§
    w²ºjóR  ã            	      óÖ   — d Z ddlmZ ddlZddlZddlZddlZddlm	Z	 ddl
mZ ddlZdddœZd	Zd
„ Zd„ Zd„ Zdddœd„Zd%ddœd„Zddœd„Zdddddœd„Zddddddd dd!œd"„Zdd#œd$„ZdS )&zÛAuditable locked recipes, official optimizer preparation and integration status.

Constrained YOLO and RF-DETR run through sensor research integration. Official
defaults remain gated despite runnable optimizer helpers.
é    )ÚannotationsN)ÚPath)ÚSimpleNamespacez8.4.163z1.11.0)ÚultralyticsÚrfdetr)Úofficial_defaultÚofficial_research_constrainedÚproject_adamw_pilotc                ó’  — t           j                             | ¦  «        }|t          |          k    r"t	          | › dt          |          › d|› �¦  «        ‚t          t           j                             | ¦  «        j        ¦  «        j	        |z  }| ||t          j        |                     ¦   «         ¦  «                             ¦   «         dœS )Nz==z required, found )ÚpackageÚversionÚfileÚsha256)Ú	importlibÚmetadatar   ÚLOCKED_VERSIONSÚRuntimeErrorr   ÚutilÚ	find_specÚoriginÚparentÚhashlibr   Ú
read_bytesÚ	hexdigest)r   Úrelativer   Úpaths       úG/root/private/project/pallet_detection_v2/pallet/experiments/recipes.pyÚ_sourcer      s¶   € ÝÔ ×(Ò(¨Ñ1Ô1€GØ•/ 'Ô*Ò*Ð*Ý˜gÐ]Ð]­¸Ô)AÐ]Ð]ÐT[Ð]Ð]Ñ^Ô^Ð^Ý•	”×(Ò(¨Ñ1Ô1Ô8Ñ9Ô9Ô@À8ÑK€DØ¨7¸HÝ”n T§_¢_Ñ%6Ô%6Ñ7Ô7×AÒAÑCÔCðEð Eð Eó    c                óŠ   — t          |t          ¦  «        st          |t          ¦  «        r|dk     rt          | › d�¦  «        ‚d S )Né   z must be a positive integer)Ú
isinstanceÚboolÚintÚ
ValueError)ÚnameÚvalues     r   Ú_positive_intr(      sM   € Ý�%�ÑÔð ?¥j°½Ñ&<Ô&<ð ?ÀÈÂ	À	Ý˜DÐ=Ð=Ð=Ñ>Ô>Ð>ð AJÀ	r   c                óz   — t          d| ¦  «         t          d|¦  «         || k     s|| z  rt          d¦  «        ‚d S )NÚ
batch_sizeÚeffective_batchz/effective_batch must be divisible by batch_size)r(   r%   )r*   r+   s     r   Ú_batchr,   #   sO   € Ý�, 
Ñ+Ô+Ð+ÝÐ# _Ñ5Ô5Ð5Ø˜Ò#Ð# ¸Ñ'CÐ#ÝÐJÑKÔKÐKð $Ð#r   é@   r!   )ÚnbsÚncc                óX  — t          ¦   «                              ¦   «                              ¦   «         D ]\  }}t          ||¦  «         Œt	          dd¦  «         t          j        | t          ||¦  «        z  ¦  «        |z  }|dk    rdndt          dd|z   z  d¦  «        f\  }}||d	|d
dœS )z¡Resolve the exact locked BaseTrainer auto branch, including >10000 boundary.

    iterations is upstream's selection estimate, not actual accumulated steps.
    r   úengine/trainer.pyi'  )ÚMuSGDç{®Gáz„?ÚAdamWr3   é   é   çÍÌÌÌÌÌì?ç        )r&   ÚlrÚmomentumÚ
iterationsÚwarmup_bias_lr)	ÚlocalsÚcopyÚitemsr(   r   ÚmathÚceilÚmaxÚround)	Únum_train_samplesÚepochsr*   r.   r/   r&   r'   r;   r9   s	            r   Úresolve_ultralytics_autorF   *   sÅ   € õ
 ‘x”x—}’}‘”×,Ò,Ñ.Ô.ð #ð #‰ˆˆeÝ�d˜EÑ"Ô"Ð"Ð"ÝˆMÐ.Ñ/Ô/Ð/Ý”Ð,­s°:¸sÑ/CÔ/CÑCÑDÔDÀvÑM€JØ",¨uÒ"4Ð"4ˆˆ¸7ÅEÈ)ÐWXÐ[]ÑW]ÑJ^Ð`aÑDbÔDbÐ:c�H€Dˆ"Ø °Ø$¸ð=ð =ð =r   r	   r5   )r*   c               óŽ  ‡— | dvrt          d| › �¦  «        ‚|t          vrt          d|› �¦  «        ‚t          d|¦  «         |dk    rt          |d¦  «         | dk    }|r¹dd	lm}m}  |d
¬¦  «                             d¬¦  «         |¦   «                              d¬¦  «        dœŠd„ dD ¦   «         }d‰d         d         ‰d         d         dddddgddddddœdd d!œ
}d"d#d$d%d&d#d'œ}d(d)d$d*d+d,d-œ}	d.d/d0d(d(d1d(d2d3d4œ	}
d5d6i|d|z  d7d(d8d9d:d:d;œd:dd7d8d6d<œd=œ}n»dd/l}d>„ d?D ¦   «         }t          t          j                             d@¦  «        j        ¦  «        j        }d|                     |dAz                       ¦   «         ¦  «        iŠdBdCdDd#dEdFg dG¢d#d#dHdHdIœdJdKdLdMœdNdOœ	}dPdQdRdSdTdUdVœ}d(dWdXd+dYdZœ}	ˆfd[„d\D ¦   «         }
d6|dd#d7d8d]œdd6d:d7d8d^œd_œ}g d`¢}|r4t          j                             da¦  «        €|                     db¦  «         d*| |dcdd|‰|dk    r|ni |||	|
|deœ}|dk    rög df¢|dg<   |sndh|dg         d<   d#|di         dj<   |                     dkg ¬l¦  «         g dm¢|dn<   do|dp         dq<   |di                              d#d#d#dr¬s¦  «         dt„ duD ¦   «         |dv<   n|di                              d9d:d:¬w¦  «         |                     dkg ¬l¦  «         dx|dg         d<   g dy¢|dn<   dz|dp         dq<   |di                              d{d:d|¬}¦  «         d~„ dD ¦   «         |dv<   |d€k    rœ|                     d�d‚d|rdndƒdEddgd„d…œd†d‡d%dQdˆœd‰d:idŠd:d‹œg d6dŒdd7d8d�œ¬Ž¦  «         d�t'          j        t          t*          ¦  «                             d�¦  «                             ¦   «         ¦  «                             ¦   «         d‘œ|d’<   |S )“a„  Return a JSON-safe specification; preserve full upstream default snapshots.

    RF-DETR defaults include both Small model config and TrainConfig. Dataset paths
    are placeholders. Research constraints use one device; global batch requires
    a separate explicit distributed design. Structural comparisons must use the
    same recipe, batch, preprocessing and selection protocol.
    >   úyolov8-cspnextÚyolo26Úyolov8úrfdetr-smallzunsupported family: zunknown recipe: r*   r	   é   rK   r   )ÚTrainConfigÚRFDETRSmallConfigz<dataset_dir>)Údataset_dirÚjson)Úmode)ÚtrainÚmodelc                ó.   — g | ]}t          d |¦  «        ‘ŒS )r   ©r   ©Ú.0Úps     r   ú
<listcomp>z#get_recipe_spec.<locals>.<listcomp>L   s.   € ð ;ð ;ð ;¨A•7˜8 QÑ'Ô'ð ;ð ;ð ;r   )z	config.pyútraining/param_groups.pyzmodels/backbone/backbone.pyztraining/module_model.pyztraining/callbacks/ema.pyztraining/trainer.pyzdatasets/coco.pyzdatasets/transforms.pyr4   rR   r9   Ú
lr_encoderçffffffæ?çš™™™™™é?ç-Cëâ6?r7   g+‡ÙÎ÷ï?zlr * lr_component_decayz0lr_encoder * lr_component_decay**2 * layer_decayzMzero for gamma/pos_embed/rel_pos/bias/norm/embeddings; otherwise weight_decayz5optimizer weight_decay, including head/decoder biases)ÚotherÚdecoderÚencoderÚencoder_decayÚother_decayzynum_layers=out_feature_indexes[-1]+1; exponent=num_layers+1-layer_id; embeddings=0; .layer.k.=k+1; otherwise num_layers+1zKmodel fused_optimizer=True, gated by upstream runtime/precision eligibility)
r&   r9   r[   Úlr_component_decayÚlr_vit_layer_decayÚweight_decayÚbetasÚgroupsÚlayer_decayÚfusedÚstepr8   éd   çš™™™™™¹?Úoptimizer_step)r&   Úwarmup_epochsÚlr_dropÚgammaÚintervalÚcosine_min_factor_if_selectedTg-²�ï§Æï?r!   zdecay * (1-exp(-updates/tau))zselected EMA by default)ÚenabledÚdecayÚtauÚupdate_intervalÚformulaÚ
evaluationÚcpuNz torchvision horizontal flip p=.5z	per-batchzRdirect random resize or resize[400,500,600] + random crop + selected target resizez(not present in upstream RF-DETR pipeline)	Úbackend_selectorÚ
aug_configÚdefault_optional_transformÚsquare_resize_div_64Úscale_jitterÚmulti_scaleÚexpanded_scalesÚresize_cropÚmosaicÚ
resolutioni€  i,  é2   ÚoffF)r*   Úgrad_accum_stepsrE   Úearly_stoppingÚearly_stopping_patiencer€   r�   r   )rƒ   r+   Ú
max_epochsÚpatienceÚsize)rS   rR   Úprotocolc                ó.   — g | ]}t          d |¦  «        ‘ŒS )r   rU   rV   s     r   rY   z#get_recipe_spec.<locals>.<listcomp>l   s4   € ð að að a°•7˜=¨!Ñ,Ô,ð að að ar   )úcfg/default.yamlr1   zutils/torch_utils.pyzdata/augment.pyr   r�   Úautoz)ceil(train_samples/max(batch,nbs))*epochsz]MuSGD lr=.01 momentum=.9 iff iterations>10000; else AdamW lr=round(.002*5/(4+nc),6), beta1=.9çü©ñÒMb@?z4weight_decay * batch * max(round(nbs/batch),1) / nbs)zMuSGD ndim in {2,4}zbias substringz$Norm module or logit_scale substringÚweightÚscaled_weight_decay)ÚpriorityÚ
bias_decayÚ
norm_decayrf   Ú
muon_decaygš™™™™™É?g      ð?z93x for head cv3/one2one_cv3, proto.semseg/SemanticSegment)ÚmuonÚsgdÚ
boosted_lrz Adam/AdamW only on eligible CUDA)	r&   Úselection_iterationsÚ	selectionÚwarmup_bias_lr_after_autorf   r“   rh   r2   rj   Úlinearr3   g      @z5round(min(warmup_epochs,max(epochs-1,0))*num_batches)zPgroup LR interpolation plus momentum/accumulation interpolation in upstream loopÚepoch)r&   Úlrfro   Úwarmup_iterationsÚwarmuprr   g§èH.ÿï?iÐ  ÚEMA)rt   ru   rv   rx   ry   c                ó.   •— i | ]}|‰d          |         “ŒS )rR   © )rW   ÚkÚdefaultss     €r   ú
<dictcomp>z#get_recipe_spec.<locals>.<dictcomp>€   s7   ø€ ð kð kð k°A˜˜8 GÔ,¨QÔ/ð kð kð kr   )Úhsv_hÚhsv_sÚhsv_vÚdegreesÚ	translateÚscaleÚshearÚperspectiveÚflipudÚfliplrrƒ   ÚmixupÚcutmixÚ
copy_pasteÚclose_mosaicr€   )ÚimgszÚbatchr.   rƒ   rE   r‹   )r+   rŒ   rƒ   rŠ   r‹   )rR   r�   )z<Project train.py does not consume this recipe specification.zjOfficial EMA, schedule/warmup, augmentation/preprocessing and trainer checkpoint state are not integrated.zuSensor adapters and project evaluation/selection must be explicitly integrated and compared under one fixed protocol.Úpytorch_lightningz…RF-DETR training/__init__.py requires pytorch_lightning, absent from this environment; official Lightning trainer cannot be imported.Únot_runnableÚrunnable)Úschema_versionÚfamilyÚrecipeÚtraining_statusÚoptimizer_preparation_statusÚsourcesÚofficial_defaultsÚresearch_overridesÚ	optimizerÚscheduleÚemaÚaugmentationÚblockers)zŒUltralytics warmup ramps accumulation from 1 to nbs/batch; exact effective batch 16 at every step requires an additional non-default policy.zgRF-DETR disables multi_scale/scale_jitter to keep training inputs at 640; those are research overrides.z\Research patience=50 does not equalize upstream min_delta or checkpoint selection semantics.Úconstraint_caveatszÍNominal batch16 uses fixed accumulation; partial epoch tails have fewer real examples with explicit16/actual loss rescaling. This is a research deviation from upstream's accumulation warmup/carry behavior.rÇ   rƒ   Ú$runnable_sensor_research_integration)r¿   rÈ   )z~Fixed nominal accumulation16, no accumulation warmup ramp; epoch tail raw loss is multiplied by16/actual_examples and flushed.zãFour float sensor channels with masked interpolation and zero missing-signal padding replace uint8 RGB/114 padding and HSV; synchronized scale=.5 and translate=.1 affine geometry plus fliplr=.5; no pilot modality perturbations.z˜Project original-coordinate AP50_95 selects EMA checkpoints; strict improvement and patience50 use project selection, not upstream fitness/tie behavior.zœSingle device, project float normalization and model adapters; no upstream OOM batch reduction, distributed trainer, RGB AMP eligibility probe or callbacks.zˆCheckpoints store FP32 EMA, full optimizer/scaler/RNG and YOLO26 criterion update count; EMA updates only on successful optimizer steps.Úsensor_research_deviationsÚfixed16_epoch_tail_rescalerÅ   ÚaccumulationÚsensor_float_spatial_v1)r©   rª   r«   Úimplementationc                óÔ   — g | ]e}d |› �t          j        t          t          ¦  «                             |¦  «                             ¦   «         ¦  «                             ¦   «         dœ‘ŒfS ©zpallet/experiments/©r   r   ©r   r   r   Ú__file__Ú	with_namer   r   ©rW   r&   s     r   rY   z#get_recipe_spec.<locals>.<listcomp>¤   s}   € ð 3Jð 3Jð 3Jð ð 6¨tÐ5Ð5Ý"œ>­$­x©.¬.×*BÒ*BÀ4Ñ*HÔ*H×*SÒ*SÑ*UÔ*UÑVÔV×`Ò`ÑbÔbðdð dð3Jð 3Jð 3Jr   )útrain.pyzyolo_recipe.pyú
dataset.pyÚproject_integration_sources)r€   r�   r   z´RF detection loss uses its native per-microbatch box normalization; the loop weights each microbatch mean by its real examples over the accumulation window and flushes epoch tails.)zÏFixed sensor letterbox with supported float/mask interpolation replaces RF square resize, random resize/crop and RGB normalization; native horizontal flip probability .5 is synchronized with boxes and masks.zÓScale jitter and expanded multi-scale are disabled for fixed640; one-device effective16 accumulation uses actual-example weighted epoch tails, a named deviation from Lightning's fixed divisor on partial windows.z«Project original-coordinate AP50_95 selects EMA strict improvement, unlike RF's separate regular/EMA tracks, smoothing/skip epochs, min_delta and checkpoint metric policy.z¸Official AveragedModel callback buffers/counter, step scheduler, grouping, fused eligibility, precision choice and clip norm .1 are used; model sensor adapter adds base-lr/decay group.zÕCUDA deformable-attention training is numerically nondeterministic across identical seeds in this environment; RNG/scheduler/optimizer/EMA resume state is preserved, but bit-exact GPU weights cannot be guaranteed.z•ONNX export supports checkpoint loading but RF near-tied query ordering in missing-mask probes can fail strict parity; report parity case explicitly.Ú&fixed16_actual_example_mean_epoch_tailÚrf_sensor_float_flip_v1g      à?)rÏ   Únative_resize_cropÚ$synchronized_sensor_flip_probabilityc                óÔ   — g | ]e}d |› �t          j        t          t          ¦  «                             |¦  «                             ¦   «         ¦  «                             ¦   «         dœ‘ŒfS rÑ   rÓ   rÖ   s     r   rY   z#get_recipe_spec.<locals>.<listcomp>·   s}   € ð 3Lð 3Lð 3Lð ð 6¨tÐ5Ð5Ý"œ>­$­x©.¬.×*BÒ*BÀ4Ñ*HÔ*H×*SÒ*SÑ*UÔ*UÑVÔV×`Ò`ÑbÔbðdð dð3Lð 3Lð 3Lr   )r×   zrfdetr_recipe.pyrØ   r
   Úrunnable_existing_train_pyÚexisting_train_pygü©ñÒMbP?zAall model parameters in one group; no encoder or bias/norm policy)r&   r9   rf   rg   rh   Ú$project_optimizer_step_warmup_cosineé   )r&   ro   Úwarmup_start_factorÚmin_lr_ratiort   zPalletDataset augmentation)rÏ   rƒ   r5   )rŒ   r*   r+   rE   r‹   )r¿   rÀ   rÄ   rÅ   rÆ   rÇ   rÈ   Úpilot_defaultszpallet/experiments/train.pyr×   rÒ   Úproject_source)r%   ÚRECIPE_NAMESr(   r,   Úrfdetr.configrM   rN   Ú
model_dumpÚyamlr   r   r   r   r   r   Ú	safe_loadÚ	read_textÚappendÚupdater   r   rÔ   rÕ   r   r   )r½   r¾   r*   Úis_rfrM   rN   rÁ   rÄ   rÅ   rÆ   rÇ   Ú	overridesrê   r   rÈ   Úspecr§   s                   @r   Úget_recipe_specrò   8   s
  ø€ ð ÐKÐKÐKÝÐ8°Ð8Ð8Ñ9Ô9Ð9Ø•\Ð!Ð!ÝÐ4¨FÐ4Ð4Ñ5Ô5Ð5Ý�, 
Ñ+Ô+Ð+ØÐ0Ò0Ð0Ýˆz˜2ÑÔÐØ�nÒ$€EØð >FØ@Ð@Ð@Ð@Ð@Ð@Ð@Ð@Ø(˜[°_ÐEÑEÔE×PÒPÐV\ÐPÑ]Ô]Ø.Ð.Ñ0Ô0×;Ò;ÀÐ;ÑHÔHðJð Jˆð;ð ;ð 2:ð ;ñ ;ô ;ˆð %¨H°WÔ,=¸dÔ,CØ#+¨GÔ#4°\Ô#BØ+.ÀcØ%)°S¸%°LØ)-Ð:SØ+]ð 2AØ/fð hð  hð %`Økð	mð 	mˆ	ð #°SÀSØ Ð.>Ðadðfð fˆà¨°sÈqØ9ÐIbðdð dˆà,1ÀØ6XØ04ÀdØ'2ÀtØ'{Ø"LðNð Nˆð  ,¨SÐ1Ø$¸"À
Ñ:JØ¨TÈbØ °UÈEð=Sð =Sð $)¸RÈsØ%'°ð6ð 6ð	7ð 7ˆ	ˆ	ð 	ˆˆˆðað að 7`ð añ aô aˆå•y”~×/Ò/°Ñ>Ô>ÔEÑFÔFÔMˆØ˜TŸ^š^¨WÐ7IÑ-I×,TÒ,TÑ,VÔ,VÑWÔWÐXˆØ#Ð=hð #BØ25ÀvØ,bð -Hð  -Hð  -HØ.1ÀØ0EÐUjð lð  lð (+°3Ø.iðkð kà@ð	Bð 	Bˆ	ð %¨TÀCØ)`ØpØ 'ð)ð )ˆð ¨¸Ø9ÈðPð Pˆðkð kð kð kð :jð kñ kô kˆð ),°jÈØ),¸ÈðMð Mà57ÀÐPUØ03Àð"Eð "EðFð Fˆ	ðIð Ið I€Hð ð a•”×)Ò)Ð*=Ñ>Ô>ÐFØ�Šð  `ñ  	aô  	að  	aØ¨6¸VØ-ÈzØ°XØ/5Ð9XÒ/XÐ/X ) )Ð^`Ø"°ÀØ(°hð@ð @€Dð Ð0Ò0Ð0ð&lð &lð &lˆÐ!Ñ"ð ð #	Lð -|ˆDÐ%Ô& qÑ)Ø-0ˆD�Ô  Ñ*Ø�KŠKÐ(NÐY[ˆKÑ\Ô\Ð\ð2\ð 2\ð 2\ˆDÐ-Ñ.ð 0LˆD�Ô˜^Ñ,Ø�Ô ×'Ò'¨b¸À"Ø8Qð (ñ Sô Sð Sð3Jð 3Jð Ið3Jñ 3Jô 3JˆDÐ.Ñ/Ð/ð
 �Ô ×'Ò'°EÈ5Ð_dÐ'ÑeÔeÐeØ�KŠKÐ(NÐY[ˆKÑ\Ô\Ð\ð -cˆDÐ%Ô& qÑ)ð2ið 2ið 2iˆDÐ-Ñ.ð 0XˆD�Ô˜^Ñ,Ø�Ô ×'Ò'Ð7PØ<AØNPð (ñ Rô Rð Rð3Lð 3Lð Kð3Lñ 3Lô 3LˆDÐ.Ñ/ð Ð&Ò&Ð&Ø�ŠÐ$@Ø1DØ'.ÀÐ6P°f°fÈ5Ø/5ÀÀeÀØ)lðnð nð 'MÐ_`Ø58È$ðPð Pà" EÐ*Ø4PÐ\aÐ!bÐ!bØ¸ÈAÐbdØ9<È"ð1Nð 1Nð 	ñ 
	Oô 
	Oð 
	Oð +HÝ”n¥T­(¡^¤^×%=Ò%=¸jÑ%IÔ%I×%TÒ%TÑ%VÔ%VÑWÔW×aÒaÑcÔcð"eð "eˆÐÑà€Kr   T©Útrainable_onlyc               óˆ  ‡‡
— ˆfd„|                       ¦   «         D ¦   «         }t          ¦   «         Š
g }|D ]´}g }|d         D ]p}t          |¦  «        }||vrt          d¦  «        ‚|‰
v rt          d||         › �¦  «        ‚‰
                     |¦  «         |                     ||         ¦  «         Œq|                     i d„ |                     ¦   «         D ¦   «         ¥d|i¥¦  «         Œµˆ
fd„|                     ¦   «         D ¦   «         }	|	r%t          dd	                     |	¦  «        › �¦  «        ‚|st          d
¦  «        ‚|S )zLReject missing, duplicated or foreign parameters; return named group report.c                óH   •— i | ]\  }}|j         s‰°t          |¦  «        |“ŒS r¥   ©Úrequires_gradÚid)rW   ÚnrX   rô   s      €r   r¨   z/validate_parameter_coverage.<locals>.<dictcomp>Î   s2   ø€ ÐgÐgÐg™T˜Q ÀÄÐgÐXfÐg•�1‘”�qÐgÐgÐgr   Úparamsz)optimizer has foreign or frozen parameterzduplicate optimizer parameter: c                ó&   — i | ]\  }}|d k    ¯||“ŒS ©rû   r¥   )rW   r¦   Úvs      r   r¨   z/validate_parameter_coverage.<locals>.<dictcomp>Û   s#   € ÐJÐJÐJ¡4 1 a¸AÀºM¸M˜!˜Q¸M¸M¸Mr   Úparameter_namesc                ó"   •— g | ]\  }}|‰v¯	|‘ŒS r¥   r¥   )rW   Úkeyr&   Úseens      €r   rY   z/validate_parameter_coverage.<locals>.<listcomp>Ü   s"   ø€ ÐIÐIÐI™	˜˜T¸ÀD¸¸ˆt¸¸¸r   zoptimizer missing parameters: z, z#model has no parameters to optimize)Únamed_parametersÚsetrù   r%   Úaddrí   r?   Újoin)rS   rh   rô   ÚexpectedÚreportÚgroupÚnamesrX   r  Úmissingr  s     `       @r   Úvalidate_parameter_coverager  Ì   s{  øø€ àgÐgÐgÐg U×%;Ò%;Ñ%=Ô%=ÐgÑgÔg€HÝ‰5Œ5€DØ€FØð 
gð 
gˆØˆØ�x”ð 	(ð 	(ˆAÝ�Q‘%”%ˆCØ˜(Ð"Ð"Ý Ð!LÑMÔMÐMØ�dˆ{ˆ{Ý Ð!RÀ8ÈCÄ=Ð!RÐ!RÑSÔSÐSØ�HŠH�S‰MŒMˆMØ�LŠL˜ #œÑ'Ô'Ð'Ð'Ø�ŠÐeÐJÐJ¨5¯;ª;©=¬=ÐJÑJÔJÐeÐL]Ð_dÐeÐeÑfÔfÐfÐfØIÐIÐIÐI X§^¢^Ñ%5Ô%5ÐIÑIÔI€GØð PÝÐN¸$¿)º)ÀGÑ:LÔ:LÐNÐNÑOÔOÐOØð @ÝÐ>Ñ?Ô?Ð?Ø€Mr   rL   r‘   )r*   r.   r/   rf   c               ó  — t          |||||¬¦  «        }t          j        |¦  «        r|dk     rt          d¦  «        ‚ddlm} t          t          ddd¬¦  «        d	|i¬
¦  «        }	||z  t          t          ||z  ¦  «        d¦  «        z  |z  }
 G d„ dt          j
        j        ¦  «        }t          | d¦  «        r || ¦  «        n| }|                     |	|d|
|d         ¬¦  «        }t          | |j        d¬¦  «         |S )zªUse upstream builder verbatim; model must support upstream MuSGD head access.

    This prepares an optimizer only. It does not implement upstream trainer semantics.
    )rD   rE   r*   r.   r/   r   z+weight_decay must be finite and nonnegative)ÚBaseTrainerr3   g–C‹lçûí?rm   )Úlr0r:   r<   r/   )ÚargsÚdatar!   c                  ó4   ‡ — e Zd Zˆ fd„Zed„ ¦   «         Zˆ xZS )ú2build_ultralytics_optimizer.<locals>.OptimizerViewc                óV   •— t          ¦   «                              ¦   «          || _        d S ©N)ÚsuperÚ__init__Údetector)Úselfr  Ú	__class__s     €r   r  z;build_ultralytics_optimizer.<locals>.OptimizerView.__init__õ   s$   ø€ Ý‰GŒG×ÒÑÔÐØ$ˆDŒMˆMˆMr   c                ó   — | j         j        S r  )r  Úyolo)r  s    r   Ústudent_modelz@build_ultralytics_optimizer.<locals>.OptimizerView.student_modelù   s   € à”=Ô%Ð%r   )Ú__name__Ú
__module__Ú__qualname__r  Úpropertyr  Ú__classcell__)r  s   @r   ÚOptimizerViewr  ô   sS   ø€ € € € € ð	%ð 	%ð 	%ð 	%ð 	%ð 
ð	&ð 	&ñ 
Œð	&ð 	&ð 	&ð 	&ð 	&r   r#  r  r�   r;   )r&   ru   r;   Fró   )rF   r@   Úisfiniter%   Úultralytics.engine.trainerr  r   rB   rC   ÚtorchÚnnÚModuleÚhasattrÚbuild_optimizerr  Úparam_groups)rS   rD   rE   r*   r.   r/   rf   Úresolvedr  Úcontextru   r#  ÚtargetrÄ   s                 r   Úbuild_ultralytics_optimizerr/  ä   si  € õ (Ð:KÐTZØ3=À3È2ðOñ Oô O€HåŒ=˜Ñ&Ô&ð H¨,¸Ò*:Ð*:ÝÐFÑGÔGÐGØ6Ð6Ð6Ð6Ð6Ð6Ý¥?°tÀeØBEð$Gñ $Gô $GØNRÐTVÈZðYñ Yô Y€Gà˜:Ñ%­­E°#¸
Ñ2BÑ,CÔ,CÀQÑ(GÔ(GÑGÈ#ÑM€Eð&ð &ð &ð &ð &�œœñ &ô &ð &õ &-¨U°FÑ%;Ô%;ÐFˆ]ˆ]˜5Ñ!Ô!Ð!À€FØ×+Ò+¨G°VÀ&ÐPUØ6>¸|Ô6Lð ,ñ Nô N€Iå  yÔ'=ÈeÐTÑTÔTÐTØÐr   Fr^   ga2U0*©#?r\   r]   )Ú
core_modelÚout_feature_indexesÚallow_adapter_parametersr9   r[   rd   re   rf   c          
     ót  ‡— t          dd¦  «         |||||dœ                     ¦   «         D ]G\  }	}
t          j        |
¦  «        r|
dk     s|	dk    r"|
dk    rt	          |	› d|	dk    rdnd› �¦  «        ‚ŒH|€"t          | d
d	¦  «        }t          |dd	¦  «        }|rt          d„ |D ¦   «         ¦  «        rt	          d¦  «        ‚t          |¦  «        t          t          |¦  «        ¦  «        k    rt	          d¦  «        ‚|€| n|}t          t          j                             d¦  «        j        ¦  «        j        }t          j                             d|dz  ¦  «        }t          j                             |¦  «        }|j                             |¦  «         t)          |||||t          |¦  «        ¬¦  «        }|                     ||¦  «        }t-          ||¦  «         d„ |D ¦   «         Šˆfd„|                      ¦   «         D ¦   «         }|r|r|                     |||ddœ¦  «         t-          | |¦  «         |S )aF  Delegate core grouping to RF-DETR; explicit adapter extension uses base LR/WD.

    For a project wrapper, pass core_model=model.model. Missing wrapper parameters
    fail by default. Opting into adapter coverage is a declared research extension,
    not a claim that the sensor adapter exists in upstream's model recipe.
    r   rZ   )r9   r[   rd   re   rf   r   rf   z must be finite and ÚnonnegativeÚpositiveNÚmodel_configr1  c              3  óx   K  — | ]5}t          |t          ¦  «        pt          |t          ¦  «         p|d k     V — Œ6dS )r   N)r"   r#   r$   )rW   Úis     r   ú	<genexpr>z,build_rfdetr_param_groups.<locals>.<genexpr>  sK   è è € Ð%zÐ%zÐab¥j°µDÑ&9Ô&9Ð&\ÅÈAÍsÑASÔASÐ=SÐ&\ÐWXÐ[\ÒW\Ð%zÐ%zÐ%zÐ%zÐ%zÐ%zr   zBout_feature_indexes must contain nonnegative integer layer indicesz/out_feature_indexes must be strictly increasingÚ$_pallet_official_rfdetr_param_groups)r9   r[   rd   re   rf   r1  c                óB   — h | ]}|d          D ]}t          |¦  «        ’ŒŒS rý   )rù   )rW   r	  rX   s      r   ú	<setcomp>z,build_rfdetr_param_groups.<locals>.<setcomp>'  s/   € ÐBÐBÐB˜°%¸´/ÐBÐB¨Q�r�!‰uŒuÐBÐBÐBÐBr   c                óD   •— g | ]}|j         ¯	t          |¦  «        ‰v¯|‘ŒS r¥   r÷   )rW   rX   Úcovereds     €r   rY   z-build_rfdetr_param_groups.<locals>.<listcomp>(  s1   ø€ ÐZÐZÐZ�a°´ÐZÅRÈÁUÄUÐRYÐEYÐEY�ÐEYÐEYÐEYr   Ú#sensor_adapter_base_lr_weight_decay)rû   r9   rf   Úrecipe_extension)r   r?   r@   r$  r%   ÚgetattrÚanyÚlistÚsortedr  r   r   r   r   r   r   Úspec_from_file_locationÚmodule_from_specÚloaderÚexec_moduler   Úget_param_dictr  Ú
parametersrí   )rS   r0  r1  r2  r9   r[   rd   re   rf   r&   r'   ÚconfigÚcorer   Úmodule_specÚofficial_moduler  rh   Úadaptersr>  s                      @r   Úbuild_rfdetr_param_groupsrP    s¥  ø€ õ ˆHÐ0Ñ1Ô1Ð1Ø °
Ø.@ÐXjØ(4ð6ð 6ç6;²e±g´gðuð u‰ˆˆeõ Œ}˜UÑ#Ô#ð 	u u¨q¢y y°T¸^Ò5KÐ5KÐPUÐYZÒPZÐPZÝ ÐsÐsÈ4ÐSaÒKaÐKa¸-¸-ÐgqÐsÐsÑtÔtÐtøØÐ"Ý˜ °Ñ5Ô5ˆÝ% fÐ.CÀTÑJÔJÐØð _¥#Ð%zÐ%zÐfyÐ%zÑ%zÔ%zÑ"zÔ"zð _ÝÐ]Ñ^Ô^Ð^ÝÐÑ Ô ¥F­3Ð/BÑ+CÔ+CÑ$DÔ$DÒDÐDÝÐJÑKÔKÐKØÐ&ˆ5ˆ5¨J€Dõ •9”>×+Ò+¨HÑ5Ô5Ô<Ñ=Ô=ÔD€GÝ”.×8Ò8Ø.°Ð:TÑ0TñVô V€Kå”n×5Ò5°kÑBÔB€OØÔ×"Ò" ?Ñ3Ô3Ð3Ý˜b¨ZÐL^Ø.@È|Ý/3Ð4GÑ/HÔ/HðJñ Jô J€Dð ×+Ò+¨D°$Ñ7Ô7€FÝ  fÑ-Ô-Ð-ØBÐB &ÐBÑBÔB€GØZÐZÐZÐZ˜5×+Ò+Ñ-Ô-ÐZÑZÔZ€HØð SÐ,ð SØ�Š °À\Ø+PðRð Rñ 	Sô 	Sð 	Så  vÑ.Ô.Ð.Ø€Mr   )rj   c               ó²   — t          | fi |¤Ž}t          j                             ||                     dd¦  «        |                     dd¦  «        |¬¦  «        S )z£Construct official default AdamW numerics; caller explicitly chooses fused.

    Runtime fused eligibility, EMA and scheduler remain trainer responsibilities.
    r9   r^   rf   )r9   rf   rj   )rP  r&  Úoptimr4   Úget)rS   rj   Úgroup_kwargsrh   s       r   Úbuild_rfdetr_optimizerrU  0  se   € õ
 ' uÐ=Ð=°Ð=Ð=€FÝŒ;×Ò˜V¨×(8Ò(8¸¸tÑ(DÔ(DØ*6×*:Ò*:¸>È4Ñ*PÔ*PÐX]ð ñ _ô _ð _r   )r	   )Ú__doc__Ú
__future__r   r   Úimportlib.metadatar   Úimportlib.utilr@   Úpathlibr   Útypesr   r&  r   rç   r   r(   r,   rF   rò   r  r/  rP  rU  r¥   r   r   ú<module>r\     s²  ððð ð
 #Ð "Ð "Ð "Ð "Ð "à €€€Ø Ð Ð Ð Ø Ð Ð Ð Ø €€€Ø Ð Ð Ð Ð Ð Ø !Ð !Ð !Ð !Ð !Ð !à €€€à"+°xÐ@Ð@€Ø[€ðEð Eð Eð?ð ?ð ?ð
Lð Lð Lð LNÐRSð =ð =ð =ð =ð =ðQÐRSð Qð Qð Qð Qð Qðh BFð ð ð ð ð ð0 QRØ$&¨1¸6ðð ð ð ð ð@ 48ÈTØ7<ÀÐRXØ14ÈÐ[_ð)ð )ð )ð )ð )ðX ,1ð _ð _ð _ð _ð _ð _ð _r   