ð˜i¨¨-¨Ð7Ñ8Ð8à/°Ô?ÐØ�?Š?�5œ=Ñ)Ô)€DÝŒ{˜4œ:Ñ&Ô&Ô*€HÝ” Ñ%Ô%¨°ªÑ9€JÝŒ~Ð5°a°a°a¸¸¸¸B¸Q¸B°hÔ?Ñ@Ô@×DÒDÈÐDÑLÔLÍuÌ~Ø˜˜˜˜A˜A˜A˜r ˜r˜Ô"ñPô Pç	‚cˆb€c�k„kñ€Hõ ”~Ð&6°q°q°q¸!¸!¸!¸Q°wÔ&?Ñ@Ô@ÐDTÐUVÐUVÐUVÐXYÐXYÐXYÐ[\ÐU\ÔD]Ð`aÒDaÑbÐØ,Ð/?Ñ?À(ÑJÈZ×MaÒMaÐbcÑMdÔMdÑdÐØ'×+Ò+¨BÑ/Ô/×2Ò2Ð9PÔ9VÐ2ÑWÔW€NØ ×&Ò&¨1Ð&Ñ-Ô-€KÝ”^Ð$;¸A¸A¸A¸q¸q¸qÀ!¸GÔ$DÑEÔE€NØ,¨~Ñ=ÐØ'×+Ò+¨BÑ/Ô/×5Ò5¸!Ð5Ñ<Ô<×?Ò?ÐF]ÔFcÐ?ÑdÔd€NÝœÐ(?ÀÀÀÀ1À1À1ÀaÀÔ(HÑIÔIÐØ+Ð.@Ñ@ÐØ%×)Ò)¨"Ñ-Ô-×3Ò3¸Ð3Ñ:Ô:×=Ò=ÐD[ÔDaÐ=ÑbÔb€Mõ ”˜J¨­e¬o¸dÑ.CÔ.CÑDÔD€IØ×(Ò(¨Ñ2Ô2×7Ò7Ñ9Ô9€Nõ ÔÐ 7¸¸¸¸1¸1¸1¸b¸q¸b¸Ô AÑBÔB€HÝ”;Ð1×;Ò;¸BÑ?Ô?ÐAXÐYZÐYZÐYZÐ\]Ð\]Ð\]Ð_aÐ`aÐ_aÐYaÔAbÐdlÑmÔm€LÝ”[Ð!3×!=Ò!=¸bÑ!AÔ!AÐCSÐTUÐTUÐTUÐWXÐWXÐWXÐZ\Ð[\ÐZ\ÐT\ÔC]Ð_gÑhÔh€NÝ”i ¨nÈÐOÑOÔO×SÒSÐTVÑWÔW€OØ&×2Ò2Ð4FÐ3FÈÑLÔLÈ~×OgÒOgÐhiÑOjÔOjÑjÐØ$×(Ò(¨Ñ,Ô,¨{Ñ:€MåœØÐ3°A°A°A°q°q°q¸!°GÔ<½eÔ>NÐOfÐghÐghÐghÐjkÐjkÐjkÐmnÐgnÔOoÑ>pÔ>pñô Ðõ Ô5ØØ	˜!˜!˜!˜Q˜Q˜Q ˜'Ô	" QÒ	&×*Ò*Ð+BÔ+HÑIÔIØðñ ô €Lð
 !×,Ò,Ð.@Ð-@À#ÑFÔF×JÒJÈ2ÑNÔNÐQ_Ñ_€MåœØÐ2°1°1°1°a°a°a¸°7Ô;½UÔ=MÐNeÐfgÐfgÐfgÐijÐijÐijÐlmÐfmÔNnÑ=oÔ=oñô Ðõ Ô4ØØ	˜!˜!˜!˜Q˜Q˜Q ˜'Ô	" QÒ	&×*Ò*Ð+BÔ+HÑIÔIØðñ ô €Kð
 ×*Ò*Ð,=Ð+=¸sÑCÔC×GÒGÈÑKÔKÈmÑ[€Là)¨!¨!¨!¨Q¨Q¨Q°¨'Ô2×5Ò5µe´mÑDÔD€KØ% a a a¨¨¨¨A gÔ.×1Ò1µ%´-Ñ@Ô@€GØ)¨!¨!¨!¨Q¨Q¨Q°¨'Ô2×5Ò5µe´mÑDÔD€KÝœ¨Ñ4Ô4µu´~ÀgÑ7NÔ7NÑNÕQVÔQ_Ð`kÑQlÔQlÑlÐØ×!Ò!Ñ#Ô#ð 
Ý�ŠØUØ Ð ×%Ò%Ñ'Ô'×,Ò,Ñ.Ô.ñ	
ô 	
ð 	
ð ,Ð.@Ñ@Ðõ ”{Ð#5×#?Ò#?ÀÑ#CÔ#CÐE\Ð]^Ð]^Ð]^Ð`aÐ`aÐ`aÐceÐdeÐceÐ]eÔEfÐhpÑqÔqÐÝœÐ%7×%AÒ%AÀ"Ñ%EÔ%EÐGWÐXYÐXYÐXYÐ[\Ð[\Ð[\Ð^`Ð_`Ð^`ÐX`ÔGaÐckÑlÔlÐØÐ1Ñ1×5Ò5µe´mÑDÔD€DØ	ˆaˆaˆa����AˆgŒ€BØ	ˆaˆaˆa����AˆgŒ€Bõ ŒkÐ,¨k½5Ô;KÈKÑ;XÔ;XÑYÔY€GÝ
Œ+Ð(¨'µ5Ô3CÀGÑ3LÔ3LÑ
MÔ
M€CÝŒkÐ,¨k½5Ô;KÈKÑ;XÔ;XÑYÔY€Gà
�+Š+‰-Œ-€CØ
�+Š+‰-Œ-€CØ	ˆr‰�C˜"‘HÑ	€BØ	ˆr‰€BØ�‰G�b˜2‘gÑ€Eð ,­e¬n¸RÑ.@Ô.@Ñ@Å5Ä>ÐRTÑCUÔCUÑUÕX]ÔXfÐglÑXmÔXmÑmÐÝÔ  Ñ-Ô-€IÝŒkÐ,¨g°yÑAÔA€GÝ
Œ+Ð(¨#¨yÑ
9Ô
9€CÝŒkÐ,¨g°yÑAÔA€GÝ	ŒÐ'¨­UÔ-=¸bÑ-AÔ-AÑ	BÔ	B€BÝ	ŒÐ'¨­UÔ-=¸bÑ-AÔ-AÑ	BÔ	B€BØ
�+Š+‰-Œ-€CØ
�+Š+‰-Œ-€CØ	ˆr‰�C˜"‘HÑ	€BØ	ˆr‰€BØ�‰G�b˜2‘gÑ€EØ+ÐØ'×+Ò+¨BÑ/Ô/×2Ò2Ð9PÔ9VÐ2ÑWÔW€NØ)×/Ò/°AÐ/Ñ6Ô6ÐØ�U˜Y×0Ò0°Ñ:Ô:×DÒDÀQÑGÔGÑGÑHÈGÐV]ÑL]Ñ^€GÝÔØ�S¥¤¨W¬]Ñ!;Ô!;Ô!?À!Ñ!CÍEÌKÐX_ÔXeÑLfÔLfÔLjðñ ô €Gð ×'Ò'Ð);Ð(;¸SÑAÔA€MØ× Ò  Ñ$Ô$Ð';Ñ;€HØ Ò"€HÝŒ{˜8¥UÔ%5°hÑ%?Ô%?ÀÑJÔJ€Hð ˜-¨°xÐ?Ð?r   ÚxÚyc                óâ   — |                      | j        ¬¦  «        }t          j        | ¦  «                             dd¬¦  «        }t          j        | |                     ¦   «         ¦  «        }||z
  S )z;Compute pairwise BCE-with-logits summed along the last dim.rS   rR   T)r   Úkeepdim)ro   rP   r{   Úsoftplusrh   r#   ÚmatmulÚt)rÆ   rÇ   Úy_floatrÊ   Údots        r   Ú_cdist_bce_with_logitsrÏ   w  sZ   € à�dŠd˜œˆdÑ!Ô!€GÝŒz˜!‰}Œ}× Ò  Q°Ð Ñ5Ô5€HÝ
Œ,�q˜'Ÿ)š)™+œ+Ñ
&Ô
&€CØ�c‰>Ðr   c           	     óò  — | j         \  }}}}|dk    s
J d¦   «         ‚|j         d         }	t          |¦  «        }
|
dk    rt          d¦  «        ‚|	dk    r*t          j        |||	f| j        | j        ¬¦  «        }||||fS ||
z  }|                      |||
||¦  «        }t          j        |||	f| j        | j        ¬¦  «        }t          j        |||	f| j        | j        ¬¦  «        }t          j        |||	f| j        | j        ¬¦  «        }t          j        |||	f| j        | j        ¬¦  «        }||z  }t          |
¦  «        D �]1}||k     	                    ¦   «          
                    d¦  «        }|                     ¦   «         dk    rŒG||         }|dk    rŒV|dd…dd…|d|…dd…f         }|                     d|¦  «        dd…d|…dd…f         }|j         d         }|                     d|¦  «                             t          j        ¦  «        }t          j        |j        ¦  «        j        }t          j        |¦  «        ||k    z  }|dd…dd…dd…f         }t          j        |dd…dd…df         ¦  «        |dd…dd…df         dk    z  t          j        |¦  «                             d	¬
¦  «        z  |                     d¦  «        z  }|                     d¬
¦  «                             t          j        ¦  «        }|                     d¬¦  «        }|dk    } |                     |¦  «                             ¦   «         }!|dd…dd…dd…dd…f                              ||d¦  «                             t          j        ¦  «        }"|                     t          j        ¦  «        }#|"                     d¦  «        |#                     d¦  «        z
  }$|$                     ¦   «                              d	¦  «        }%|                     d¦  «        }&|%                     |& d¦  «        }%|%                     d	¦  «        }'|'|                     d¦  «        z                       |!                     d¦  «        ¦  «        }(t          j        |(ddd¬¦  «        }(|(                     |||¦  «                             | j        ¦  «        |dd…dd…|f<   |dd…dd…dd…df                              ||¦  «                             t          j        ¦  «        })|dd…dd…dd…df                              ||¦  «                             t          j        ¦  «        }*|dd…dd…dd…df                              ||¦  «                             t          j        ¦  «        }+|)},|*}-|+}.|,                     ¦   «         }/|.                     ¦   «         }0t          j        |"¦  «                             d	¬
¦  «        }1|1t          j        |)¦  «        z  t          j        |*¦  «        z  t          j        |+¦  «        z  }2|$d         }3|$d         }4|/                     d¦  «        |3z  |-                     d¦  «        |4z  z   }5|0                     d¦  «        |4z  }6|5|5z  |6|6z  z   }7|&|2                     d¦  «        z  t          j        |5¦  «        z  t          j        |6¦  «        z  t          j        |7¦  «        z  }8d|7|                     |¦  «                             d|d¦  «        z  z  |,|.z                        d¦  «        z
  }9t          j        |9dt          j        |9j        ¦  «        j        dz  t          j        |9j        ¦  «        j        ¬¦  «        }9|9                     |8 d¦  «        }9|9                     d	¦  «        }:|:|                     d¦  «        z                       |||¦  «        };|;                      |                      d¦  «         d¦  «         |;                     | j        ¦  «        |dd…dd…|f<   |dd…dd…dd…df                              ||¦  «        }<|dd…dd…df         dk                         | j        ¦  «                             ||¦  «        }=|dd…dd…dd…df                              ||¦  «        }>|dd…dd…df         dk                         | j        ¦  «                             ||¦  «        }?tC          |<|=¦  «                             |||¦  «        tE          |¦  «        z  |dd…dd…|f<   tC          |>|?¦  «                             |||¦  «        tE          |¦  «        z  |dd…dd…|f<   �Œ3||||fS )uŸ  Compute many-to-many keypoint matching costs.

    Unlike ``compute_l1_keypoint_loss``, this function does not guard ``cost_findable``/
    ``cost_visible`` against a non-finite prediction: ``_cdist_bce_with_logits`` has no
    finiteness mask, so a single non-finite logit poisons the whole cost row for that query
    against every target, which then feeds the Hungarian assignment. Tracked separately since
    it needs a different fix shape (the helper's signature would need to accept a mask).

    Args:
        all_pred_keypoints: Predicted keypoints of shape ``(B, Q, K_total, >=7)``.
        target_keypoints: Ground truth keypoints with shape ``(N, Kmax, 3)``.
        target_classes: Class ids for each target with shape ``(N,)``.
        target_areas: Target box areas with shape ``(N,)``.
        num_keypoints_per_class: Number of keypoints per class.

    Returns:
        Tuple ``(cost_l1, cost_findable, cost_visible, cost_nll)`` each of shape ``(B, Q, N)``.

    Example:
        Compute matching costs for 1 image with 4 decoder queries against 2 ground-truth
        targets, each with 17 COCO keypoints:

        .. code-block:: python

            B, Q, K_total, pred_dim = 1, 4, 17, 7
            n_targets = 2

            all_pred_keypoints = torch.randn(B, Q, K_total, pred_dim)

            target_keypoints = torch.rand(n_targets, K_total, 3)
            target_keypoints[:, :, 2] = 2.0   # mark all keypoints fully visible

            target_classes = torch.zeros(n_targets, dtype=torch.long)  # single class (person)
            target_areas   = torch.tensor([0.05, 0.12])

            cost_l1, cost_findable, cost_visible, cost_nll = compute_keypoint_matching_cost(
                all_pred_keypoints,
                target_keypoints,
                target_classes,
                target_areas,
                num_keypoints_per_class=[17],
            )
            # Each output tensor has shape (1, 4, 2) â€” (B, Q, N).
            # cost_l1[:, q, n]  is the area-normalized L1 cost for query q against target n.
            # cost_nll[:, q, n] incorporates the Cholesky precision uncertainty.
            # These cost matrices are passed to the Hungarian matcher.
    rL   rM   r   zQnum_keypoints_per_class must be non-empty when computing keypoint matching costs.)rO   rP   rR   NrQ   r   r7   rT   rX   r]   rY   rZ   r[   ).r   ).rR   r\   r   )#ra   rb   rc   r#   rl   rO   rP   rj   ÚrangeÚnonzeroÚsqueezeÚnumelÚindex_selectro   rp   rq   rr   rs   rt   r8   rh   ru   rx   ry   rg   Úabsr}   Údivr‚   r�   rd   rU   Úmasked_fill_rÏ   Úfloat)@rD   rE   rF   rG   rH   Úbr   Útotal_num_keypointsr…   rƒ   r†   rl   rˆ   ÚpredÚcost_l1Úcost_findableÚcost_visibleÚcost_nllÚflat_bqrŒ   Útarget_indicesÚnum_kptsÚpred_by_classÚtarget_by_classÚn_targets_by_classÚareasr�   r‘   Ú	target_xyÚvisibleÚvalid_per_targetÚ	nll_denomÚhas_visibleÚ	area_sqrtÚpred_xy_flatÚtarget_xy_f32ÚdiffÚ
per_kpt_l1Úvisible_btkÚ
scaled_locÚloc_costr«   r¬   r­   rµ   r¶   r·   r¸   r¹   Ú	finite_xyÚfinite_predr³   r´   rº   r»   r¼   Úkeypoint_maskÚnll_kÚnll_sumÚmean_nllÚpred_findableÚtarget_findableÚpred_visibleÚtarget_visibles@                                                                   r   Úcompute_keypoint_matching_costrÿ     sÙ	  € ðn 5GÔ4LÑ1€A€{Ð'¨Ø�qŠ=ˆ=ˆ=Ðe‰=Œ=ˆ=Ø Ô& qÔ)€IÝÐ-Ñ.Ô.€Kà�aÒÐÝÐlÑmÔmÐmà�A‚~€~Ý”Ø�˜YÐ'Ø%Ô,Ø$Ô*ð
ñ 
ô 
ˆð
 �e˜U EÐ)Ð)à +Ñ-€DØ×"Ò" 1 k°;ÀÀhÑOÔO€DåŒkØ	
ˆK˜Ð#Ø!Ô(Ø Ô&ðñ ô €Gõ
 ”KØ	
ˆK˜Ð#Ø!Ô(Ø Ô&ðñ ô €Mõ
 ”;Ø	
ˆK˜Ð#Ø!Ô(Ø Ô&ðñ ô €Lõ
 Œ{Ø	
ˆK˜Ð#Ø!Ô(Ø Ô&ðñ ô €Hð �+‰o€GÝ˜;Ñ'Ô'ð sñ sˆ	Ø(¨IÒ5×>Ò>Ñ@Ô@×HÒHÈÑKÔKˆØ×ÒÑ!Ô! QÒ&Ð&ØØ*¨9Ô5ˆØ�qŠ=ˆ=Øà˜Q˜Q˜Q    9¨i¨x¨i¸¸¸Ð:Ô;ˆØ*×7Ò7¸¸>ÑJÔJÈ1È1È1ÈiÈxÈiÐYZÐYZÐYZÈ?Ô[ˆØ,Ô2°1Ô5Ðà×)Ò)¨!¨^Ñ<Ô<×?Ò?ÅÄÑNÔNˆÝ”;˜uœ{Ñ+Ô+Ô/ˆÝ”^ EÑ*Ô*¨e°hÒ.>Ñ?ˆ
Ø# A A A q q q¨"¨1¨" HÔ-ˆ	åŒN˜?¨1¨1¨1¨a¨a¨a°¨7Ô3Ñ4Ô4Ø˜q˜q˜q ! ! ! Q˜wÔ'¨!Ò+ñ-åŒn˜YÑ'Ô'×+Ò+°Ð+Ñ3Ô3ñ4ð ×"Ò" 1Ñ%Ô%ñ&ð 	ð #Ÿ;š;¨1˜;Ñ-Ô-×0Ò0µ´Ñ?Ô?ÐØ$×*Ò*¨qÐ*Ñ1Ô1ˆ	Ø&¨Ò*ˆà—O’O HÑ-Ô-×2Ò2Ñ4Ô4ˆ	ð % Q Q Q¨¨¨¨1¨1¨1¨b¨q¨b [Ô1×9Ò9¸'À8ÈQÑOÔO×RÒRÕSXÔS`ÑaÔaˆØ!Ÿš¥U¤]Ñ3Ô3ˆð ×%Ò% aÑ(Ô(¨=×+BÒ+BÀ1Ñ+EÔ+EÑEˆØ—X’X‘Z”Z—^’^ BÑ'Ô'ˆ
Ø×'Ò'¨Ñ*Ô*ˆØ×+Ò+¨[¨L¸#Ñ>Ô>ˆ
Ø—^’^ BÑ'Ô'ˆ
à ×!4Ò!4°QÑ!7Ô!7Ñ7×<Ò<¸Y×=PÒ=PÐQRÑ=SÔ=SÑTÔTˆÝÔ# H°#¸cÈ#ÐNÑNÔNˆØ(0×(8Ò(8ØØØñ)
ô )
÷ Š"ÐÔ%Ñ
&Ô
&ð	 	����1�1�1�nÐ$Ñ%ð $ A A A q q q¨!¨!¨!¨Q JÔ/×7Ò7¸ÀÑJÔJ×MÒMÍeÌmÑ\Ô\ˆØ    1 1 1 a a a¨ 
Ô+×3Ò3°G¸XÑFÔF×IÒIÍ%Ì-ÑXÔXˆØ# A A A q q q¨!¨!¨!¨Q JÔ/×7Ò7¸ÀÑJÔJ×MÒMÍeÌmÑ\Ô\ˆàˆØˆØˆØ�kŠk‰mŒmˆØ�kŠk‰mŒmˆå”N <Ñ0Ô0×4Ò4¸Ð4Ñ<Ô<ˆ	Ø¥%¤.°Ñ"=Ô"=Ñ=ÅÄÈwÑ@WÔ@WÑWÕZ_ÔZhÐitÑZuÔZuÑuˆà�&Œ\ˆØ�&Œ\ˆØ�]Š]˜1ÑÔ Ñ" S§]¢]°1Ñ%5Ô%5¸Ñ%:Ñ:ˆØ�]Š]˜1ÑÔ Ñ"ˆØ�R‘˜"˜r™'Ñ!ˆð ˜+×/Ò/°Ñ2Ô2Ñ2µU´^ÀBÑ5GÔ5GÑGÍ%Ì.ÐY[ÑJ\ÔJ\Ñ\Õ_dÔ_mÐnsÑ_tÔ_tÑtð 	ð �u˜uŸš¨xÑ8Ô8×=Ò=¸aÐASÐUVÑWÔWÑWÑXØ�gÑß
Š)�A‰,Œ,ñˆõ Ô Ø�s¥5¤;¨u¬{Ñ#;Ô#;Ô#?À!Ñ#CÍEÌKÐX]ÔXcÑLdÔLdÔLhð
ñ 
ô 
ˆð ×!Ò! = .°#Ñ6Ô6ˆØ—)’)˜B‘-”-ˆà˜i×1Ò1°!Ñ4Ô4Ñ4×=Ò=¸aÀÐN`ÑaÔaˆØ×Ò˜{×4Ò4°QÑ7Ô7Ð7¸Ñ=Ô=Ð=Ø)1¯ªÐ5GÔ5MÑ)NÔ)Nˆ����A�A�A�~Ð%Ñ&à% a a a¨¨¨¨A¨A¨A¨q jÔ1×9Ò9¸'À8ÑLÔLˆà˜Q˜Q˜Q    1˜WÔ%¨Ò)ßŠRÐ"Ô(Ñ)Ô)ßŠWØ"Øñô ð 	ð % Q Q Q¨¨¨¨1¨1¨1¨a ZÔ0×8Ò8¸À(ÑKÔKˆà˜Q˜Q˜Q    1˜WÔ%¨Ò)ßŠRÐ"Ô(Ñ)Ô)ßŠWØ"Øñô ð 	õ /EÀ]ÐTcÑ.dÔ.d×.lÒ.lØØØñ/
ô /
õ �(‰OŒOñ	/ˆ�a�a�a˜˜˜˜NÐ*Ñ+õ
 .DÀLÐR`Ñ-aÔ-a×-iÒ-iØØØñ.
ô .
õ �(‰OŒOñ	.ˆ�Q�Q�Q˜˜˜˜>Ð)Ñ*Ñ*ð �M <°Ð9Ð9r   )r   rÿ   rÅ   )r   r   r   r   r   r   r   r   )rD   r   rE   r   rF   r   rG   r   rH   rI   r   rJ   )rÆ   r   rÇ   r   r   r   )rB   Ú
__future__r   Úcollections.abcr   Útypingr   r#   Útorch.nn.functionalr   Ú
functionalr{   r   Úrfdetr.utilities.loggerr   re   r   Ú__annotations__r   ÚModuler   rÅ   rÏ   rÿ   Ú__all__r   r   r   ú<module>r	     sY  ððð ð ð  #Ð "Ð "Ð "Ð "Ð "à $Ð $Ð $Ð $Ð $Ð $Ø Ð Ð Ð Ð Ð à €€€Ø Ð Ð Ð Ð Ð Ð Ð Ð Ø Ð Ð Ð Ð Ð Ð Ð à .Ð .Ð .Ð .Ð .Ð .à	ˆ‰Œ€ð Ð Ð Ð Ð Ñ ð,ð ,ð ,ð ,ð878ð 78ð 78ð 78ð 78 "¤)ñ 78ô 78ð 78ðtw@ð w@ð w@ð w@ðtð ð ð ðW:ð W:ð W:ð W:ðtð ð €€€r   