§
    |¯ºjþ  ã                  ód   — d Z ddlmZ ddlZddlmZ d„ Z G d„ d¦  «        Z G d	„ d
¦  «        ZdS )zBLocked RF-DETR numerics with declared sensor research adaptations.é    )ÚannotationsNé   )Úbuild_rfdetr_optimizerc                ó„   — | dk    s*|dk    s$d|cxk    rdk    rn n|dk    s|s|dk    s|dk    rt          d¦  «        ‚d S )	Ni€  é   r   i,  é2   r   g{®Gáz„?z½RF official_research_constrained requires size=640, effective_batch=16, 1<=epochs<=300, patience=50, mosaic=False, warmup_epochs=0; min_lr_ratio=.01 is unused by the official step scheduler)Ú
ValueError)ÚsizeÚeffective_batchÚepochsÚpatienceÚmosaicÚwarmup_epochsÚmin_lr_ratios          úM/root/private/project/pallet_detection_v2/pallet/experiments/rfdetr_recipe.pyÚvalidate_constraintsr   	   s}   € à�Šˆ�¨"Ò,Ð,°A¸Ð4FÐ4FÒ4FÐ4FÀ3Ò4FÐ4FÐ4FÐ4FÐ4FØ˜2Š~ˆ~ ˆ~¨=¸AÒ+=Ð+=ÀÐQTÒATÐATÝð Uñ Vô Vð 	Vð BUÐATó    c                  óV   — e Zd ZdZd„ Zed„ ¦   «         Zed„ ¦   «         Zd„ Zd„ Z	d„ Z
dS )	ÚRFResearchEMAzHUse the installed EMA callback's averaging, counter and state unchanged.c                ó  — ddl m} ddlm}  |dddd¬¦  «        | _         ||d| j        j        ¬	¦  «                             ¦   «         | j        _        | j         	                    ¦   «         D ]}| 
                    d
¦  «         Œd S )Nr   )ÚRFDETREMACallback)ÚAveragedModelg-²�ï§Æï?éd   Tr   )ÚdecayÚtauÚuse_buffersÚupdate_interval_steps)r   Úmulti_avg_fnF)Úrfdetr.training.callbacks.emar   Útorch.optim.swa_utilsr   ÚcallbackÚ_multi_avg_fnÚevalÚ_average_modelÚemaÚ
parametersÚrequires_grad_)ÚselfÚmodelr   r   Ú	parameters        r   Ú__init__zRFResearchEMA.__init__   s¿   € ØCÐCÐCÐCÐCÐCØ7Ð7Ð7Ð7Ð7Ð7à)Ð)°¸#È4Ø@AðCñ Cô CˆŒà'4 }Ø˜t°$´-Ô2Mð(Oñ (Oô (OßOSÊtÉvÌvð 	ŒÔ$àœ×,Ò,Ñ.Ô.ð 	,ð 	,ˆIØ×$Ò$ UÑ+Ô+Ð+Ð+ð	,ð 	,r   c                ó$   — | j         j        j        S ©N)r!   r$   Úmodule©r(   s    r   r%   zRFResearchEMA.ema   s   € àŒ}Ô+Ô2Ð2r   c                ób   — t          | j        j        j                             ¦   «         ¦  «        S r-   )Úintr!   r$   Ú
n_averagedÚitemr/   s    r   ÚupdateszRFResearchEMA.updates#   s$   € å�4”=Ô/Ô:×?Ò?ÑAÔAÑBÔBÐBr   c                ó<   — | j                              ||¦  «         d S r-   )r!   Ú_update_ema_for_step)r(   r)   Úsuccessful_stepss      r   ÚupdatezRFResearchEMA.update'   s!   € ØŒ×*Ò*¨5Ð2BÑCÔCÐCÐCÐCr   c                ó4   — | j                              ¦   «         S r-   )r!   Ú
state_dictr/   s    r   r:   zRFResearchEMA.state_dict*   s   € ØŒ}×'Ò'Ñ)Ô)Ð)r   c                ón   — | j                              |¦  «         | j                              |¦  «         d S r-   )r!   Úload_state_dictÚ_restore_pending_average_state)r(   Ústater)   s      r   r<   zRFResearchEMA.load_state_dict-   s4   € ØŒ×%Ò% eÑ,Ô,Ð,ØŒ×4Ò4°UÑ;Ô;Ð;Ð;Ð;r   N)Ú__name__Ú
__module__Ú__qualname__Ú__doc__r+   Úpropertyr%   r4   r8   r:   r<   © r   r   r   r      s’   € € € € € ØRÐRð	,ð 	,ð 	,ð ð3ð 3ñ „Xð3ð ðCð Cñ „XðCðDð Dð Dð*ð *ð *ð<ð <ð <ð <ð <r   r   c                  óD   — e Zd ZdZdZdZd„ Zd„ Zd„ Zd„ Z	d„ Z
d	„ Zd
„ ZdS )ÚRFResearchRecipea‘  Official groups, managed step scheduler and EMA on project sensor tensors.

    Full equal microbatch windows average the official criterion loss exactly
    once, matching Lightning automatic optimization. Epoch tails use an actual
    example-weighted average, rather than Lightning's fixed accumulation divisor.
    Criterion box normalization remains per microbatch, as in upstream detection.
    çš™™™™™¹?Ú&fixed16_actual_example_mean_epoch_tailc               ó¶  — ddl m} ddlm} || _         |d||d|z  d¬¦  «        }	t          t          |                     ¦   «         ¦  «        j        o9|t          j
        k    o)|j        j        ot          j                             ¦   «         ¦  «        | _        t!          ||j        d| j        ¬	¦  «        | _        t%          d
||	j        z   d
z
  |	j        z  ¦  «        | _        d| _         || j        |	|| j        z  | j        d¦  «        | _        t/          |¦  «        | _        d| _        d|	j        |	j        |	j        ddg| j        ddd| j        ddœ| _        d S )Nr   )ÚTrainConfig)Ú_build_managed_schedulerú.r   g        )Údataset_dirr   Ú
batch_sizeÚgrad_accum_stepsr   T)Ú
core_modelÚallow_adapter_parametersÚfusedr   ÚAdamWgÍÌÌÌÌÌì?g+‡ÙÎ÷ï?rG   Úoptimizer_stepr   Ú#sensor_adapter_base_lr_weight_decay)ÚnameÚlrÚ
lr_encoderÚweight_decayÚbetasrR   Úclip_max_normÚscheduler_intervalÚlr_drop_epochsÚsteps_per_epochÚadapter_extension)Úrfdetr.configrJ   Úrfdetr.training.module_modelrK   r)   ÚboolÚnextr&   Úis_cudaÚtorchÚbfloat16Úmodel_configÚfused_optimizerÚcudaÚis_bf16_supportedrR   r   Ú	optimizerÚmaxrO   r^   Úwarmup_iterationsÚ	schedulerr   r%   Úoptimizer_updatesrW   rX   rY   Úresolved_optimizer)
r(   r)   Úsamplesr   rN   Únum_batchesÚ	amp_dtyperJ   rK   Úconfigs
             r   r+   zRFResearchRecipe.__init__=   s�  € Ø-Ð-Ð-Ð-Ð-Ð-ØIÐIÐIÐIÐIÐIàˆŒ
Ø�¨°VÈ
Ø.0°JÑ.>ÈbðRñ Rô Rˆå�$˜u×/Ò/Ñ1Ô1Ñ2Ô2Ô:ð d¸yÍEÌNÒ?Zð dØ#Ô0Ô@ðdÝEJÄZ×EaÒEaÑEcÔEcñeô eˆŒ
å/Ø˜eœkÀDÐPTÔPZð\ñ \ô \ˆŒå" 1 {°VÔ5LÑ'LÈqÑ'PÐU[ÔUlÑ&lÑmÔmˆÔØ!"ˆÔØ1Ð1ØŒN˜F F¨TÔ-AÑ$AÀ4ÔCWÐYZñ\ô \ˆŒå  Ñ'Ô'ˆŒØ!"ˆÔØ+2¸&¼)Ø17Ô1BØ39Ô3FÐRTÐVZÐQ[Ø,0¬JÈØ9IØ58ÈTÔMaØ8]ð#_ð #_ˆÔÐÐr   c                ó   — d S r-   rD   r/   s    r   Úbegin_epochzRFResearchRecipe.begin_epochV   ó   € Øˆr   c                ó   — d S r-   rD   )r(   Ú	iterationÚepochs      r   ÚwarmupzRFResearchRecipe.warmupY   rw   r   c                óª   — |rP| xj         dz  c_         | j                             ¦   «          | j                             | j        | j         ¦  «         d S d S )Nr   )ro   rn   Ústepr%   r8   r)   )r(   Ú
successfuls     r   Ú
after_stepzRFResearchRecipe.after_step\   s`   € Øð 	@ØÐ"Ô" aÑ'Ð"Ô"ØŒN×ÒÑ!Ô!Ð!ØŒH�OŠO˜DœJ¨Ô(>Ñ?Ô?Ð?Ð?Ð?ð	@ð 	@r   c                ó   — d S r-   rD   r/   s    r   Ú	end_epochzRFResearchRecipe.end_epochb   rw   r   c                ó®   — | j         j                              ¦   «         | j         j        | j                              ¦   «         | j        | j        | j        ddœS )Nr%   )r%   Úema_updatesÚema_callbackro   rp   Úaccumulation_policyÚselection_model)r%   r:   r4   ro   rp   r…   r/   s    r   r:   zRFResearchRecipe.state_dicte   sP   € Ø”x”|×.Ò.Ñ0Ô0ÀÄÔAQØ $¤× 3Ò 3Ñ 5Ô 5ÈDÔLbØ&*Ô&=Ø'+Ô'?Ø#(ð	*ð *ð 	*r   c                ó¶   — | j                              |d         | j        ¦  «         |d         | _        | j         j        |d         k    rt          d¦  «        ‚d S )Nr„   ro   rƒ   z0RF EMA counter differs from saved callback state)r%   r<   r)   ro   r4   r	   )r(   r>   s     r   r<   z RFResearchRecipe.load_state_dictl   s[   € ØŒ× Ò   ~Ô!6¸¼
ÑCÔCÐCØ!&Ð':Ô!;ˆÔØŒ8Ô˜u ]Ô3Ò3Ð3ÝÐOÑPÔPÐPð 4Ð3r   N)r?   r@   rA   rB   r[   r…   r+   rv   r{   r   r�   r:   r<   rD   r   r   rF   rF   2   sŸ   € € € € € ðð ð €MØBÐð_ð _ð _ð2ð ð ðð ð ð@ð @ð @ðð ð ð*ð *ð *ðQð Qð Qð Qð Qr   rF   )	rB   Ú
__future__r   re   Úrecipesr   r   r   rF   rD   r   r   ú<module>rŠ      s¯   ðØ HÐ HØ "Ð "Ð "Ð "Ð "Ð "à €€€à +Ð +Ð +Ð +Ð +Ð +ðVð Vð Vð<ð <ð <ð <ð <ñ <ô <ð <ð@>Qð >Qð >Qð >Qð >Qñ >Qô >Qð >Qð >Qð >Qr   