§
    ­¸jz‹  ã                  ó*  — U d Z ddlmZ ddlZddlZddlmZmZ ddlZddl	m
c mZ ddlmZ ddl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 ddlmZmZmZ  e¦   «         Zg d¢Z dZ!dZ"de#d<   d4d„Z$d5d„Z%d6d!„Z&d7d&„Z'd8d*„Z(d+d,œd9d2„Z)d:d3„Z*dS );a8  Shared weight-loading and LoRA application utilities.

Provides the canonical implementations of pretrained checkpoint loading and LoRA adapter injection, used by both the L1
inference facade (``rfdetr.detr``) and the L2 LightningModule (``rfdetr.training.module_model``).

The weight-loading logic is taken from ``RFDETRModelModule._load_pretrain_weights`` in ``module_model.py`` (more
complete: Pydantic-aware user-override detection, auto-alignment for fine-tuned checkpoints) and augmented with class-
name extraction from ``detr.py:_load_pretrain_weights_into``.
é    )ÚannotationsN)ÚAnyÚcast)ÚTensor)Údownload_pretrain_weightsÚvalidate_pretrain_weights)ÚModelConfig)ÚBackbone)ÚDinoV2)ÚLWDETR)Ú
get_logger)Ú_ckpt_args_getÚremap_projector_to_cross_attnÚ!validate_checkpoint_compatibility)Úload_pretrain_weightsÚ
apply_loraÚinterpolate_position_embeddingszembeddings.position_embeddings)zrefpoint_embed.weightzquery_feat.weightztuple[str, ...]Ú_QUERY_PARAM_SUFFIXESÚnameÚstrÚreturnÚboolc                ó6   — |                       t          ¦  «        S )a  Return whether *name* is a state-dict key for a packed query parameter.

    Single source of truth for the ``_QUERY_PARAM_SUFFIXES`` suffix match, so every call site agrees on which keys are
    reshaped per group.

    Args:
        name: A checkpoint / model ``state_dict`` key.

    Returns:
        ``True`` when the key ends with one of ``_QUERY_PARAM_SUFFIXES``.

    Examples:
        >>> _is_query_param("transformer.refpoint_embed.weight")
        True
        >>> _is_query_param("class_embed.weight")
        False
    )Úendswithr   )r   s    úe/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/rfdetr/models/weights.pyÚ_is_query_paramr   -   s   € ð$ �=Š=Õ.Ñ/Ô/Ð/ó    Útensorr   Úckpt_num_queriesÚintÚckpt_group_detrÚtarget_num_queriesÚtarget_group_detrc                óÊ  ‡ ‡‡— ‰dk    s|dk    s|dk    s|dk    rt          d‰› d|› d|› d|› d�	¦  «        ‚‰|z  }‰ j        d         |k    r6t                               d‰||‰ j        d         ¦  «         ‰ d||z  …         S |‰k    r||k    r‰ S t	          ||¦  «        }t	          |‰¦  «        Šˆˆˆ fd	„t          |¦  «        D ¦   «         }t          j        |d¬
¦  «        S )u˜	  Slice a ``refpoint_embed`` / ``query_feat`` weight preserving per-group structure.

    ``LWDETR`` packs query embeddings as ``nn.Embedding(num_queries * group_detr, ...)`` where group ``g`` occupies the
    contiguous slot range ``[g * num_queries, (g + 1) * num_queries)`` (see ``LWDETR.__init__`` and ``LWDETR.forward``
    in ``models/lwdetr.py``).  When ``num_queries`` decreases and ``group_detr > 1``, a flat ``tensor[:
    target_num_queries * target_group_detr]`` slice silently scrambles groups: the tail of group 0 winds up in what
    should be group 1's slots, and so on.  At inference only group 0 is read so the bug is invisible, but for
    training-resume it corrupts groups 1+.

    This helper does the right thing per group:

    * ``num_queries`` decrease (``target_num_queries < ckpt_num_queries``) â†’
      keep the first ``target_num_queries`` slots of each retained group.
    * ``group_detr`` decrease (``target_group_detr < ckpt_group_detr``) â†’
      drop tail groups; retained groups stay pretrained.
    * Either dimension expands, or one shrinks while the other expands â†’
      return whatever per-group sub-tensor can be built (``min(target, ckpt)`` along each axis). The result has fewer
      rows than the model expects, so ``load_state_dict`` will raise a shape mismatch immediately.

    When the tensor's flat length disagrees with ``ckpt_num_queries * ckpt_group_detr`` (corrupt or unexpected
    checkpoint shape), fall back to the legacy flat slice so loading continues with the same behavior the codebase had
    before this fix.

    Args:
        tensor: The checkpoint tensor for ``refpoint_embed.weight`` or
            ``query_feat.weight``.
        ckpt_num_queries: ``num_queries`` recorded in the checkpoint's training args.
        ckpt_group_detr: ``group_detr`` recorded in the checkpoint's training args.
        target_num_queries: ``num_queries`` configured for the model.
        target_group_detr: ``group_detr`` configured for the model.

    Returns:
        A tensor whose layout matches the model's configured packing for the decrease-or-equal cases, or a per-group
        sub-tensor built from ``min(target, ckpt)`` along each axis for the expansion case (which ``load_state_dict``
        will then reject on shape mismatch).

    Raises:
        ValueError: If any of ``ckpt_num_queries``, ``ckpt_group_detr``,
            ``target_num_queries``, or ``target_group_detr`` is â‰¤ 0.
    r   zX_slice_query_param_per_group: all dimension args must be positive; got ckpt_num_queries=z, ckpt_group_detr=z, target_num_queries=z, target_group_detr=ú.u²   _slice_query_param_per_group: checkpoint args claim %d Ã— %d = %d rows but tensor has %d rows; falling back to flat slice. Per-group structure may be scrambled if group_detr > 1.Nc                ó6   •— g | ]}‰|‰z  |‰z  ‰z   …         ‘ŒS © r'   )Ú.0Úgr   Úkeep_per_groupr   s     €€€r   ú
<listcomp>z0_slice_query_param_per_group.<locals>.<listcomp>‹   s5   ø€ ÐoÐoÐoÐWXˆf�QÐ)Ñ)¨AÐ0@Ñ,@À>Ñ,QÐQÔRÐoÐoÐor   ©Údim)Ú
ValueErrorÚshapeÚloggerÚwarningÚminÚrangeÚtorchÚcat)	r   r   r!   r"   r#   Úexpected_totalÚkeep_groupsÚpiecesr*   s	   ``      @r   Ú_slice_query_param_per_groupr9   B   so  øøø€ ð^ ˜1ÒÐ °1Ò 4Ð 4Ð8JÈaÒ8OÐ8OÐSdÐhiÒSiÐSiÝð_Ø$4ð_ð _ØHWð_ð _à"4ð_ð _àJ[ð_ð _ð _ñ
ô 
ð 	
ð &¨Ñ7€NØ„|�A„˜.Ò(Ð(å�Šð2ð ØØØŒL˜ŒOñ	
ô 	
ð 	
ð Ð>Ð*Ð->Ñ>Ð>Ô?Ð?àÐ-Ò-Ð-Ð2CÀÒ2VÐ2VØˆåÐ'¨Ñ9Ô9€KÝÐ+Ð-=Ñ>Ô>€NØoÐoÐoÐoÐoÐoÕ\aÐbmÑ\nÔ\nÐoÑoÔo€FÝŒ9�V Ð#Ñ#Ô#Ð#r   Úkeysú	list[str]c                óR   ‡‡— ddgt           ¢d‘d‘d‘R Šdˆfd
„Šˆfd„| D ¦   «         S )u;  Return *keys* with intentional-reinit/trim entries removed.

    Matching is boundary-aware: a pattern matches a key when the pattern appears at the start of the key or immediately
    after a module separator (``.``).  This prevents substring collisions where a pattern like ``"class_embed."`` would
    inadvertently match a key belonging to an unrelated module (e.g. ``"class_embed_projection.weight"`` is safe because
    ``class_embed_projection.`` â‰  ``class_embed.``, but using a plain ``in`` check against longer ambiguous strings is
    fragile by design).
    zclass_embed.zbbox_embed.zenc_out_class_embed.zenc_out_bbox_embed.zkeypoint_head.keypoint_proj.Úkeyr   r   r   c                ó<   •‡ — t          ˆ fd„‰D ¦   «         ¦  «        S )Nc              3  óP   •K  — | ] }‰                      |¦  «        pd |› �‰v V — Œ!dS )r%   N)Ú
startswith)r(   Úpatr=   s     €r   ú	<genexpr>zD_filter_intentional_keys.<locals>._is_intentional.<locals>.<genexpr>ª   s?   øè è € Ð[Ð[¸s�3—>’> #Ñ&Ô&Ð:¨)¨c¨)¨)°sÐ*:Ð[Ð[Ð[Ð[Ð[Ð[r   )Úany)r=   Úintentional_patternss   `€r   Ú_is_intentionalz1_filter_intentional_keys.<locals>._is_intentional©   s(   øø€ ÝÐ[Ð[Ð[Ð[ÐFZÐ[Ñ[Ô[Ñ[Ô[Ð[r   c                ó*   •— g | ]} ‰|¦  «        °|‘ŒS r'   r'   )r(   ÚkrE   s     €r   r+   z,_filter_intentional_keys.<locals>.<listcomp>¬   s(   ø€ Ð6Ð6Ð6�! ? ?°1Ñ#5Ô#5Ð6ˆAÐ6Ð6Ð6r   )r=   r   r   r   )r   )r:   rE   rD   s    @@r   Ú_filter_intentional_keysrH   �   s�   øø€ ð 	Øð
-õ 
ð
-ð 	ð	
-ð
 	ð
-ð 	'ð
-ð 
-Ðð\ð \ð \ð \ð \ð \ð 7Ð6Ð6Ð6�tÐ6Ñ6Ô6Ð6r   Úincompatibler   Úpretrain_weights_pathÚNonec                óz  — t          | dd¦  «        }t          | dd¦  «        }	 |rd„ |D ¦   «         ng }|rd„ |D ¦   «         ng }n# t          $ r Y dS w xY wg }t          |¦  «        D ]O}|dk    s|                     d¦  «        rt                               d||¦  «         Œ:|                     |¦  «         ŒPt          |¦  «        }|s|sdS g }	|r]d	                     |dd
…         ¦  «        }
t          |¦  «        d
k    r|
dz  }
|	                     t          |¦  «        › d|
› d�¦  «         |r]d	                     |dd
…         ¦  «        }
t          |¦  «        d
k    r|
dz  }
|	                     t          |¦  «        › d|
› d�¦  «         t           	                    d|d                     |	¦  «        ¦  «         dS )uF  Emit a ``logger.warning`` when ``load_state_dict`` left non-trivial gaps.

    ``load_state_dict(strict=False)`` silently ignores keys that the model has but the checkpoint does not
    (``missing_keys``) and keys present in the checkpoint but absent from the model (``unexpected_keys``).  When this
    happens for parameters outside the head / query embeddings â€” which the loader intentionally reinitialises or trims â€”
    the corresponding model weights were left at their random initial values and the user is silently getting a much
    weaker model. This helper surfaces that condition with a single, actionable warning. Same-key shape mismatches do
    not reach this function â€” they raise :class:`RuntimeError` directly from ``load_state_dict`` and are therefore
    impossible to miss.

    Two separate exclusions are applied before deciding whether to warn, for two different reasons:

    * :func:`_filter_intentional_keys` drops head / query-embedding keys in *both* directions, because the loader
      itself reinitialises or trims them.  Matching there is module-prefix based.
    * ``_kp_active_mask`` is dropped from ``missing_keys`` only, because it is a deterministic schema buffer rather
      than a learned parameter â€” the model always rebuilds it from the configured keypoint schema, so a checkpoint
      predating keypoint support is not actually missing anything.  Matching is on the exact terminal key so a
      similarly-named real parameter still warns, and the ``unexpected_keys`` direction is deliberately left alone.

    Args:
        incompatible: The ``_IncompatibleKeys`` namedtuple returned by
            :meth:`torch.nn.Module.load_state_dict`.
        pretrain_weights_path: Path to the checkpoint that was loaded; included
            in the warning so the user can identify which load partially succeeded.
    Úmissing_keysNÚunexpected_keysc                ó,   — g | ]}t          |¦  «        ‘ŒS r'   ©r   ©r(   rG   s     r   r+   z)_warn_on_partial_load.<locals>.<listcomp>Ì   s   € Ð9Ð9Ð9 1�˜A™œÐ9Ð9Ð9r   c                ó,   — g | ]}t          |¦  «        ‘ŒS r'   rP   rQ   s     r   r+   z)_warn_on_partial_load.<locals>.<listcomp>Í   s   € Ð?Ð?Ð? a�3˜q™6œ6Ð?Ð?Ð?r   Ú_kp_active_maskz._kp_active_maskuN   Checkpoint %r has no %r â€” rebuilding it from the configured keypoint schema.z, é   z, ...z> model parameter(s) not in checkpoint (left at random init): [ú]z+ checkpoint key(s) not consumed by model: [uü   Pretrained weights at %r loaded only partially â€” this typically produces lower accuracy. %s. Check that the model configuration (encoder, hidden_dim, out_feature_indexes, projector_scale, ...) matches the architecture the checkpoint was trained with.ú )
ÚgetattrÚ	TypeErrorrH   r   r0   ÚdebugÚappendÚjoinÚlenr1   )rI   rJ   Úmissing_keys_rawÚunexpected_keys_rawrM   rN   Úmissingr=   Ú
unexpectedÚpartsÚsamples              r   Ú_warn_on_partial_loadrc   ¯   s4  € õ4 ˜|¨^¸TÑBÔBÐÝ! ,Ð0AÀ4ÑHÔHÐðØ=MÐUÐ9Ð9Ð(8Ð9Ñ9Ô9Ð9ÐSUˆØCVÐ^Ð?Ð?Ð+>Ð?Ñ?Ô?Ð?Ð\^ˆˆøÝð ð ð àˆˆðøøøð €GÝ'¨Ñ5Ô5ð 
ð 
ˆØÐ#Ò#Ð# s§|¢|Ð4FÑ'GÔ'GÐ#õ �LŠLØ`Ø%Øñô ð ð
 Ø�Š�sÑÔÐÐÝ)¨/Ñ:Ô:€JØð ˜:ð Øˆà€EØð oØ—’˜7 2 A 2œ;Ñ'Ô'ˆÝˆw‰<Œ<˜!ÒÐØ�gÑˆFØ�Š�˜G™œÐmÐmÐdjÐmÐmÐmÑnÔnÐnØð _Ø—’˜: b q bœ>Ñ*Ô*ˆÝˆz‰?Œ?˜QÒÐØ�gÑˆFØ�Š�˜J™œÐ]Ð]ÐTZÐ]Ð]Ð]Ñ^Ô^Ð^å
‡N‚Nð	'ð 	Ø�Š�‰Œñô ð ð ð s   ¤ A Á
AÁAÚcheckpoint_stateúdict[str, Any]Úpe_sizec           
     óØ  — ||z  }d„ | D ¦   «         }|D �]Ô}| |         }|j         d         dz
  }||k    rŒ"t          t          j        |¦  «        ¦  «        }t          t          j        |¦  «        ¦  «        }||z  |k    s	||z  |k    r%t                               d|› d|› d|› d�¦  «         Œ›|j         d         }	|dd…dd…f         }
|dd…dd…f         }|                     d|||	¦  «                             d	d
dd¦  «        }t          j	        | 
                    ¦   «         ||fdd|j        j        dk    ¬¦  «                             |j        ¦  «        }|                     d	dd
d¦  «                             d||	¦  «        }t          j        |
|gd¬¦  «        | |<   t                               d|t%          |j         ¦  «        t%          | |         j         ¦  «        ¦  «         �ŒÖdS )ut  Interpolate DINOv2 positional embeddings in *checkpoint_state* to match *pe_size*.

    When the model is configured with a custom ``resolution`` that differs from the checkpoint's training resolution,
    the DINOv2 backbone's ``position_embeddings`` parameter has an incompatible shape.
    ``load_state_dict(strict=False)`` does **not** skip shape mismatches on matching keys â€” it raises ``RuntimeError``.

    This function bicubic-interpolates every PE tensor in the checkpoint whose shape differs from the target grid,
    modifying *checkpoint_state* in-place before ``load_state_dict`` is called.

    Args:
        checkpoint_state: The ``"model"`` sub-dict from a loaded checkpoint.
        pe_size: Target grid side length in patches (number of patches per spatial
            dimension, assuming a square grid).  Typically ``model_config.positional_encoding_size``.
    c                óF   — g | ]}|                      t          ¦  «        ¯|‘ŒS r'   )r   Ú_PE_KEY_SUFFIXrQ   s     r   r+   z3interpolate_position_embeddings.<locals>.<listcomp>  s)   € ÐIÐIÐI�Q¨a¯jªj½Ñ.HÔ.HÐIˆqÐIÐIÐIr   é   zSkipping PE interpolation for z,: grid size is not a perfect square (source z	, target z).éÿÿÿÿNr   é   é   ÚbicubicFÚmps)ÚsizeÚmodeÚalign_cornersÚ	antialiasr,   u1   Interpolated positional embeddings %s: %s â†’ %s.)r/   r    ÚmathÚisqrtr0   r1   ÚreshapeÚpermuteÚFÚinterpolateÚfloatÚdeviceÚtypeÚtoÚdtyper4   r5   rY   Útuple)rd   rf   Ún_targetÚpe_keysr=   Úckpt_peÚn_sourceÚh_srcÚh_tgtr-   Úclass_tokenÚpatch_pes               r   r   r   û   sD  € ð$ ˜Ñ €HàIÐIÐ*ÐIÑIÔI€GØð #
ñ #
ˆØ" 3Ô'ˆØ”= Ô# aÑ'ˆØ�xÒÐØå•D”J˜xÑ(Ô(Ñ)Ô)ˆÝ•D”J˜xÑ(Ô(Ñ)Ô)ˆØ�5‰=˜HÒ$Ð$¨°©¸Ò(AÐ(AÝ�NŠNð^°ð ^ð ^Ø>Fð^ð ^ØQYð^ð ^ð ^ñô ð ð àŒm˜BÔˆØ˜a˜a˜a  ! ˜e”nˆØ˜1˜1˜1˜a˜b˜b˜5”>ˆà×#Ò# A u¨e°SÑ9Ô9×AÒAÀ!ÀQÈÈ1ÑMÔMˆÝ”=Ø�NŠNÑÔØ˜�ØØØ”oÔ*¨eÒ3ð
ñ 
ô 
÷ Š"ˆWŒ]Ñ
Ô
ð 	ð ×#Ò# A q¨!¨QÑ/Ô/×7Ò7¸¸8ÀSÑIÔIˆå %¤	¨;¸Ð*AÀqÐ IÑ IÔ IÐ˜ÑÝ�ŠØ?ØÝ�'”-Ñ Ô ÝÐ" 3Ô'Ô-Ñ.Ô.ñ		
ô 	
ð 	
ñ 	
ð=#
ð #
r   F©ÚtrustÚnn_modelr   Úmodel_configr	   r‰   c          	     ó0  ‡.— |}|j         €g S t          |j         ¦  «        }g }ddlm} t	          |¦  «         t
          j                             |¦  «        s+t           	                    d¦  «         t	          |d¬¦  «         t          |d¬¦  «         	  |||¬	¦  «        }nH# t          $ r; t                               d
¦  «         t	          |d¬¦  «          |||¬	¦  «        }Y nw xY wd|vrÎd|v rÊt                               d¦  «         d}d}	i }
|d                              ¦   «         D ]b\  }}|                     |¦  «        rH|t!          |¦  «        d…         }|                     |	¦  «        r|t!          |	¦  «        d…         }||
|<   Œc|
st#          d|›d�¦  «        ‚|
|d<   d|vrd|v r|d         |d<   d|v r`t%          |d         d¦  «        }|rHt'          |t          ¦  «        r|g}n/	 t)          |¦  «        }d„ |D ¦   «         }n# t*          $ r g }Y nw xY wt-          ||¦  «         dt/          |dt1          ¦   «         ¦  «        v }|j        }|d         d         j        d         }|dz   }||k    r+||k     r|s|dz
  }|}||_        |                      |¦  «         |                     d¦  «        }|�t%          |d¦  «        nd}|�t%          |d¦  «        nd}	 |�t;          |¦  «        nd}|�t;          |¦  «        nd}n5# t*          t"          f$ r! t           	                    d¦  «         d}d}Y nw xY w|du |du k    r¢t=          d„ |d         D ¦   «         d¦  «        }|�€|d         |         j        d         }d}|� |dk    r||z  dk    r||z  }d|d|f\  }}}}n!|�|dk    r||z  dk    r||z  }d|d|f\  }}}}|� t           	                    d||||||¦  «         |j        |j         z  Š.|j         dk    rW|�|€StC          ˆ.fd„|d                              ¦   «         D ¦   «         ¦  «        r t           	                    d |j         ¦  «         tE          |d          #                    ¦   «         ¦  «        D ]^}tI          |¦  «        rM|d         |         } |�*|�(tK          | |||j        |j         ¬!¦  «        |d         |<   ŒK| d‰.…         |d         |<   Œ_tM          |d         | ¦  «        |d<   d"t/          |dt1          ¦   «         ¦  «        v }!|!�sQt/          |d#d¦  «        �r?tO          |d"¦  «        �r.|d                              d$¦  «        }"t'          |"tP          ¦  «        r¼|"j)        d%k    r±d&„ |" *                    d¬'¦  «         +                    ¦   «         D ¦   «         }#tE          t/          |d"g ¦  «        pg ¦  «        }$tC          d(„ |#D ¦   «         ¦  «        st           	                    d)|#¦  «         nl|#|$k    r#t                               d*|$|#¦  «         |#|_,        nBt'          |"tP          ¦  «        r-t           	                    d+t[          |"j        ¦  «        ¦  «         tE          t/          |d"g ¦  «        pg ¦  «        }%d}&d}'t/          |d#d¦  «        r¦tO          | d,¦  «        r–|  .                    |d         ¦  «        }&|&ryt/          | d-d¦  «        }(t_          |(¦  «        r
 |(¦   «         n|%})|&|)k    rGtO          | d.¦  «        r7t           	                    d/|)|&¦  «         |  0                    |&¦  «         |&|%k    }'|d                              d$¦  «        }*tO          | d¦  «        r|  1                    ¦   «         ni }+t'          |+td          ¦  «        r|+                     d$¦  «        nd},t'          |*tP          ¦  «        rƒt'          |,tP          ¦  «        rn|*j        |,j        k    r^|'s\t           	                    d0t[          |*j        ¦  «        t[          |,j        ¦  «        ¦  «         |d          3                    d$d¦  «         ti          |d         |j5        ¦  «         |  6                    |d         d¬¦  «        }-to          |-|¦  «         |'r@tO          | d.¦  «        r0t           	                    d1|%¦  «         |  0                    |%¦  «         ||k     r|r|                      |¦  «         |dz   |k     r|                      |dz   ¦  «         |S )2a‰  Load pretrained checkpoint weights into *nn_model* in-place.

    Canonical implementation shared by the L1 facade (``_build_model_context`` in ``rfdetr.detr``) and the L2
    LightningModule (``RFDETRModelModule.__init__`` in ``rfdetr.training.module_model``).

    Uses the Pydantic-aware logic from ``module_model.py``:

    - When the user did **not** explicitly set ``num_classes`` (it is left unset),
      the checkpoint class count is treated as authoritative and the model head is auto-aligned to it.
    - When the user **did** explicitly set ``num_classes`` (to any value, including the
      class default) larger than the checkpoint provides, the head is temporarily aligned to the checkpoint for
      loading, then expanded back to the configured size.
    - When the checkpoint has more classes than configured (backbone-pretrain
      scenario), both reinitializations are applied: expand to checkpoint size for loading, then trim to configured
      size.

    Class names stored in the checkpoint ``args`` are extracted and returned.

    Args:
        nn_model: The model whose weights will be updated in-place.
        model_config: Pydantic ``ModelConfig`` instance. Must have
            ``pretrain_weights``, ``num_classes``, ``num_queries``, and ``group_detr`` attributes.
        trust: Forwarded to :func:`rfdetr.utilities.io._safe_torch_load` as its ``trust``
            argument. Set ``True`` only when ``model_config.pretrain_weights`` points to a
            checkpoint the caller explicitly trusts (mirrors ``RFDETR.from_checkpoint(...,
            trust_checkpoint=True)``); the safe-load default otherwise applies here too.

    Returns:
        List of class name strings from the checkpoint, or an empty list if none are present or if
        ``model_config.pretrain_weights`` is ``None``.

    Raises:
        Exception: If the checkpoint file cannot be loaded even after a re-download.
    Nr   )Ú_safe_torch_loadz<Pretrain weights not found after initial download; retrying.T)Ú
redownloadF)Ústrictrˆ   z/Failed to load pretrain weights, re-downloadingÚmodelÚ
state_dictz=Normalizing PTL .ckpt checkpoint format (state_dict -> model)zmodel.z
_orig_mod.zThe checkpoint at zÚ appears to be in PyTorch Lightning format ('state_dict' key present, 'model' key absent), but 'state_dict' contains no keys with the expected 'model.' prefix. The checkpoint may be corrupt or in an unsupported format.ÚargsÚhyper_parametersÚclass_namesc                ó<   — g | ]}t          |t          ¦  «        ¯|‘ŒS r'   )Ú
isinstancer   )r(   r   s     r   r+   z)load_pretrain_weights.<locals>.<listcomp>¨  s(   € Ð"VÐ"VÐ"V¨DÅ
È4ÕQTÑ@UÔ@UÐ"V 4Ð"VÐ"VÐ"Vr   Únum_classesÚmodel_fields_setzclass_embed.biasrj   Únum_queriesÚ
group_detrz}load_pretrain_weights: checkpoint args.num_queries / args.group_detr not coercible to int; falling back to legacy flat slice.c              3  ó8   K  — | ]}t          |¦  «        ¯|V — Œd S )N)r   rQ   s     r   rB   z(load_pretrain_weights.<locals>.<genexpr>Ü  s/   è è € ÐBÐB�1­¸qÑ/AÔ/AÐBˆQÐBÐBÐBÐBÐBÐBr   u]   load_pretrain_weights: args.%s absent; inferred ckpt_%s=%d from tensor rows %d Ã· ckpt_%s=%d.c              3  ó^   •K  — | ]'\  }}t          |¦  «        ¯|j        d          ‰k    V — Œ(dS ©r   N)r   r/   )r(   r   r   Útarget_query_rowss      €r   rB   z(load_pretrain_weights.<locals>.<genexpr>ÿ  sV   øè è € ð 
ð 
Ù4@°D¸&ÕcrÐswÑcxÔcxð
ØŒL˜ŒOÐ/Ò/ð
ð 
ð 
ð 
ð 
ð 
r   zÙload_pretrain_weights: checkpoint lacks args.num_queries / args.group_detr; falling back to flat slice. With group_detr=%d this may scramble per-group query structure if the checkpoint was trained with group_detr > 1.)r   r!   r"   r#   Únum_keypoints_per_classÚuse_grouppose_keypointsrS   rm   c                ó,   — g | ]}t          |¦  «        ‘ŒS r'   )r    ©r(   Úns     r   r+   z)load_pretrain_weights.<locals>.<listcomp>/  s   € ÐRÐRÐR¨!�s 1™vœvÐRÐRÐRr   r,   c              3  ó"   K  — | ]
}|d k    V — ŒdS r�   r'   r¢   s     r   rB   z(load_pretrain_weights.<locals>.<genexpr>1  s&   è è € Ð6Ð6 �q˜1’uÐ6Ð6Ð6Ð6Ð6Ð6r   u›   load_pretrain_weights: _kp_active_mask in checkpoint has no active slots (schema=%s) â€” skipping auto-align to avoid overwriting config with empty schema.u�   load_pretrain_weights: auto-aligning num_keypoints_per_class %s â†’ %s (inferred from checkpoint _kp_active_mask; user did not set explicitly).u›   load_pretrain_weights: _kp_active_mask has unexpected shape %s (expected 2-D) â€” skipping auto-align; schema mismatch may cause APâ‰ˆ0 on keypoint models.Ú+get_num_keypoints_per_class_from_checkpointÚget_num_keypoints_per_classÚreinitialize_keypoint_headz\load_pretrain_weights: temporarily resizing keypoint schema from %s to checkpoint schema %s.z‘load_pretrain_weights: dropping checkpoint _kp_active_mask with shape %s because current model expects %s (derived from current keypoint schema).zUload_pretrain_weights: restoring configured keypoint schema %s after checkpoint load.)8Úpretrain_weightsr   Úrfdetr.utilities.ior�   r   ÚosÚpathÚisfiler0   r1   r   Ú	ExceptionÚinforY   Úitemsr@   r\   r.   r   r–   ÚiterrX   r   rW   Úsetr—   r/   Úreinitialize_detection_headÚgetr    Únextr™   rš   rC   Úlistr:   r   r9   r   Úhasattrr   ÚndimÚsumÚtolistrŸ   r   r¥   Úcallabler§   r‘   ÚdictÚpopr   Úpositional_encoding_sizeÚload_state_dictrc   )/rŠ   r‹   r‰   Úmcr¨   r”   r�   Ú
checkpointÚprefixÚcompile_prefixÚmodel_staterG   ÚvÚstrippedÚraw_class_namesÚiteratorÚuser_overroder—   Úcheckpoint_num_classesÚconfigured_num_classes_plus_bgÚ	ckpt_argsÚckpt_num_queries_rawÚckpt_group_detr_rawr   r!   Ú_first_query_keyÚ_nÚ_absentÚ	_inferredÚ_knownÚ
_known_valr   r   Ú_user_overrode_kp_schemaÚ_early_kp_maskÚ_ckpt_kp_schemaÚ_cfg_kp_schemaÚconfigured_keypoint_schemaÚcheckpoint_keypoint_schemaÚ%should_restore_config_keypoint_schemaÚget_model_schemaÚmodel_keypoint_schemaÚckpt_kp_active_maskÚmodel_state_dictÚmodel_kp_active_maskrI   rž   s/                                                 @r   r   r   6  s‰  ø€ ðP 
€BØ	ÔÐ"Øˆ	õ ˜2Ô.Ñ/Ô/ÐØ€Kà4Ð4Ð4Ð4Ð4Ð4õ Ð.Ñ/Ô/Ð/õ Œ7�>Š>Ð*Ñ+Ô+ð EÝ�ŠÐUÑVÔVÐVÝ!Ð"2¸tÐDÑDÔDÐDÝÐ.°uÐ=Ñ=Ô=Ð=ðEØ%Ð%Ð&6¸eÐDÑDÔDˆ
ˆ
øÝð Eð Eð EÝ�ŠÐEÑFÔFÐFÝ!Ð"2¸tÐDÑDÔDÐDØ%Ð%Ð&6¸eÐDÑDÔDˆ
ˆ
ˆ
ðEøøøð �jÐ Ð  \°ZÐ%?Ð%?Ý�ŠÐTÑUÔUÐUØˆð &ˆØˆØ˜|Ô,×2Ò2Ñ4Ô4ð 	*ð 	*‰DˆAˆqØ�|Š|˜FÑ#Ô#ð *Ø�S ™[œ[˜]˜]Ô+�Ø×&Ò& ~Ñ6Ô6ð ?Ø'­¨NÑ(;Ô(;Ð(=Ð(=Ô>�HØ()�˜HÑ%øØð 	ÝðMÐ%5ð Mð Mð Mñô ð ð *ˆ
�7Ñð ˜Ð#Ð#Ð(:¸jÐ(HÐ(HØ!+Ð,>Ô!?ˆJ�vÑð �ÐÐÝ(¨°FÔ);¸]ÑKÔKˆØð 	Wõ ˜/­3Ñ/Ô/ð 	WØ.Ð/��ðWÝ# OÑ4Ô4�Hð
 #WÐ"V°HÐ"VÑ"VÔ"V�K�Køõ	 !ð %ð %ð %à"$�K�K�Kð%øøøõ & j°"Ñ5Ô5Ð5ð "¥W¨RÐ1CÅSÁUÄUÑ%KÔ%KÐK€MØ”.€Kà'¨Ô0Ð1CÔDÔJÈ1ÔMÐØ%0°1¡_Ð"ØÐ!?Ò?Ð?à!Ð$BÒBÐBà ð -ð 5°qÑ8�Ø1GÐ.Ø!,�”ð 	×,Ò,Ð-CÑDÔDÐDð
 —’˜vÑ&Ô&€IØGPÐG\�>¨)°]ÑCÔCÐCÐbfÐØENÐEZ�.¨°LÑAÔAÐAÐ`dÐð	Ø8LÐ8X�3Ð3Ñ4Ô4Ð4Ð^bÐØ6IÐ6U�#Ð1Ñ2Ô2Ð2Ð[_ˆˆøÝ•zÐ"ð ð ð Ý�Šð9ñ	
ô 	
ð 	
ð  ÐØˆˆˆðøøøð 	˜DÐ  o¸Ð&=Ò>Ð>ÝØBÐB˜
 7Ô+ÐBÑBÔBØñ
ô 
Ðð Ð'Ø˜GÔ$Ð%5Ô6Ô<¸QÔ?ˆBØ"&ˆGØÐ+Ð0@À1Ò0DÐ0DÈÐN^ÑI^ÐbcÒIcÐIcØ"$Ð(8Ñ"8�Ø9EÀÐXeÐgwÐ9wÑ6�˜ F¨J¨JØ Ð,°À1Ò1DÐ1DÈÈoÑI]ÐabÒIbÐIbØ#%¨Ñ#8Ð Ø9FÐHXÐZfÐhwÐ9wÑ6�˜ F¨JØÐ"Ý—’ØsØØØØØØñô ð ð œ¨¬Ñ6Ðð 	Œ˜ÒÐØÐ%¨Ð)@Ýð 
ð 
ð 
ð 
ØDNÈwÔDW×D]ÒD]ÑD_ÔD_ð
ñ 
ô 
ñ 
ô 
ð *Aõ
 	�Šð>ð ŒMñ	
ô 	
ð 	
õ �Z Ô(×-Ò-Ñ/Ô/Ñ0Ô0ð Gð GˆÝ˜4Ñ Ô ð 	GØ Ô(¨Ô.ˆFØÐ+°Ð0KÝ,HØØ%5Ø$3Ø')¤~Ø&(¤mð-ñ -ô -�
˜7Ô# DÑ)Ð)ð -3Ð3EÐ4EÐ3EÔ,F�
˜7Ô# DÑ)øå7¸
À7Ô8KÈXÑVÔV€JˆwÑð  9½GÀBÐHZÕ\_Ñ\aÔ\aÑ<bÔ<bÐbÐà$ñå�BÐ1°5Ñ9Ô9ñõ �BÐ1Ñ2Ô2ñð
 $ GÔ,×0Ò0Ð1BÑCÔCˆÝ�n¥fÑ-Ô-ð 	°.Ô2EÈÒ2JÐ2JØRÐR¨~×/AÒ/AÀaÐ/AÑ/HÔ/H×/OÒ/OÑ/QÔ/QÐRÑRÔRˆOÝ!¥'¨"Ð.GÈÑ"LÔ"LÐ"RÐPRÑSÔSˆNÝÐ6Ð6 oÐ6Ñ6Ô6Ñ6Ô6ð =Ý—’ðià#ñô ð ð ð
 ! NÒ2Ð2Ý—’ð_à"Ø#ñ	ô ð ð .=�Ô*øÝ˜­Ñ/Ô/ð 	Ý�NŠNð`å�nÔ*Ñ+Ô+ñô ð õ "&¥g¨bÐ2KÈRÑ&PÔ&PÐ&VÐTVÑ!WÔ!WÐØ!%ÐØ,1Ð)ÝˆrÐ,¨eÑ4Ô4ð q½ØÐ?ñ:ô :ð qð &.×%YÒ%YÐZdÐelÔZmÑ%nÔ%nÐ"Ø%ð 
	qÝ& xÐ1NÐPTÑUÔUÐÝ:BÐCSÑ:TÔ:TÐ$tÐ$4Ð$4Ñ$6Ô$6Ð$6ÐZtÐ!Ø)Ð-BÒBÐBÅwÈxÐYuÑGvÔGvÐBÝ—’ØrØ)Ø.ñô ð ð
 ×3Ò3Ð4NÑOÔOÐOØ8RÐVpÒ8pÐ5ð % WÔ-×1Ò1Ð2CÑDÔDÐÝ07¸À,Ñ0OÔ0OÐW�x×*Ò*Ñ,Ô,Ð,ÐUWÐÝFPÐQaÕcgÑFhÔFhÐrÐ+×/Ò/Ð0AÑBÔBÐBÐnrÐåÐ&­Ñ/Ô/ð9åÐ+­VÑ4Ô4ð9ð  Ô%Ð)=Ô)CÒCÐCØ5ð Dõ 	�ŠðWåÐ%Ô+Ñ,Ô,ÝÐ&Ô,Ñ-Ô-ñ		
ô 	
ð 	
ð 	�7Ô×ÒÐ 1°4Ñ8Ô8Ð8Ý# J¨wÔ$7¸Ô9TÑUÔUÐUØ×+Ò+¨J°wÔ,?ÈÐ+ÑNÔN€LÝ˜,Ð(8Ñ9Ô9Ð9à,ð Hµ¸ÐC_Ñ1`Ô1`ð HÝ�ŠØcØ&ñ	
ô 	
ð 	
ð 	×+Ò+Ð,FÑGÔGÐGð Ð >Ò>Ð>À=Ð>Ø×,Ò,Ð-KÑLÔLÐLð �Q�Ð/Ò/Ð/Ø×,Ò,¨[¸1©_Ñ=Ô=Ð=àÐs7   ÂB" Â"AC'Ã&C'Ç2H ÈHÈHË#&L
 Ì
/L<Ì;L<c                ó>  — 	 ddl m}m} ddlm} n"# t
          $ r}t          d¦  «        |‚d}~ww xY w |dddg d¢¬	¦  «        }t          t          | j        d         ¦  «        } |t          ||j	        ¦  «        |¦  «        }t          t          |¦  «        |_	        dS )
aý  Apply LoRA adapters to the backbone encoder of *nn_model*.

    Replaces ``nn_model.backbone[0].encoder`` in-place with a PEFT-wrapped encoder using DoRA with rank 16 and alpha 16.

    Args:
        nn_model: LWDETR model whose backbone encoder will receive LoRA adapters.

    Raises:
        ImportError: If ``peft`` is not installed.
            Install via the RF-DETR extras, for example::

                pip install "rfdetr[lora]"
                # or
                pip install "rfdetr[train]"
    r   )Ú
LoraConfigÚget_peft_model)ÚPreTrainedModelz„LoRA requires the 'peft' dependency. Install it via RF-DETR extras, e.g.: pip install "rfdetr[lora]" or pip install "rfdetr[train]".Né   T)	Úq_projÚv_projÚk_projÚqkvÚqueryr=   ÚvalueÚ	cls_tokenÚregister_tokens)ÚrÚ
lora_alphaÚuse_doraÚtarget_modules)Úpeftrá   râ   Útransformersrã   ÚImportErrorr   r
   ÚbackboneÚencoderr   )rŠ   rá   râ   rã   ÚexcÚlora_configrô   Úpeft_encoders           r   r   r   ‡  sñ   € ð Ø3Ð3Ð3Ð3Ð3Ð3Ð3Ð3Ø0Ð0Ð0Ð0Ð0Ð0Ð0øÝð ð ð ÝðIñ
ô 
ð ð		øøøøðøøøð �*Ø
ØØð

ð 

ð 

ð	ñ ô €Kõ  •H˜hÔ/°Ô2Ñ3Ô3€Hð "�>¥$ ¸Ô8HÑ"IÔ"IÈ;ÑWÔW€LÝ�F LÑ1Ô1€HÔÐÐs   ‚ ‘
0›+«0)r   r   r   r   )r   r   r   r    r!   r    r"   r    r#   r    r   r   )r:   r;   r   r;   )rI   r   rJ   r   r   rK   )rd   re   rf   r    r   rK   )rŠ   r   r‹   r	   r‰   r   r   r;   )rŠ   r   r   rK   )+Ú__doc__Ú
__future__r   rt   rª   Útypingr   r   r4   Útorch.nn.functionalÚnnÚ
functionalrx   r   Úrfdetr.assets.model_weightsr   r   Úrfdetr.configr	   Úrfdetr.models.backbone.backboner
   Úrfdetr.models.backbone.dinov2r   Úrfdetr.models.lwdetrr   Úrfdetr.utilities.loggerr   Úrfdetr.utilities.state_dictr   r   r   r0   Ú__all__ri   r   Ú__annotations__r   r9   rH   rc   r   r   r   r'   r   r   ú<module>r     s   ððð ð ð #Ð "Ð "Ð "Ð "Ð "à €€€Ø 	€	€	€	Ø Ð Ð Ð Ð Ð Ð Ð à €€€Ø Ð Ð Ð Ð Ð Ð Ð Ð Ø Ð Ð Ð Ð Ð à \Ð \Ð \Ð \Ð \Ð \Ð \Ð \Ø %Ð %Ð %Ð %Ð %Ð %Ø 4Ð 4Ð 4Ð 4Ð 4Ð 4Ø 0Ð 0Ð 0Ð 0Ð 0Ð 0Ø 'Ð 'Ð 'Ð 'Ð 'Ð 'Ø .Ð .Ð .Ð .Ð .Ð .Ø xÐ xÐ xÐ xÐ xÐ xÐ xÐ xÐ xÐ xà	ˆ‰Œ€à
TÐ
TÐ
T€à1€ð *XÐ Ð WÐ WÐ WÑ Wð0ð 0ð 0ð 0ð*J$ð J$ð J$ð J$ðZ7ð 7ð 7ð 7ð@Ið Ið Ið IðX8
ð 8
ð 8
ð 8
ð~ ð	Nð Nð Nð Nð Nð Nðb
/2ð /2ð /2ð /2ð /2ð /2r   