§
    ­¸j¡2  ã            	       ó*  — d Z ddlmZmZ ddlmZ ddlmZ ddlm	Z	  e	¦   «         Z
d"ded	ed
edefd„Zd#dededefd„Zdeeef         deeeef         df         fd„Zdeeeef                  deee                  fd„Zdedej        deeeef                  fd„Zdedej        deeeef                  fd„Zdeeef         deeeef                  deeef         fd„Zdeeef         ddfd„Zdeeef         ded eee                  ddfd!„ZdS )$zFunctions to get params dict.é    )ÚAnyÚcast)Únn)ÚJoiner)Ú
get_loggerç      ð?é   ÚnameÚlr_decay_rateÚ
num_layersÚreturnc                 óP  — |dz   }|                       d¦  «        rYd| v sd| v rd}nNd| v rJd| vrFt          | |                      d¦  «        d…                              d	¦  «        d
         ¦  «        dz   }t                               d| › d||dz   |z
  z  › �¦  «         ||dz   |z
  z  S )zñCalculate lr decay rate for different ViT blocks.

    Args:
        name: parameter name.
        lr_decay_rate: base lr decay rate.
        num_layers: number of ViT blocks.

    Returns:
        lr decay rate for the given parameter.
    é   Úbackbonez
.pos_embedz.patch_embedr   z.blocks.z
.residual.Nú.é   úname: z, lr_decay: )Ú
startswithÚintÚfindÚsplitÚloggerÚdebug)r
   r   r   Úlayer_ids       úl/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/rfdetr/training/param_groups.pyÚget_vit_lr_decay_rater      sÔ   € ð ˜A‰~€HØ‡‚�zÑ"Ô"ð LØ˜4ÐÐ >°TÐ#9Ð#9ØˆHˆHØ˜4ÐÐ L¸Ð$<Ð$<Ý˜4 §	¢	¨*Ñ 5Ô 5Ð 7Ð 7Ô8×>Ò>¸sÑCÔCÀAÔFÑGÔGÈ!ÑKˆHÝ
‡L‚LÐZ˜$ÐZÐZ¨M¸jÈ1¹nÈxÑ>WÑ,XÐZÐZÑ[Ô[Ð[Ø˜Z¨!™^¨hÑ6Ñ7Ð7ó    Úweight_decay_ratec                 ór   — d| v sd| v sd| v sd| v sd| v rd}t                                d| › d|› �¦  «         |S )	zÛCalculate weight decay rate for different ViT parameters.

    Args:
        name: parameter name.
        weight_decay_rate: base weight decay rate.

    Returns:
        weight decay rate for the given parameter.
    ÚgammaÚ	pos_embedÚrel_posÚbiasÚnormg        r   z, weight_decay rate: )r   r   )r
   r   s     r   Úget_vit_weight_decay_rater%   -   sf   € ð 	�4ˆˆ˜[¨DÐ0Ð0°iÀ4Ð6GÐ6GÈVÐW[È^È^ÐagÐkoÐaoÐaoØÐÝ
‡L‚LÐH˜$ÐHÐHÐ5FÐHÐHÑIÔIÐIØÐr   Úparam_group.c                 ór   — t          t          d„ |                      ¦   «         D ¦   «         ¦  «        ¦  «        S )al  Return the hyperparameter overrides of ``param_group``, without its parameters.

    Values are compared by ``repr`` so a group carrying an unhashable setting (a third-party
    optimizer's list-valued option, say) still yields a key. Distinct floats keep distinct ``repr``,
    and anything without a value-based ``repr`` lands in its own bucket, which under-merges rather
    than merging two differently configured parameters.

    Args:
        param_group: A single optimizer parameter group.

    Returns:
        The group's non-``params`` items, sorted by key so groups configured identically compare equal.
    c              3   óJ   K  — | ]\  }}|d k    ¯|t          |¦  «        fV — ŒdS )ÚparamsN)Úrepr©Ú.0ÚkeyÚvalues      r   ú	<genexpr>z&_hyperparameter_key.<locals>.<genexpr>K   s=   è è € ÐdÐd©z¨s°EÐTWÐ[cÒTcÐTc˜�d 5™kœkÐ*ÐTcÐTcÐTcÐTcÐdÐdr   )ÚtupleÚsortedÚitems)r&   s    r   Ú_hyperparameter_keyr3   =   s5   € õ •ÐdÐd¸[×=NÒ=NÑ=PÔ=PÐdÑdÔdÑdÔdÑeÔeÐer   Úparam_dictsc                 óæ   — i }g }t          | ¦  «        D ]\\  }}t          |¦  «        }|                     |¦  «        }|€g x}||<   |                     |¦  «         |                     |¦  «         Œ]|S )a  Bucket ``param_dicts`` indices by hyperparameter overrides, in first-appearance order.

    Args:
        param_dicts: Single-parameter groups as built by :func:`_build_param_dicts`.

    Returns:
        One list of ``param_dicts`` indices per distinct hyperparameter combination.
    )Ú	enumerater3   ÚgetÚappend)r4   ÚbucketsÚorderedÚindexr&   r-   Úbuckets          r   Ú_merge_bucketsr=   N   s‡   € ð =?€GØ!€GÝ'¨Ñ4Ô4ð ð Ñˆˆ{Ý! +Ñ.Ô.ˆØ—’˜SÑ!Ô!ˆØˆ>Ø$&Ð&ˆF�W˜S‘\Ø�NŠN˜6Ñ"Ô"Ð"Ø�Š�eÑÔÐÐØ€Nr   ÚargsÚmodel_without_ddpc                 ó²  ‡ ‡	‡
— t          |j        t          ¦  «        sJ ‚t          d|j        d         ¦  «        }|                     ‰ d¬¦  «        Š	d„ ‰	                     ¦   «         D ¦   «         }dŠ
ˆ
fd„|                     ¦   «         D ¦   «         }ˆ fd„|D ¦   «         }ˆ	ˆ
fd	„|                     ¦   «         D ¦   «         }ˆ fd
„|D ¦   «         }||z   |z   }|S )a  Build one single-parameter group per trainable parameter, with its LR/weight-decay overrides.

    Args:
        args: Namespace supplying the learning-rate and weight-decay knobs.
        model_without_ddp: The model whose parameters the optimizer will own.

    Returns:
        One group per trainable parameter, ordered head/neck parameters first, then backbone, then
        decoder. :func:`get_param_dict` merges these; the order is also the parameter order of
        checkpoints written before that merge.
    r   r   z
backbone.0)Úprefixc                 ó   — g | ]\  }}|‘ŒS © rC   )r,   Ú_Ú
param_dicts      r   ú
<listcomp>z&_build_param_dicts.<locals>.<listcomp>r   s   € ÐeÐeÐe©m¨a°˜zÐeÐeÐer   ztransformer.decoderc                 ó0   •— g | ]\  }}‰|v ¯	|j         ¯|‘ŒS rC   ©Úrequires_grad)r,   ÚnÚpÚdecoder_keys      €r   rF   z&_build_param_dicts.<locals>.<listcomp>u   s1   ø€ ÐqÐqÐq™D˜A˜qÈKÐ[\ÐL\ÐL\ÐabÔapÐL\�aÐL\ÐL\ÐL\r   c                 ó4   •— g | ]}|‰j         ‰j        z  d œ‘ŒS ©)r)   Úlr)rO   Úlr_component_decay©r,   Úparamr>   s     €r   rF   z&_build_param_dicts.<locals>.<listcomp>w   s,   ø€ ÐuÐuÐuÐ]b¨°d´gÀÔ@WÑ6WÐXÐXÐuÐuÐur   c                 ó8   •— g | ]\  }}|‰v¯	‰|v¯|j         ¯|‘ŒS rC   rH   )r,   rJ   rK   Úbackbone_named_param_lr_pairsrL   s      €€r   rF   z&_build_param_dicts.<locals>.<listcomp>y   sF   ø€ ð ð ð áˆAˆqØÐ2Ð2Ð2°{È!Ð7KÐ7KÐPQÔP_Ð7Kð 	
à7KÐ7KÐ7Kr   c                 ó$   •— g | ]}|‰j         d œ‘ŒS rN   )rO   rQ   s     €r   rF   z&_build_param_dicts.<locals>.<listcomp>~   s#   ø€ ÐTÐTÐT¸e E°´Ð9Ð9ÐTÐTÐTr   )Ú
isinstancer   r   r   Úget_named_param_lr_pairsr2   Únamed_parameters)r>   r?   r   Úbackbone_param_lr_pairsÚdecoder_paramsÚdecoder_param_lr_pairsÚother_paramsÚother_param_dictsÚfinal_param_dictsrT   rL   s   `        @@r   Ú_build_param_dictsr_   c   s(  øøø€ õ Ð'Ô0µ&Ñ9Ô9Ð9Ð9Ð9Ý�EÐ,Ô5°aÔ8Ñ9Ô9€HØ$,×$EÒ$EÀdÐS_Ð$EÑ$`Ô$`Ð!ØeÐeÐ?\×?bÒ?bÑ?dÔ?dÐeÑeÔeÐà'€KØqÐqÐqÐqÐ$5×$FÒ$FÑ$HÔ$HÐqÑqÔq€NàuÐuÐuÐuÐftÐuÑuÔuÐðð ð ð ð à%×6Ò6Ñ8Ô8ðñ ô €Lð
 UÐTÐTÐTÀ|ÐTÑTÔTÐà)Ð,CÑCÐF\Ñ\ÐàÐr   c                 óZ   ‡— t          | |¦  «        Šˆfd„t          ‰¦  «        D ¦   «         S )a]  Build optimizer parameter groups with layer-wise LR (and backbone weight-decay) overrides.

    Parameters that end up configured identically share one group: ``torch.optim``'s foreach and
    fused kernels batch a group's parameters into a single multi-tensor launch, so one group per
    parameter would run ~500 single-tensor launches (and ~500 Python iterations of the optimizer's
    per-group loop) per step instead of one launch per distinct configuration.

    Args:
        args: Namespace supplying ``lr``, ``lr_encoder``, ``lr_component_decay``,
            ``lr_vit_layer_decay``, ``weight_decay``, and ``out_feature_indexes``.
        model_without_ddp: The model whose parameters the optimizer will own.

    Returns:
        Optimizer parameter groups, each holding every trainable parameter that shares its
        hyperparameter overrides.
    c                 óˆ   •— g | ]>}i d „ ‰|d                                        ¦   «         D ¦   «         ¥dˆfd„|D ¦   «         i¥‘Œ?S )c                 ó&   — i | ]\  }}|d k    ¯||“ŒS ©r)   rC   r+   s      r   ú
<dictcomp>z-get_param_dict.<locals>.<listcomp>.<dictcomp>™   s$   € Ð^Ð^Ð^™j˜c 5ÈcÐU]ÊoÈoˆs�EÈoÈoÈor   r   r)   c                 ó,   •— g | ]}‰|         d          ‘ŒS rc   rC   )r,   r;   r4   s     €r   rF   z-get_param_dict.<locals>.<listcomp>.<listcomp>š   s#   ø€ ÐHÐHÐH¸�{ 5Ô)¨(Ô3ÐHÐHÐHr   )r2   )r,   r<   r4   s     €r   rF   z"get_param_dict.<locals>.<listcomp>—   sz   ø€ ð ð ð ð
 ð		
Ø^Ð^¨K¸¸q¼	Ô,B×,HÒ,HÑ,JÔ,JÐ^Ñ^Ô^ð	
àÐHÐHÐHÐHÀÐHÑHÔHð	
ð 	
ðð ð r   )r_   r=   )r>   r?   r4   s     @r   Úget_param_dictrf   …   sJ   ø€ õ" % TÐ+<Ñ=Ô=€Kðð ð ð õ
 % [Ñ1Ô1ðñ ô ð r   Úscheduler_kwargsÚunmerged_param_groupsc                 ó¬  ‡‡— |                       d¦  «        Št          ‰t          ¦  «        r t          ‰¦  «        t          |¦  «        k    r| S t	          |¦  «        }g }|D ]]}ˆfd„|D ¦   «         Št          ˆfd„‰dd…         D ¦   «         ¦  «        rt          d¦  «        ‚|                     ‰d         ¦  «         Œ^t          | ¦  «        }||d<   |S )aq  Collapse legacy ``LambdaLR`` callbacks onto the merged optimizer groups.

    Legacy RF-DETR versions created one optimizer group per parameter. A dotted
    ``LambdaLR`` configuration could therefore provide one ``lr_lambda`` callback
    per parameter, while the current optimizer merges parameters with identical
    hyperparameters before the scheduler constructor validates that list. This
    helper returns a copy with one callback per merged group when every callback
    in that group is the same object.

    Args:
        scheduler_kwargs: Explicit scheduler constructor keyword arguments.
        unmerged_param_groups: Legacy single-parameter groups in their original order.

    Returns:
        Scheduler keyword arguments compatible with the merged group layout.

    Raises:
        ValueError: If one merged group would need distinct ``lr_lambda`` callbacks.
    Ú	lr_lambdac                 ó    •— g | ]
}‰|         ‘ŒS rC   rC   )r,   r;   Ú
lr_lambdass     €r   rF   z5regroup_unmerged_scheduler_kwargs.<locals>.<listcomp>½   s   ø€ Ð;Ð;Ð;¨5�Z Ô&Ð;Ð;Ð;r   c              3   ó,   •K  — | ]}|‰d          uV — ŒdS )r   NrC   )r,   ÚcallbackÚ	callbackss     €r   r/   z4regroup_unmerged_scheduler_kwargs.<locals>.<genexpr>¾   s,   øè è € ÐJÐJ°ˆx˜y¨œ|Ð+ÐJÐJÐJÐJÐJÐJr   r   NzÃCannot merge optimizer groups with distinct LambdaLR callbacks. Use one shared callback for parameters that share optimizer hyperparameters, or keep those parameters in separate optimizer groups.r   )	r7   rV   ÚlistÚlenr=   ÚanyÚ
ValueErrorr8   Údict)rg   rh   r9   Úregrouped_lambdasr<   Úregrouped_kwargsro   rl   s         @@r   Ú!regroup_unmerged_scheduler_kwargsrw       s  øø€ ð, "×%Ò% kÑ2Ô2€JÝ�j¥$Ñ'Ô'ð  ­3¨z©?¬?½cÐBWÑ>XÔ>XÒ+XÐ+XØÐåÐ2Ñ3Ô3€GØ#%ÐØð /ð /ˆØ;Ð;Ð;Ð;°FÐ;Ñ;Ô;ˆ	ÝÐJÐJÐJÐJ¸IÀaÀbÀb¼MÐJÑJÔJÑJÔJð 	ÝðIñô ð ð
 	× Ò  ¨1¤Ñ.Ô.Ð.Ð.åÐ,Ñ-Ô-ÐØ$5Ð�[Ñ!ØÐr   Ú
checkpointNc                 ó�  — |                       d¦  «        pg }t          |                       d¦  «        pg ¦  «        D �]ˆ\  }}|                      d¦  «        pg }|rt          d„ |D ¦   «         ¦  «        rŒ9t          |¦  «        }|                      di ¦  «        }i }g }d}	|D ]Š}
d„ ||
d                                       ¦   «         D ¦   «         }g }|
D ]?}||         d         d         }||v r||         ||	<   |                     |	¦  «         |	d	z  }	Œ@||d<   |                     |¦  «         Œ‹||d<   ||d<   |t          |¦  «        k     r$t          ||         t          |¦  «        |¦  «         t          |¦  «        t          |¦  «        k     r6t           	                    d
t          |¦  «        t          |¦  «        ¦  «         �ŒŠdS )uû  Rewrite one-group-per-parameter optimizer/scheduler state onto the merged parameter groups.

    :func:`get_param_dict` used to emit one parameter group per parameter, so ``torch.optim`` numbered
    the saved optimizer state by each parameter's position in that layout, and every per-group
    scheduler list (``base_lrs``, ``_last_lr``, ``lr_lambdas``, ``min_lrs``) had one entry per
    parameter. Groups now hold every parameter sharing their hyperparameters â€” a group count
    ``Optimizer.load_state_dict`` rejects outright â€” so reindex the saved state rather than fail the
    resume.

    Each optimizer's scheduler state is collapsed alongside it, matched by position the way
    PyTorch Lightning stores the two lists.

    The merged layout is derived from the saved groups themselves: bucketing them by the
    hyperparameters they recorded, in the order they were saved, reproduces the buckets
    :func:`get_param_dict` builds for the same run, since those are the same parameters in the same
    order. State already saved in the merged layout has groups holding several parameters and is left
    untouched.

    Args:
        checkpoint: Checkpoint dict carrying ``optimizer_states`` (and optionally ``lr_schedulers``),
            mutated in-place.
    Úlr_schedulersÚoptimizer_statesÚparam_groupsc              3   óH   K  — | ]}t          |d          ¦  «        dk    V — ŒdS )r)   r   N)rq   )r,   Úsaved_groups     r   r/   z3regroup_unmerged_optimizer_state.<locals>.<genexpr>å   s4   è è € Ð"cÐ"cÀ{¥3 {°8Ô'<Ñ#=Ô#=ÀÒ#BÐ"cÐ"cÐ"cÐ"cÐ"cÐ"cr   Ústater   c                 ó&   — i | ]\  }}|d k    ¯||“ŒS rc   rC   r+   s      r   rd   z4regroup_unmerged_optimizer_state.<locals>.<dictcomp>í   s*   € ÐlÐlÐl©:¨3°Ð\_ÐckÒ\kÐ\k˜C Ð\kÐ\kÐ\kr   r)   r   zXRegrouped resumed optimizer state from %d single-parameter groups onto %d merged groups.N)
r7   r6   rr   r=   r2   r8   rq   Ú_regroup_scheduler_listsr   Úinfo)rx   Úscheduler_statesr;   Úoptimizer_stateÚsaved_groupsr9   Úsaved_stateÚmerged_stateÚmerged_groupsÚslotr<   Úmerged_groupÚslotsÚunmerged_indexÚsaved_ids                  r   Ú regroup_unmerged_optimizer_staterŽ   Ë   s  € ð. "—~’~ oÑ6Ô6Ð<¸"ÐÝ"+¨J¯NªNÐ;MÑ,NÔ,NÐ,TÐRTÑ"UÔ"Uð ñ ÑˆˆØ&×*Ò*¨>Ñ:Ô:Ð@¸bˆØð 	�sÐ"cÐ"cÐVbÐ"cÑ"cÔ"cÑcÔcð 	ØÝ  Ñ.Ô.ˆØ%×)Ò)¨'°2Ñ6Ô6ˆØ')ˆØ.0ˆØˆØð 
	/ð 
	/ˆFØlÐl¸ÀfÈQÄiÔ9P×9VÒ9VÑ9XÔ9XÐlÑlÔlˆLØˆEØ"(ð ð �Ø'¨Ô7¸ÔAÀ!ÔD�Ø˜{Ð*Ð*Ø)4°XÔ)>�L Ñ&Ø—’˜TÑ"Ô"Ð"Ø˜‘	��Ø%*ˆL˜Ñ"Ø× Ò  Ñ.Ô.Ð.Ð.Ø#/ˆ˜Ñ Ø*7ˆ˜Ñ'Ø•3Ð'Ñ(Ô(Ò(Ð(Ý$Ð%5°eÔ%<½cÀ,Ñ>OÔ>OÐQXÑYÔYÐYÝˆ}ÑÔ¥ LÑ 1Ô 1Ò1Ð1Ý�KŠKØjÝ�LÑ!Ô!Ý�MÑ"Ô"ñô ð ùð3ð r   Úscheduler_stateÚunmerged_countr9   c                 ó<  ‡— |                       ¦   «         D ]…\  }Š|dk    rAt          ‰t          ¦  «        r,‰D ](}t          |t          ¦  «        rt	          |||¦  «         Œ)ŒLt          ‰t          ¦  «        r$t          ‰¦  «        |k    rˆfd„|D ¦   «         | |<   Œ†dS )au  Collapse a scheduler's per-parameter-group lists the way its optimizer's groups were collapsed.

    A composite scheduler (``SequentialLR``, ``ChainedScheduler``) nests its wrapped schedulers' own
    state dicts under a ``_schedulers`` list rather than holding per-group lists at the top level, so
    those nested dicts need the same collapse applied recursively.

    Args:
        scheduler_state: Saved scheduler state, mutated in-place.
        unmerged_count: Number of parameter groups the scheduler state was saved with.
        buckets: Saved-group indices per merged group, as produced by :func:`_merge_buckets`.
    Ú_schedulersc                 ó,   •— g | ]}‰|d                   ‘ŒS )r   rC   )r,   r<   r.   s     €r   rF   z,_regroup_scheduler_lists.<locals>.<listcomp>  s"   ø€ Ð#KÐ#KÐ#K¸ E¨&°¬)Ô$4Ð#KÐ#KÐ#Kr   N)r2   rV   rp   rt   r�   rq   )r�   r�   r9   r-   Únested_stater.   s        @r   r�   r�     sÍ   ø€ ð &×+Ò+Ñ-Ô-ð Lð L‰
ˆˆUØ�-ÒÐ¥J¨uµdÑ$;Ô$;ÐØ %ð Tð T�Ý˜l­DÑ1Ô1ð TÝ,¨\¸>È7ÑSÔSÐSøðTõ ˜�tÑ$Ô$ð 	L­¨U©¬°~Ò)EÐ)EØ#KÐ#KÐ#KÐ#KÀ7Ð#KÑ#KÔ#KˆO˜CÑ øðLð Lr   )r   r	   )r   )Ú__doc__Útypingr   r   Útorchr   Úrfdetr.models.backboner   Úrfdetr.utilities.loggerr   r   ÚstrÚfloatr   r   r%   rt   r0   r3   rp   r=   ÚModuler_   rf   rw   rŽ   r�   rC   r   r   ú<module>r�      s„  ðð $Ð #à Ð Ð Ð Ð Ð Ð Ð à Ð Ð Ð Ð Ð à )Ð )Ð )Ð )Ð )Ð )Ø .Ð .Ð .Ð .Ð .Ð .à	ˆ‰Œ€ð8ð 8 ð 8°Eð 8ÈSð 8ÐZ_ð 8ð 8ð 8ð 8ð0ð  Cð ¸Eð ÈEð ð ð ð ð f T¨#¨s¨(¤^ð f¸¸eÀCÈÀH¼oÈsÐ>RÔ8Sð fð fð fð fð"  T¨#¨s¨(¤^Ô 4ð ¸¸dÀ3¼i¼ð ð ð ð ð*˜Sð °R´Yð À4ÈÈSÐRUÈXÌÔCWð ð ð ð ðD˜ð °´ð ¸tÀDÈÈcÈÄNÔ?Sð ð ð ð ð6(Ø˜3 ˜8”nð(Ø=AÀ$ÀsÈCÀxÄ.Ô=Qð(à	ˆ#ˆsˆ(„^ð(ð (ð (ð (ðV5°°c¸3°h´ð 5ÀDð 5ð 5ð 5ð 5ðpL¨d°3¸°8¬nð LÈcð LÐ\`ÐaeÐfiÔajÔ\kð LÐptð Lð Lð Lð Lð Lð Lr   